Merge f298643fcf into 02b1436f39

README.md

@@ -73,3 +73,6 @@ The output folder should be an empty folder where the output model will be writt
 * weighted_loss: If ``true``, this means that you want to apply weighted categorical_crossentropy as loss fucntion. Be carefull if you set to ``true``the parameter "is_loss_soft_dice" should be ``false``
 * data_is_provided: If you have already provided the input data you can set this to ``true``. Be sure that the train and eval data are in "dir_output". Since when once we provide training data we resize and augment them and then we write them in sub-directories train and eval in "dir_output". 
 * dir_train: This is the directory of "images" and "labels" (dir_train should include two subdirectories with names of images and labels ) for raw images and labels. Namely they are not prepared (not resized and not augmented) yet for training the model. When we run this tool these raw data will be transformed to suitable size needed for the model and they will be written in "dir_output" in train and eval directories. Each of train and eval include "images" and "labels" sub-directories.
+
+#### Additional documentation
+Please check the [wiki](https://github.com/qurator-spk/sbb_pixelwise_segmentation/wiki).

config_params.json

@@ -1,30 +1,58 @@
 {
-    "n_classes" : 3,
+    "backbone_type" : "transformer",
-    "n_epochs" : 2,
+    "task": "segmentation",
+    "n_classes" : 2,
+    "n_epochs" : 0,
     "input_height" : 448,
-    "input_width" : 672,
+    "input_width" : 448,
     "weight_decay" : 1e-6,
-    "n_batch" : 2,
+    "n_batch" : 1,
     "learning_rate": 1e-4,
-    "patches" : true,
+    "patches" : false,
     "pretraining" : true,
-    "augmentation" : false,
+    "augmentation" : true,
     "flip_aug" : false,
     "blur_aug" : false,
-    "scaling" : true,
+    "scaling" : false,
-    "binarization" : false,
+    "adding_rgb_background": true,
+    "adding_rgb_foreground": true,
+    "add_red_textlines": false,
+    "channels_shuffling": false,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : true,
     "scaling_bluring" : false,
     "scaling_binarization" : false,
     "scaling_flip" : false,
     "rotation": false,
     "rotation_not_90": false,
+    "transformer_num_patches_xy": [56, 56],
+    "transformer_patchsize_x": 4,
+    "transformer_patchsize_y": 4,
+    "transformer_projection_dim": 64,
+    "transformer_mlp_head_units": [128, 64],
+    "transformer_layers": 1,
+    "transformer_num_heads": 1,
+    "transformer_cnn_first": false,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "shuffle_indexes" : [ [0,2,1], [1,2,0], [1,0,2] , [2,1,0]],
+    "thetha" : [5, -5],
+    "number_of_backgrounds_per_image": 2,
     "continue_training": false,
-    "index_start": 0,
+    "index_start" : 0,
-    "dir_of_start_model": " ", 
+    "dir_of_start_model" : " ", 
     "weighted_loss": false,
     "is_loss_soft_dice": false,
     "data_is_provided": false,
-    "dir_train": "/train",
+    "dir_train": "/home/vahid/Documents/test/sbb_pixelwise_segmentation/test_label/pageextractor_test/train_new",
-    "dir_eval": "/eval",
+    "dir_eval": "/home/vahid/Documents/test/sbb_pixelwise_segmentation/test_label/pageextractor_test/eval_new",
-    "dir_output": "/output"
+    "dir_output": "/home/vahid/Documents/test/sbb_pixelwise_segmentation/test_label/pageextractor_test/output_new",
+    "dir_rgb_backgrounds": "/home/vahid/Documents/1_2_test_eynollah/set_rgb_background",
+    "dir_rgb_foregrounds": "/home/vahid/Documents/1_2_test_eynollah/out_set_rgb_foreground",
+    "dir_img_bin": "/home/vahid/Documents/test/sbb_pixelwise_segmentation/test_label/pageextractor_test/train_new/images_bin"
+    
 }

custom_config_page2label.json

@@ -0,0 +1,8 @@
+{
+"use_case": "textline",
+"textregions":{ "rest_as_paragraph": 1, "header":2 , "heading":2 , "marginalia":3 },
+"imageregion":4,
+"separatorregion":5,
+"graphicregions" :{"rest_as_decoration":6},
+"columns_width":{"1":1000, "2":1300, "3":1600, "4":2000, "5":2300, "6":2500}
+}

generate_gt_for_training.py

@@ -0,0 +1,221 @@
+import click
+import json
+from gt_gen_utils import *
+from tqdm import tqdm
+
+@click.group()
+def main():
+    pass
+
+@main.command()
+@click.option(
+    "--dir_xml",
+    "-dx",
+    help="directory of GT page-xml files",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--dir_images",
+    "-di",
+    help="directory of org images. If print space cropping or scaling is needed for labels it would be great to provide the original images to apply the same function on them. So if -ps is not set true or in config files no columns_width key is given this argumnet can be ignored. File stems in this directory should be the same as those in dir_xml.",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--dir_out_images",
+    "-doi",
+    help="directory where the output org images after undergoing a process (like print space cropping or scaling) will be written.",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--dir_out",
+    "-do",
+    help="directory where ground truth label images would be written",
+    type=click.Path(exists=True, file_okay=False),
+)
+
+@click.option(
+    "--config",
+    "-cfg",
+    help="config file of prefered layout or use case.",
+    type=click.Path(exists=True, dir_okay=False),
+)
+
+@click.option(
+    "--type_output",
+    "-to",
+    help="this defines how output should be. A 2d image array or a 3d image array encoded with RGB color. Just pass 2d or 3d. The file will be saved one directory up. 2D image array is 3d but only information of one channel would be enough since all channels have the same values.",
+)
+@click.option(
+    "--printspace",
+    "-ps",
+    is_flag=True,
+    help="if this parameter set to true, generated labels and in the case of provided org images cropping will be imposed and cropped labels and images will be written in output directories.",
+)
+
+def pagexml2label(dir_xml,dir_out,type_output,config, printspace, dir_images, dir_out_images):
+    if config:
+        with open(config) as f:
+            config_params = json.load(f)
+    else:
+        print("passed")
+        config_params = None
+    gt_list = get_content_of_dir(dir_xml)
+    get_images_of_ground_truth(gt_list,dir_xml,dir_out,type_output, config, config_params, printspace, dir_images, dir_out_images)
+    
+@main.command()
+@click.option(
+    "--dir_imgs",
+    "-dis",
+    help="directory of images with high resolution.",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--dir_out_images",
+    "-dois",
+    help="directory where degraded images will be written.",
+    type=click.Path(exists=True, file_okay=False),
+)
+
+@click.option(
+    "--dir_out_labels",
+    "-dols",
+    help="directory where original images will be written as labels.",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--scales",
+    "-scs",
+    help="json dictionary where the scales are written.",
+    type=click.Path(exists=True, dir_okay=False),
+)
+def image_enhancement(dir_imgs, dir_out_images, dir_out_labels, scales):
+    ls_imgs = os.listdir(dir_imgs)
+    with open(scales) as f:
+        scale_dict = json.load(f)
+    ls_scales = scale_dict['scales']
+
+    for img in tqdm(ls_imgs):
+        img_name = img.split('.')[0]
+        img_type = img.split('.')[1]
+        image = cv2.imread(os.path.join(dir_imgs, img))
+        for i, scale in enumerate(ls_scales):
+            height_sc = int(image.shape[0]*scale)
+            width_sc = int(image.shape[1]*scale)
+            
+            image_down_scaled = resize_image(image, height_sc, width_sc)
+            image_back_to_org_scale = resize_image(image_down_scaled, image.shape[0], image.shape[1])
+            
+            cv2.imwrite(os.path.join(dir_out_images, img_name+'_'+str(i)+'.'+img_type), image_back_to_org_scale)
+            cv2.imwrite(os.path.join(dir_out_labels, img_name+'_'+str(i)+'.'+img_type), image)
+    
+    
+@main.command()
+@click.option(
+    "--dir_xml",
+    "-dx",
+    help="directory of GT page-xml files",
+    type=click.Path(exists=True, file_okay=False),
+)
+
+@click.option(
+    "--dir_out_modal_image",
+    "-domi",
+    help="directory where ground truth images would be written",
+    type=click.Path(exists=True, file_okay=False),
+)
+
+@click.option(
+    "--dir_out_classes",
+    "-docl",
+    help="directory where ground truth classes would be written",
+    type=click.Path(exists=True, file_okay=False),
+)
+
+@click.option(
+    "--input_height",
+    "-ih",
+    help="input height",
+)
+@click.option(
+    "--input_width",
+    "-iw",
+    help="input width",
+)
+@click.option(
+    "--min_area_size",
+    "-min",
+    help="min area size of regions considered for reading order training.",
+)
+
+def machine_based_reading_order(dir_xml, dir_out_modal_image, dir_out_classes, input_height, input_width, min_area_size):
+    xml_files_ind = os.listdir(dir_xml)
+    input_height = int(input_height)
+    input_width = int(input_width)
+    min_area = float(min_area_size)
+
+    indexer_start= 0#55166
+    max_area = 1
+    #min_area = 0.0001
+
+    for ind_xml in tqdm(xml_files_ind):
+        indexer = 0
+        #print(ind_xml)
+        #print('########################')
+        xml_file = os.path.join(dir_xml,ind_xml )
+        f_name = ind_xml.split('.')[0]
+        _, _, _, file_name, id_paragraph, id_header,co_text_paragraph,co_text_header,tot_region_ref,x_len, y_len,index_tot_regions,img_poly = read_xml(xml_file)
+        
+        id_all_text = id_paragraph + id_header
+        co_text_all = co_text_paragraph + co_text_header
+        
+        
+        _, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = find_new_features_of_contours(co_text_header)
+        
+        img_header_and_sep = np.zeros((y_len,x_len), dtype='uint8')
+
+        for j in range(len(cy_main)):
+            img_header_and_sep[int(y_max_main[j]):int(y_max_main[j])+12,int(x_min_main[j]):int(x_max_main[j]) ] = 1
+
+
+        texts_corr_order_index  = [index_tot_regions[tot_region_ref.index(i)] for i in id_all_text ]
+        texts_corr_order_index_int = [int(x) for x in texts_corr_order_index]
+        
+
+        co_text_all, texts_corr_order_index_int = filter_contours_area_of_image(img_poly, co_text_all, texts_corr_order_index_int, max_area, min_area)
+        
+        arg_array = np.array(range(len(texts_corr_order_index_int)))
+        
+        labels_con = np.zeros((y_len,x_len,len(arg_array)),dtype='uint8')
+        for i in range(len(co_text_all)):
+            img_label = np.zeros((y_len,x_len,3),dtype='uint8')
+            img_label=cv2.fillPoly(img_label, pts =[co_text_all[i]], color=(1,1,1))
+            
+            img_label[:,:,0][img_poly[:,:,0]==5] = 2
+            img_label[:,:,0][img_header_and_sep[:,:]==1] = 3
+            
+            labels_con[:,:,i] = img_label[:,:,0]
+        
+        for i in range(len(texts_corr_order_index_int)):
+            for j in range(len(texts_corr_order_index_int)):
+                if i!=j:
+                    input_multi_visual_modal = np.zeros((input_height,input_width,3)).astype(np.int8)
+                    final_f_name = f_name+'_'+str(indexer+indexer_start)
+                    order_class_condition = texts_corr_order_index_int[i]-texts_corr_order_index_int[j]
+                    if order_class_condition<0:
+                        class_type = 1
+                    else:
+                        class_type = 0
+
+                    input_multi_visual_modal[:,:,0] = resize_image(labels_con[:,:,i], input_height, input_width)
+                    input_multi_visual_modal[:,:,1] = resize_image(img_poly[:,:,0], input_height, input_width)
+                    input_multi_visual_modal[:,:,2] = resize_image(labels_con[:,:,j], input_height, input_width)
+
+                    np.save(os.path.join(dir_out_classes,final_f_name+'.npy' ), class_type)
+                    
+                    cv2.imwrite(os.path.join(dir_out_modal_image,final_f_name+'.png' ), input_multi_visual_modal)
+                    indexer = indexer+1
+
+    
+    
+if __name__ == "__main__":
+    main()

inference.py

@@ -0,0 +1,644 @@
+import sys
+import os
+import numpy as np
+import warnings
+import cv2
+import seaborn as sns
+from tensorflow.keras.models import load_model
+import tensorflow as tf
+from tensorflow.keras import backend as K
+from tensorflow.keras import layers
+import tensorflow.keras.losses
+from tensorflow.keras.layers import *
+from models import *
+from gt_gen_utils import *
+import click
+import json
+from tensorflow.python.keras import backend as tensorflow_backend
+import xml.etree.ElementTree as ET
+import matplotlib.pyplot as plt
+
+
+with warnings.catch_warnings():
+    warnings.simplefilter("ignore")
+    
+__doc__=\
+"""
+Tool to load model and predict for given image.
+"""
+
+class sbb_predict:
+    def __init__(self,image, model, task, config_params_model, patches, save, save_layout, ground_truth, xml_file, out, min_area):
+        self.image=image
+        self.patches=patches
+        self.save=save
+        self.save_layout=save_layout
+        self.model_dir=model
+        self.ground_truth=ground_truth
+        self.task=task
+        self.config_params_model=config_params_model
+        self.xml_file = xml_file
+        self.out = out
+        if min_area:
+            self.min_area = float(min_area)
+        else:
+            self.min_area = 0
+
+    def resize_image(self,img_in,input_height,input_width):
+        return cv2.resize( img_in, ( input_width,input_height) ,interpolation=cv2.INTER_NEAREST)
+    
+    
+    def color_images(self,seg):
+        ann_u=range(self.n_classes)
+        if len(np.shape(seg))==3:
+            seg=seg[:,:,0]
+            
+        seg_img=np.zeros((np.shape(seg)[0],np.shape(seg)[1],3)).astype(np.uint8)
+        colors=sns.color_palette("hls", self.n_classes)
+        
+        for c in ann_u:
+            c=int(c)
+            segl=(seg==c)
+            seg_img[:,:,0][seg==c]=c
+            seg_img[:,:,1][seg==c]=c
+            seg_img[:,:,2][seg==c]=c
+        return seg_img
+    
+    def otsu_copy_binary(self,img):
+        img_r=np.zeros((img.shape[0],img.shape[1],3))
+        img1=img[:,:,0]
+
+        #print(img.min())
+        #print(img[:,:,0].min())
+        #blur = cv2.GaussianBlur(img,(5,5))
+        #ret3,th3 = cv2.threshold(blur,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+        retval1, threshold1 = cv2.threshold(img1, 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+
+        
+
+        img_r[:,:,0]=threshold1
+        img_r[:,:,1]=threshold1
+        img_r[:,:,2]=threshold1
+        #img_r=img_r/float(np.max(img_r))*255
+        return img_r
+    
+    def otsu_copy(self,img):
+        img_r=np.zeros((img.shape[0],img.shape[1],3))
+        #img1=img[:,:,0]
+
+        #print(img.min())
+        #print(img[:,:,0].min())
+        #blur = cv2.GaussianBlur(img,(5,5))
+        #ret3,th3 = cv2.threshold(blur,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+        _, threshold1 = cv2.threshold(img[:,:,0], 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+        _, threshold2 = cv2.threshold(img[:,:,1], 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+        _, threshold3 = cv2.threshold(img[:,:,2], 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)
+
+        
+
+        img_r[:,:,0]=threshold1
+        img_r[:,:,1]=threshold2
+        img_r[:,:,2]=threshold3
+        ###img_r=img_r/float(np.max(img_r))*255
+        return img_r
+    
+    def soft_dice_loss(self,y_true, y_pred, epsilon=1e-6): 
+
+        axes = tuple(range(1, len(y_pred.shape)-1))
+        
+        numerator = 2. * K.sum(y_pred * y_true, axes)
+    
+        denominator = K.sum(K.square(y_pred) + K.square(y_true), axes)
+        return 1.00 - K.mean(numerator / (denominator + epsilon)) # average over classes and batch
+    
+    def weighted_categorical_crossentropy(self,weights=None):
+
+        def loss(y_true, y_pred):
+            labels_floats = tf.cast(y_true, tf.float32)
+            per_pixel_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels_floats,logits=y_pred)
+        
+            if weights is not None:
+                weight_mask = tf.maximum(tf.reduce_max(tf.constant(
+                    np.array(weights, dtype=np.float32)[None, None, None])
+                    * labels_floats, axis=-1), 1.0)
+                per_pixel_loss = per_pixel_loss * weight_mask[:, :, :, None]
+            return tf.reduce_mean(per_pixel_loss)
+        return self.loss
+
+
+    def IoU(self,Yi,y_predi):
+        ## mean Intersection over Union
+        ## Mean IoU = TP/(FN + TP + FP)
+    
+        IoUs = []
+        Nclass = np.unique(Yi)
+        for c in Nclass:
+            TP = np.sum( (Yi == c)&(y_predi==c) )
+            FP = np.sum( (Yi != c)&(y_predi==c) )
+            FN = np.sum( (Yi == c)&(y_predi != c)) 
+            IoU = TP/float(TP + FP + FN)
+            if self.n_classes>2:
+                print("class {:02.0f}: #TP={:6.0f}, #FP={:6.0f}, #FN={:5.0f}, IoU={:4.3f}".format(c,TP,FP,FN,IoU))
+            IoUs.append(IoU)
+        if self.n_classes>2:
+            mIoU = np.mean(IoUs)
+            print("_________________")
+            print("Mean IoU: {:4.3f}".format(mIoU))
+            return mIoU
+        elif self.n_classes==2:
+            mIoU = IoUs[1]
+            print("_________________")
+            print("IoU: {:4.3f}".format(mIoU))
+            return mIoU
+            
+    def start_new_session_and_model(self):
+        
+        config = tf.compat.v1.ConfigProto()
+        config.gpu_options.allow_growth = True
+
+        session = tf.compat.v1.Session(config=config)  # tf.InteractiveSession()
+        tensorflow_backend.set_session(session)
+        #tensorflow.keras.layers.custom_layer = PatchEncoder
+        #tensorflow.keras.layers.custom_layer = Patches
+        self.model = load_model(self.model_dir , compile=False,custom_objects = {"PatchEncoder": PatchEncoder, "Patches": Patches})
+        #config = tf.ConfigProto()
+        #config.gpu_options.allow_growth=True
+    
+        #self.session = tf.InteractiveSession()
+        #keras.losses.custom_loss = self.weighted_categorical_crossentropy
+        #self.model = load_model(self.model_dir , compile=False)
+
+        
+        ##if self.weights_dir!=None:
+            ##self.model.load_weights(self.weights_dir)
+            
+        if (self.task != 'classification' and self.task != 'reading_order'):
+            self.img_height=self.model.layers[len(self.model.layers)-1].output_shape[1]
+            self.img_width=self.model.layers[len(self.model.layers)-1].output_shape[2]
+            self.n_classes=self.model.layers[len(self.model.layers)-1].output_shape[3]
+        
+    def visualize_model_output(self, prediction, img, task):
+        if task == "binarization":
+            prediction = prediction * -1
+            prediction = prediction + 1
+            added_image = prediction * 255
+            layout_only = None
+        else:
+            unique_classes = np.unique(prediction[:,:,0])
+            rgb_colors = {'0' : [255, 255, 255],
+                        '1' : [255, 0, 0],
+                        '2' : [255, 125, 0],
+                        '3' : [255, 0, 125],
+                        '4' : [125, 125, 125],
+                        '5' : [125, 125, 0],
+                        '6' : [0, 125, 255],
+                        '7' : [0, 125, 0],
+                        '8' : [125, 125, 125],
+                        '9' : [0, 125, 255],
+                        '10' : [125, 0, 125],
+                        '11' : [0, 255, 0],
+                        '12' : [0, 0, 255],
+                        '13' : [0, 255, 255],
+                        '14' : [255, 125, 125],
+                        '15' : [255, 0, 255]}
+        
+            layout_only = np.zeros(prediction.shape)
+        
+            for unq_class in unique_classes:
+                rgb_class_unique = rgb_colors[str(int(unq_class))]
+                layout_only[:,:,0][prediction[:,:,0]==unq_class] = rgb_class_unique[0]
+                layout_only[:,:,1][prediction[:,:,0]==unq_class] = rgb_class_unique[1]
+                layout_only[:,:,2][prediction[:,:,0]==unq_class] = rgb_class_unique[2]
+        
+        
+        
+            img = self.resize_image(img, layout_only.shape[0], layout_only.shape[1])
+        
+            layout_only = layout_only.astype(np.int32)
+            img = img.astype(np.int32)
+        
+            
+            
+            added_image = cv2.addWeighted(img,0.5,layout_only,0.1,0)
+            
+        return added_image, layout_only
+
+    def predict(self):
+        self.start_new_session_and_model()
+        if self.task == 'classification':
+            classes_names = self.config_params_model['classification_classes_name']
+            img_1ch = img=cv2.imread(self.image, 0)
+
+            img_1ch = img_1ch / 255.0
+            img_1ch = cv2.resize(img_1ch, (self.config_params_model['input_height'], self.config_params_model['input_width']), interpolation=cv2.INTER_NEAREST)
+            img_in = np.zeros((1, img_1ch.shape[0], img_1ch.shape[1], 3))
+            img_in[0, :, :, 0] = img_1ch[:, :]
+            img_in[0, :, :, 1] = img_1ch[:, :]
+            img_in[0, :, :, 2] = img_1ch[:, :]
+                      
+            label_p_pred = self.model.predict(img_in, verbose=0)
+            index_class = np.argmax(label_p_pred[0])
+            
+            print("Predicted Class: {}".format(classes_names[str(int(index_class))]))
+        elif self.task == 'reading_order':
+            img_height = self.config_params_model['input_height']
+            img_width = self.config_params_model['input_width']
+            
+            tree_xml, root_xml, bb_coord_printspace, file_name, id_paragraph, id_header, co_text_paragraph, co_text_header, tot_region_ref, x_len, y_len, index_tot_regions, img_poly = read_xml(self.xml_file)
+            _, cy_main, x_min_main, x_max_main, y_min_main, y_max_main, _ = find_new_features_of_contours(co_text_header)
+            
+            img_header_and_sep = np.zeros((y_len,x_len), dtype='uint8')
+            
+
+            for j in range(len(cy_main)):
+                img_header_and_sep[int(y_max_main[j]):int(y_max_main[j])+12,int(x_min_main[j]):int(x_max_main[j]) ] = 1
+                
+            co_text_all = co_text_paragraph + co_text_header
+            id_all_text = id_paragraph + id_header
+            
+
+            ##texts_corr_order_index  = [index_tot_regions[tot_region_ref.index(i)] for i in id_all_text ]
+            ##texts_corr_order_index_int = [int(x) for x in texts_corr_order_index]
+            texts_corr_order_index_int = list(np.array(range(len(co_text_all))))
+            
+            #print(texts_corr_order_index_int)
+            
+            max_area = 1
+            #print(np.shape(co_text_all[0]), len( np.shape(co_text_all[0]) ),'co_text_all')
+            #co_text_all = filter_contours_area_of_image_tables(img_poly, co_text_all, _, max_area, min_area)
+            #print(co_text_all,'co_text_all')
+            co_text_all, texts_corr_order_index_int = filter_contours_area_of_image(img_poly, co_text_all, texts_corr_order_index_int, max_area, self.min_area)
+            
+            #print(texts_corr_order_index_int)
+            
+            #co_text_all = [co_text_all[index] for index in texts_corr_order_index_int]
+            id_all_text = [id_all_text[index] for index in texts_corr_order_index_int]
+            
+            labels_con = np.zeros((y_len,x_len,len(co_text_all)),dtype='uint8')
+            for i in range(len(co_text_all)):
+                img_label = np.zeros((y_len,x_len,3),dtype='uint8')
+                img_label=cv2.fillPoly(img_label, pts =[co_text_all[i]], color=(1,1,1))
+                labels_con[:,:,i] = img_label[:,:,0]
+                
+            if bb_coord_printspace:
+                #bb_coord_printspace[x,y,w,h,_,_]
+                x = bb_coord_printspace[0]
+                y = bb_coord_printspace[1]
+                w = bb_coord_printspace[2]
+                h = bb_coord_printspace[3]
+                labels_con = labels_con[y:y+h, x:x+w, :]
+                img_poly = img_poly[y:y+h, x:x+w, :]
+                img_header_and_sep = img_header_and_sep[y:y+h, x:x+w]
+                
+
+                
+            img3= np.copy(img_poly)
+            labels_con = resize_image(labels_con, img_height, img_width)
+
+            img_header_and_sep = resize_image(img_header_and_sep, img_height, img_width)
+
+            img3= resize_image (img3, img_height, img_width)
+            img3 = img3.astype(np.uint16)
+            
+            inference_bs = 1#4
+
+            input_1= np.zeros( (inference_bs, img_height, img_width,3))
+
+
+            starting_list_of_regions = []
+            starting_list_of_regions.append( list(range(labels_con.shape[2])) )
+
+            index_update = 0
+            index_selected = starting_list_of_regions[0]
+            
+            scalibility_num = 0
+            while index_update>=0:
+                ij_list = starting_list_of_regions[index_update] 
+                i = ij_list[0]
+                ij_list.pop(0)
+                
+                
+                pr_list = []
+                post_list = []
+                
+                batch_counter = 0
+                tot_counter = 1
+                
+                tot_iteration = len(ij_list)
+                full_bs_ite= tot_iteration//inference_bs
+                last_bs = tot_iteration % inference_bs
+                
+                jbatch_indexer =[]
+                for j in ij_list:
+                    img1= np.repeat(labels_con[:,:,i][:, :, np.newaxis], 3, axis=2)
+                    img2 = np.repeat(labels_con[:,:,j][:, :, np.newaxis], 3, axis=2)
+
+                    
+                    img2[:,:,0][img3[:,:,0]==5] = 2
+                    img2[:,:,0][img_header_and_sep[:,:]==1] = 3
+                    
+                    
+                    
+                    img1[:,:,0][img3[:,:,0]==5] = 2
+                    img1[:,:,0][img_header_and_sep[:,:]==1] = 3
+                    
+                    #input_1= np.zeros( (height1, width1,3))
+                    
+
+                    jbatch_indexer.append(j)
+                        
+                    input_1[batch_counter,:,:,0] = img1[:,:,0]/3.
+                    input_1[batch_counter,:,:,2] = img2[:,:,0]/3.
+                    input_1[batch_counter,:,:,1] = img3[:,:,0]/5.
+                    #input_1[batch_counter,:,:,:]= np.zeros( (batch_counter, height1, width1,3))
+                    batch_counter = batch_counter+1
+                    
+                    #input_1[:,:,0] = img1[:,:,0]/3.
+                    #input_1[:,:,2] = img2[:,:,0]/3.
+                    #input_1[:,:,1] = img3[:,:,0]/5.
+                    
+                    if batch_counter==inference_bs or ( (tot_counter//inference_bs)==full_bs_ite and tot_counter%inference_bs==last_bs):
+                        y_pr = self.model.predict(input_1 , verbose=0)
+                        scalibility_num = scalibility_num+1
+                        
+                        if batch_counter==inference_bs:
+                            iteration_batches = inference_bs
+                        else:
+                            iteration_batches = last_bs
+                        for jb in range(iteration_batches):
+                            if y_pr[jb][0]>=0.5:
+                                post_list.append(jbatch_indexer[jb])
+                            else:
+                                pr_list.append(jbatch_indexer[jb])
+                                
+                        batch_counter = 0
+                        jbatch_indexer = []
+                        
+                    tot_counter = tot_counter+1
+                        
+                starting_list_of_regions, index_update = update_list_and_return_first_with_length_bigger_than_one(index_update, i, pr_list, post_list,starting_list_of_regions)
+            
+            
+            index_sort = [i[0] for i in starting_list_of_regions ]
+            
+            id_all_text = np.array(id_all_text)[index_sort]
+            
+            alltags=[elem.tag for elem in root_xml.iter()]
+            
+            
+            
+            link=alltags[0].split('}')[0]+'}'
+            name_space = alltags[0].split('}')[0]
+            name_space = name_space.split('{')[1]
+            
+            page_element = root_xml.find(link+'Page')
+            
+            """
+            ro_subelement = ET.SubElement(page_element, 'ReadingOrder')
+            #print(page_element, 'page_element')
+            
+            #new_element = ET.Element('ReadingOrder')
+            
+            new_element_element = ET.Element('OrderedGroup')
+            new_element_element.set('id', "ro357564684568544579089")
+            
+            for index, id_text in enumerate(id_all_text):
+                new_element_2 = ET.Element('RegionRefIndexed')
+                new_element_2.set('regionRef', id_all_text[index])
+                new_element_2.set('index', str(index_sort[index]))
+            
+                new_element_element.append(new_element_2)
+            
+            ro_subelement.append(new_element_element)
+            """
+            ##ro_subelement = ET.SubElement(page_element, 'ReadingOrder')
+            
+            ro_subelement = ET.Element('ReadingOrder')
+            
+            ro_subelement2 = ET.SubElement(ro_subelement, 'OrderedGroup')
+            ro_subelement2.set('id', "ro357564684568544579089")
+            
+            for index, id_text in enumerate(id_all_text):
+                new_element_2 = ET.SubElement(ro_subelement2, 'RegionRefIndexed')
+                new_element_2.set('regionRef', id_all_text[index])
+                new_element_2.set('index', str(index))
+            
+            if (link+'PrintSpace' in alltags) or  (link+'Border' in alltags):
+                page_element.insert(1, ro_subelement)
+            else:
+                page_element.insert(0, ro_subelement)
+            
+            alltags=[elem.tag for elem in root_xml.iter()]
+            
+            ET.register_namespace("",name_space)
+            tree_xml.write(os.path.join(self.out, file_name+'.xml'),xml_declaration=True,method='xml',encoding="utf8",default_namespace=None)
+            #tree_xml.write('library2.xml')
+            
+        else:
+            if self.patches:
+                #def textline_contours(img,input_width,input_height,n_classes,model):
+                
+                img=cv2.imread(self.image)
+                self.img_org = np.copy(img)
+                
+                if img.shape[0] < self.img_height:
+                    img = cv2.resize(img, (img.shape[1], self.img_width), interpolation=cv2.INTER_NEAREST)
+
+                if img.shape[1] < self.img_width:
+                    img = cv2.resize(img, (self.img_height, img.shape[0]), interpolation=cv2.INTER_NEAREST)
+                margin = int(0.1 * self.img_width)
+                width_mid = self.img_width - 2 * margin
+                height_mid = self.img_height - 2 * margin
+                img = img / float(255.0)
+
+                img_h = img.shape[0]
+                img_w = img.shape[1]
+
+                prediction_true = np.zeros((img_h, img_w, 3))
+                nxf = img_w / float(width_mid)
+                nyf = img_h / float(height_mid)
+
+                nxf = int(nxf) + 1 if nxf > int(nxf) else int(nxf)
+                nyf = int(nyf) + 1 if nyf > int(nyf) else int(nyf)
+                
+                for i in range(nxf):
+                    for j in range(nyf):
+                        if i == 0:
+                            index_x_d = i * width_mid
+                            index_x_u = index_x_d + self.img_width
+                        else:
+                            index_x_d = i * width_mid
+                            index_x_u = index_x_d + self.img_width
+                        if j == 0:
+                            index_y_d = j * height_mid
+                            index_y_u = index_y_d + self.img_height
+                        else:
+                            index_y_d = j * height_mid
+                            index_y_u = index_y_d + self.img_height
+
+                        if index_x_u > img_w:
+                            index_x_u = img_w
+                            index_x_d = img_w - self.img_width
+                        if index_y_u > img_h:
+                            index_y_u = img_h
+                            index_y_d = img_h - self.img_height
+
+                        img_patch = img[index_y_d:index_y_u, index_x_d:index_x_u, :]
+                        label_p_pred = self.model.predict(img_patch.reshape(1, img_patch.shape[0], img_patch.shape[1], img_patch.shape[2]),
+                                                                verbose=0)
+                        
+                        if self.task == 'enhancement':
+                            seg = label_p_pred[0, :, :, :]
+                            seg = seg * 255
+                        elif self.task == 'segmentation' or self.task == 'binarization':
+                            seg = np.argmax(label_p_pred, axis=3)[0]
+                            seg = np.repeat(seg[:, :, np.newaxis], 3, axis=2)
+                            
+
+                        if i == 0 and j == 0:
+                            seg = seg[0 : seg.shape[0] - margin, 0 : seg.shape[1] - margin]
+                            prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + 0 : index_x_u - margin, :] = seg
+                        elif i == nxf - 1 and j == nyf - 1:
+                            seg = seg[margin : seg.shape[0] - 0, margin : seg.shape[1] - 0]
+                            prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + margin : index_x_u - 0, :] = seg
+                        elif i == 0 and j == nyf - 1:
+                            seg = seg[margin : seg.shape[0] - 0, 0 : seg.shape[1] - margin]
+                            prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + 0 : index_x_u - margin, :] = seg
+                        elif i == nxf - 1 and j == 0:
+                            seg = seg[0 : seg.shape[0] - margin, margin : seg.shape[1] - 0]
+                            prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + margin : index_x_u - 0, :] = seg
+                        elif i == 0 and j != 0 and j != nyf - 1:
+                            seg = seg[margin : seg.shape[0] - margin, 0 : seg.shape[1] - margin]
+                            prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + 0 : index_x_u - margin, :] = seg
+                        elif i == nxf - 1 and j != 0 and j != nyf - 1:
+                            seg = seg[margin : seg.shape[0] - margin, margin : seg.shape[1] - 0]
+                            prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + margin : index_x_u - 0, :] = seg
+                        elif i != 0 and i != nxf - 1 and j == 0:
+                            seg = seg[0 : seg.shape[0] - margin, margin : seg.shape[1] - margin]
+                            prediction_true[index_y_d + 0 : index_y_u - margin, index_x_d + margin : index_x_u - margin, :] = seg
+                        elif i != 0 and i != nxf - 1 and j == nyf - 1:
+                            seg = seg[margin : seg.shape[0] - 0, margin : seg.shape[1] - margin]
+                            prediction_true[index_y_d + margin : index_y_u - 0, index_x_d + margin : index_x_u - margin, :] = seg
+                        else:
+                            seg = seg[margin : seg.shape[0] - margin, margin : seg.shape[1] - margin]
+                            prediction_true[index_y_d + margin : index_y_u - margin, index_x_d + margin : index_x_u - margin, :] = seg
+                prediction_true = prediction_true.astype(int)
+                prediction_true = cv2.resize(prediction_true, (self.img_org.shape[1], self.img_org.shape[0]), interpolation=cv2.INTER_NEAREST)
+                return prediction_true
+
+            else:
+
+                img=cv2.imread(self.image)
+                self.img_org = np.copy(img)
+
+                width=self.img_width
+                height=self.img_height
+
+                img=img/255.0
+                img=self.resize_image(img,self.img_height,self.img_width)
+                
+
+                label_p_pred=self.model.predict(
+                    img.reshape(1,img.shape[0],img.shape[1],img.shape[2]))
+
+                if self.task == 'enhancement':
+                    seg = label_p_pred[0, :, :, :]
+                    seg = seg * 255
+                elif self.task == 'segmentation' or self.task == 'binarization':
+                    seg = np.argmax(label_p_pred, axis=3)[0]
+                    seg = np.repeat(seg[:, :, np.newaxis], 3, axis=2)
+                    
+                prediction_true = seg.astype(int)
+
+                prediction_true = cv2.resize(prediction_true, (self.img_org.shape[1], self.img_org.shape[0]), interpolation=cv2.INTER_NEAREST)
+                return prediction_true
+
+
+
+    def run(self):
+        res=self.predict()
+        if (self.task == 'classification' or self.task == 'reading_order'):
+            pass
+        elif self.task == 'enhancement':
+            if self.save:
+                cv2.imwrite(self.save,res)
+        else:
+            img_seg_overlayed, only_layout  = self.visualize_model_output(res, self.img_org, self.task)
+            if self.save:
+                cv2.imwrite(self.save,img_seg_overlayed)
+            if self.save_layout:
+                cv2.imwrite(self.save_layout, only_layout)
+                
+        if self.ground_truth:
+            gt_img=cv2.imread(self.ground_truth)
+            self.IoU(gt_img[:,:,0],res[:,:,0])
+        
+@click.command()
+@click.option(
+    "--image",
+    "-i",
+    help="image filename",
+    type=click.Path(exists=True, dir_okay=False),
+)
+@click.option(
+    "--out",
+    "-o",
+    help="output directory where xml with detected reading order will be written.",
+    type=click.Path(exists=True, file_okay=False),
+)
+@click.option(
+    "--patches/--no-patches",
+    "-p/-nop",
+    is_flag=True,
+    help="if this parameter set to true, this tool will try to do inference in patches.",
+)
+@click.option(
+    "--save",
+    "-s",
+    help="save prediction as a png file in current folder.",
+)
+@click.option(
+    "--save_layout",
+    "-sl",
+    help="save layout prediction only as a png file in current folder.",
+)
+@click.option(
+    "--model",
+    "-m",
+    help="directory of models",
+    type=click.Path(exists=True, file_okay=False),
+    required=True,
+)
+@click.option(
+    "--ground_truth",
+    "-gt",
+    help="ground truth directory if you want to see the iou of prediction.",
+)
+@click.option(
+    "--xml_file",
+    "-xml",
+    help="xml file with layout coordinates that reading order detection will be implemented on. The result will be written in the same xml file.",
+)
+
+@click.option(
+    "--min_area",
+    "-min",
+    help="min area size of regions considered for reading order detection. The default value is zero and means that all text regions are considered for reading order.",
+)
+def main(image, model, patches, save, save_layout, ground_truth, xml_file, out, min_area):
+    with open(os.path.join(model,'config.json')) as f:
+        config_params_model = json.load(f)
+    task = config_params_model['task']
+    if (task != 'classification' and task != 'reading_order'):
+        if not save:
+            print("Error: You used one of segmentation or binarization task but not set -s, you need a filename to save visualized output with -s")
+            sys.exit(1)
+    x=sbb_predict(image, model, task, config_params_model, patches, save, save_layout, ground_truth, xml_file, out, min_area)
+    x.run()
+
+if __name__=="__main__":
+    main()
+
+    
+    
+    

models.py

@@ -1,12 +1,106 @@
+import tensorflow as tf
+from tensorflow import keras
 from tensorflow.keras.models import *
 from tensorflow.keras.layers import *
 from tensorflow.keras import layers
 from tensorflow.keras.regularizers import l2
 
+##mlp_head_units = [512, 256]#[2048, 1024]
+###projection_dim = 64
+##transformer_layers = 2#8
+##num_heads = 1#4
 resnet50_Weights_path = './pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5'
 IMAGE_ORDERING = 'channels_last'
 MERGE_AXIS = -1
 
+def mlp(x, hidden_units, dropout_rate):
+    for units in hidden_units:
+        x = layers.Dense(units, activation=tf.nn.gelu)(x)
+        x = layers.Dropout(dropout_rate)(x)
+    return x
+
+class Patches(layers.Layer):
+    def __init__(self, patch_size_x, patch_size_y):#__init__(self, **kwargs):#:__init__(self, patch_size):#__init__(self, **kwargs):
+        super(Patches, self).__init__()
+        self.patch_size_x = patch_size_x
+        self.patch_size_y = patch_size_y
+
+    def call(self, images):
+        #print(tf.shape(images)[1],'images')
+        #print(self.patch_size,'self.patch_size')
+        batch_size = tf.shape(images)[0]
+        patches = tf.image.extract_patches(
+            images=images,
+            sizes=[1, self.patch_size_y, self.patch_size_x, 1],
+            strides=[1, self.patch_size_y, self.patch_size_x, 1],
+            rates=[1, 1, 1, 1],
+            padding="VALID",
+        )
+        #patch_dims = patches.shape[-1]
+        patch_dims = tf.shape(patches)[-1]
+        patches = tf.reshape(patches, [batch_size, -1, patch_dims])
+        return patches
+    def get_config(self):
+
+        config = super().get_config().copy()
+        config.update({
+            'patch_size_x': self.patch_size_x,
+            'patch_size_y': self.patch_size_y,
+        })
+        return config
+
+class Patches_old(layers.Layer):
+    def __init__(self, patch_size):#__init__(self, **kwargs):#:__init__(self, patch_size):#__init__(self, **kwargs):
+        super(Patches, self).__init__()
+        self.patch_size = patch_size
+
+    def call(self, images):
+        #print(tf.shape(images)[1],'images')
+        #print(self.patch_size,'self.patch_size')
+        batch_size = tf.shape(images)[0]
+        patches = tf.image.extract_patches(
+            images=images,
+            sizes=[1, self.patch_size, self.patch_size, 1],
+            strides=[1, self.patch_size, self.patch_size, 1],
+            rates=[1, 1, 1, 1],
+            padding="VALID",
+        )
+        patch_dims = patches.shape[-1]
+        #print(patches.shape,patch_dims,'patch_dims')
+        patches = tf.reshape(patches, [batch_size, -1, patch_dims])
+        return patches
+    def get_config(self):
+
+        config = super().get_config().copy()
+        config.update({
+            'patch_size': self.patch_size,
+        })
+        return config
+
+    
+class PatchEncoder(layers.Layer):
+    def __init__(self, num_patches, projection_dim):
+        super(PatchEncoder, self).__init__()
+        self.num_patches = num_patches
+        self.projection = layers.Dense(units=projection_dim)
+        self.position_embedding = layers.Embedding(
+            input_dim=num_patches, output_dim=projection_dim
+        )
+
+    def call(self, patch):
+        positions = tf.range(start=0, limit=self.num_patches, delta=1)
+        encoded = self.projection(patch) + self.position_embedding(positions)
+        return encoded
+    def get_config(self):
+
+        config = super().get_config().copy()
+        config.update({
+            'num_patches': self.num_patches,
+            'projection': self.projection,
+            'position_embedding': self.position_embedding,
+        })
+        return config
+    
     
 def one_side_pad(x):
     x = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING)(x)
@@ -100,7 +194,7 @@ def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2))
     return x
 
 
-def resnet50_unet_light(n_classes, input_height=224, input_width=224, weight_decay=1e-6, pretraining=False):
+def resnet50_unet_light(n_classes, input_height=224, input_width=224, taks="segmentation", weight_decay=1e-6, pretraining=False):
     assert input_height % 32 == 0
     assert input_width % 32 == 0
 
@@ -191,14 +285,17 @@ def resnet50_unet_light(n_classes, input_height=224, input_width=224, weight_dec
     o = Activation('relu')(o)
 
     o = Conv2D(n_classes, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
-    o = (BatchNormalization(axis=bn_axis))(o)
+    if task == "segmentation":
-    o = (Activation('softmax'))(o)
+        o = (BatchNormalization(axis=bn_axis))(o)
+        o = (Activation('softmax'))(o)
+    else:
+        o = (Activation('sigmoid'))(o)
 
     model = Model(img_input, o)
     return model
 
 
-def resnet50_unet(n_classes, input_height=224, input_width=224, weight_decay=1e-6, pretraining=False):
+def resnet50_unet(n_classes, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False):
     assert input_height % 32 == 0
     assert input_width % 32 == 0
 
@@ -286,9 +383,374 @@ def resnet50_unet(n_classes, input_height=224, input_width=224, weight_decay=1e-
     o = Activation('relu')(o)
 
     o = Conv2D(n_classes, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(o)
-    o = (BatchNormalization(axis=bn_axis))(o)
+    if task == "segmentation":
-    o = (Activation('softmax'))(o)
+        o = (BatchNormalization(axis=bn_axis))(o)
+        o = (Activation('softmax'))(o)
+    else:
+        o = (Activation('sigmoid'))(o)
 
     model = Model(img_input, o)
 
     return model
+
+
+def vit_resnet50_unet(n_classes, patch_size_x, patch_size_y, num_patches, mlp_head_units=[128, 64], transformer_layers=8, num_heads =4, projection_dim = 64, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False):
+    inputs = layers.Input(shape=(input_height, input_width, 3))
+    
+    #transformer_units = [
+        #projection_dim * 2,
+        #projection_dim,
+    #]  # Size of the transformer layers
+    IMAGE_ORDERING = 'channels_last'
+    bn_axis=3
+
+    x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(inputs)
+    x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x)
+    f1 = x
+
+    x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x)
+    x = Activation('relu')(x)
+    x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x)
+    
+    x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='b')
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='c')
+    f2 = one_side_pad(x)
+
+    x = conv_block(x, 3, [128, 128, 512], stage=3, block='a')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='b')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='c')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='d')
+    f3 = x 
+
+    x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')
+    f4 = x 
+
+    x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')
+    f5 = x 
+    
+    if pretraining:
+        model = Model(inputs, x).load_weights(resnet50_Weights_path)
+
+    #num_patches = x.shape[1]*x.shape[2]
+    
+    #patch_size_y = input_height / x.shape[1]
+    #patch_size_x = input_width / x.shape[2]
+    #patch_size = patch_size_x * patch_size_y
+    patches = Patches(patch_size_x, patch_size_y)(x)
+    # Encode patches.
+    encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)
+    
+    for _ in range(transformer_layers):
+        # Layer normalization 1.
+        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
+        # Create a multi-head attention layer.
+        attention_output = layers.MultiHeadAttention(
+            num_heads=num_heads, key_dim=projection_dim, dropout=0.1
+        )(x1, x1)
+        # Skip connection 1.
+        x2 = layers.Add()([attention_output, encoded_patches])
+        # Layer normalization 2.
+        x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
+        # MLP.
+        x3 = mlp(x3, hidden_units=mlp_head_units, dropout_rate=0.1)
+        # Skip connection 2.
+        encoded_patches = layers.Add()([x3, x2])
+    
+    encoded_patches = tf.reshape(encoded_patches, [-1, x.shape[1], x.shape[2] , int( projection_dim / (patch_size_x * patch_size_y) )])
+
+    v1024_2048 = Conv2D( 1024 , (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(encoded_patches)
+    v1024_2048 = (BatchNormalization(axis=bn_axis))(v1024_2048)
+    v1024_2048 = Activation('relu')(v1024_2048)
+    
+    o = (UpSampling2D( (2, 2), data_format=IMAGE_ORDERING))(v1024_2048)
+    o = (concatenate([o, f4],axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(512, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o ,f3], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(256, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, f2], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(128, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, f1], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(64, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, inputs],axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(32, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = Conv2D(n_classes, (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(o)
+    if task == "segmentation":
+        o = (BatchNormalization(axis=bn_axis))(o)
+        o = (Activation('softmax'))(o)
+    else:
+        o = (Activation('sigmoid'))(o)
+
+    model = Model(inputs=inputs, outputs=o)
+    
+    return model
+
+def vit_resnet50_unet_transformer_before_cnn(n_classes, patch_size_x, patch_size_y, num_patches, mlp_head_units=[128, 64], transformer_layers=8, num_heads =4, projection_dim = 64, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False):
+    inputs = layers.Input(shape=(input_height, input_width, 3))
+    
+    ##transformer_units = [
+        ##projection_dim * 2,
+        ##projection_dim,
+    ##]  # Size of the transformer layers
+    IMAGE_ORDERING = 'channels_last'
+    bn_axis=3
+    
+    patches = Patches(patch_size_x, patch_size_y)(inputs)
+    # Encode patches.
+    encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)
+    
+    for _ in range(transformer_layers):
+        # Layer normalization 1.
+        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
+        # Create a multi-head attention layer.
+        attention_output = layers.MultiHeadAttention(
+            num_heads=num_heads, key_dim=projection_dim, dropout=0.1
+        )(x1, x1)
+        # Skip connection 1.
+        x2 = layers.Add()([attention_output, encoded_patches])
+        # Layer normalization 2.
+        x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
+        # MLP.
+        x3 = mlp(x3, hidden_units=mlp_head_units, dropout_rate=0.1)
+        # Skip connection 2.
+        encoded_patches = layers.Add()([x3, x2])
+    
+    encoded_patches = tf.reshape(encoded_patches, [-1, input_height, input_width , int( projection_dim / (patch_size_x * patch_size_y) )])
+    
+    encoded_patches = Conv2D(3, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay), name='convinput')(encoded_patches)
+    
+    x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(encoded_patches)
+    x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x)
+    f1 = x
+
+    x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x)
+    x = Activation('relu')(x)
+    x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x)
+    
+    x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='b')
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='c')
+    f2 = one_side_pad(x)
+
+    x = conv_block(x, 3, [128, 128, 512], stage=3, block='a')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='b')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='c')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='d')
+    f3 = x 
+
+    x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')
+    f4 = x 
+
+    x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')
+    f5 = x 
+    
+    if pretraining:
+        model = Model(encoded_patches, x).load_weights(resnet50_Weights_path)
+
+    v1024_2048 = Conv2D( 1024 , (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(x)
+    v1024_2048 = (BatchNormalization(axis=bn_axis))(v1024_2048)
+    v1024_2048 = Activation('relu')(v1024_2048)
+    
+    o = (UpSampling2D( (2, 2), data_format=IMAGE_ORDERING))(v1024_2048)
+    o = (concatenate([o, f4],axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(512, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o ,f3], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(256, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, f2], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(128, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, f1], axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(64, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = (UpSampling2D((2, 2), data_format=IMAGE_ORDERING))(o)
+    o = (concatenate([o, inputs],axis=MERGE_AXIS))
+    o = (ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING))(o)
+    o = (Conv2D(32, (3, 3), padding='valid', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay)))(o)
+    o = (BatchNormalization(axis=bn_axis))(o)
+    o = Activation('relu')(o)
+    
+    o = Conv2D(n_classes, (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(o)
+    if task == "segmentation":
+        o = (BatchNormalization(axis=bn_axis))(o)
+        o = (Activation('softmax'))(o)
+    else:
+        o = (Activation('sigmoid'))(o)
+
+    model = Model(inputs=inputs, outputs=o)
+    
+    return model
+
+def resnet50_classifier(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False):
+    include_top=True
+    assert input_height%32 == 0
+    assert input_width%32 == 0
+
+    
+    img_input = Input(shape=(input_height,input_width , 3 ))
+
+    if IMAGE_ORDERING == 'channels_last':
+        bn_axis = 3
+    else:
+        bn_axis = 1
+
+    x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input)
+    x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x)
+    f1 = x
+
+    x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x)
+    x = Activation('relu')(x)
+    x = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x)
+    
+
+    x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='b')
+    x = identity_block(x, 3, [64, 64, 256], stage=2, block='c')
+    f2 = one_side_pad(x )
+
+
+    x = conv_block(x, 3, [128, 128, 512], stage=3, block='a')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='b')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='c')
+    x = identity_block(x, 3, [128, 128, 512], stage=3, block='d')
+    f3 = x 
+
+    x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')
+    x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')
+    f4 = x 
+
+    x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')
+    x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')
+    f5 = x 
+
+    if pretraining:
+        Model(img_input, x).load_weights(resnet50_Weights_path)
+
+    x = AveragePooling2D((7, 7), name='avg_pool')(x)
+    x = Flatten()(x)
+    
+    ##
+    x = Dense(256, activation='relu', name='fc512')(x)
+    x=Dropout(0.2)(x)
+    ##
+    x = Dense(n_classes, activation='softmax', name='fc1000')(x)
+    model = Model(img_input, x)
+    
+    
+
+
+    return model
+
+def machine_based_reading_order_model(n_classes,input_height=224,input_width=224,weight_decay=1e-6,pretraining=False):
+    assert input_height%32 == 0
+    assert input_width%32 == 0
+
+    img_input = Input(shape=(input_height,input_width , 3 ))
+
+    if IMAGE_ORDERING == 'channels_last':
+        bn_axis = 3
+    else:
+        bn_axis = 1
+
+    x1 = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input)
+    x1 = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x1)
+
+    x1 = BatchNormalization(axis=bn_axis, name='bn_conv1')(x1)
+    x1 = Activation('relu')(x1)
+    x1 = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x1)
+    
+    x1 = conv_block(x1, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
+    x1 = identity_block(x1, 3, [64, 64, 256], stage=2, block='b')
+    x1 = identity_block(x1, 3, [64, 64, 256], stage=2, block='c')
+
+    x1 = conv_block(x1, 3, [128, 128, 512], stage=3, block='a')
+    x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='b')
+    x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='c')
+    x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='d')
+
+    x1 = conv_block(x1, 3, [256, 256, 1024], stage=4, block='a')
+    x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='b')
+    x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='c')
+    x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='d')
+    x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='e')
+    x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='f')
+
+    x1 = conv_block(x1, 3, [512, 512, 2048], stage=5, block='a')
+    x1 = identity_block(x1, 3, [512, 512, 2048], stage=5, block='b')
+    x1 = identity_block(x1, 3, [512, 512, 2048], stage=5, block='c')
+    
+    if pretraining:
+        Model(img_input , x1).load_weights(resnet50_Weights_path)
+    
+    x1 = AveragePooling2D((7, 7), name='avg_pool1')(x1)
+    flattened = Flatten()(x1)
+    
+    o = Dense(256, activation='relu', name='fc512')(flattened)
+    o=Dropout(0.2)(o)
+    
+    o = Dense(256, activation='relu', name='fc512a')(o)
+    o=Dropout(0.2)(o)
+
+    o = Dense(n_classes, activation='sigmoid', name='fc1000')(o)
+    model = Model(img_input , o)
+
+    return model

requirements.txt

@@ -6,3 +6,6 @@ tqdm
 imutils
 numpy
 scipy
+scikit-learn
+shapely
+click

scales_enhancement.json

@@ -0,0 +1,3 @@
+{
+    "scales" : [ 0.5, 0.55, 0.6, 0.65, 0.7, 0.75,  0.8, 0.85,  0.9]
+}

train.md

@@ -0,0 +1,576 @@
+# Documentation for Training Models
+
+This repository assists users in preparing training datasets, training models, and performing inference with trained models. We cover various use cases including pixel-wise segmentation, image classification, image enhancement, and machine-based reading order. For each use case, we provide guidance on how to generate the corresponding training dataset.
+All these use cases are now utilized in the Eynollah workflow. 
+As mentioned, the following three tasks can be accomplished using this repository:
+
+* Generate training dataset
+* Train a model
+* Inference with the trained model
+
+## Generate training dataset
+The script generate_gt_for_training.py is used for generating training datasets. As the results of the following command demonstrate, the dataset generator provides three different commands:
+
+`python generate_gt_for_training.py --help`
+
+
+These three commands are:
+
+* image-enhancement
+* machine-based-reading-order
+* pagexml2label
+
+
+### image-enhancement
+
+Generating a training dataset for image enhancement is quite straightforward. All that is needed is a set of high-resolution images. The training dataset can then be generated using the following command:
+
+`python generate_gt_for_training.py image-enhancement -dis "dir of high resolution images" -dois "dir where degraded images will be written" -dols "dir where the corresponding high resolution image will be written as label" -scs "degrading scales json file"`
+
+The scales JSON file is a dictionary with a key named 'scales' and values representing scales smaller than 1. Images are downscaled based on these scales and then upscaled again to their original size. This process causes the images to lose resolution at different scales. The degraded images are used as input images, and the original high-resolution images serve as labels. The enhancement model can be trained with this generated dataset. The scales JSON file looks like this:
+
+```yaml
+{
+    "scales": [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9]
+}
+```
+
+### machine-based-reading-order
+
+For machine-based reading order, we aim to determine the reading priority between two sets of text regions. The model's input is a three-channel image: the first and last channels contain information about each of the two text regions, while the middle channel encodes prominent layout elements necessary for reading order, such as separators and headers. To generate the training dataset, our script requires a page XML file that specifies the image layout with the correct reading order.
+
+For output images, it is necessary to specify the width and height. Additionally, a minimum text region size can be set to filter out regions smaller than this minimum size. This minimum size is defined as the ratio of the text region area to the image area, with a default value of zero. To run the dataset generator, use the following command:
+
+
+`python generate_gt_for_training.py machine-based-reading-order -dx "dir of GT xml files" -domi "dir where output images will be written" -docl "dir where the labels will be written" -ih "height" -iw "width" -min "min area ratio"`
+
+### pagexml2label
+
+pagexml2label is designed to generate labels from GT page XML files for various pixel-wise segmentation use cases, including 'layout,' 'textline,' 'printspace,' 'glyph,' and 'word' segmentation.
+To train a pixel-wise segmentation model, we require images along with their corresponding labels. Our training script expects a PNG image where each pixel corresponds to a label, represented by an integer. The background is always labeled as zero, while other elements are assigned different integers. For instance, if we have ground truth data with four elements including the background, the classes would be labeled as 0, 1, 2, and 3 respectively.
+
+In binary segmentation scenarios such as textline or page extraction, the background is encoded as 0, and the desired element is automatically encoded as 1 in the PNG label.
+
+To specify the desired use case and the elements to be extracted in the PNG labels, a custom JSON file can be passed. For example, in the case of 'textline' detection, the JSON file would resemble this:
+
+```yaml
+{
+"use_case": "textline"
+}
+```
+
+In the case of layout segmentation a possible custom config json file can be like this:
+
+```yaml
+{
+"use_case": "layout",
+"textregions":{"rest_as_paragraph":1 , "drop-capital": 1, "header":2, "heading":2, "marginalia":3},
+"imageregion":4,
+"separatorregion":5,
+"graphicregions" :{"rest_as_decoration":6 ,"stamp":7}
+}
+```
+
+A possible custom config json file  for layout segmentation where the "printspace" is wished to be a class:
+
+```yaml
+{
+"use_case": "layout",
+"textregions":{"rest_as_paragraph":1 , "drop-capital": 1, "header":2, "heading":2, "marginalia":3},
+"imageregion":4,
+"separatorregion":5,
+"graphicregions" :{"rest_as_decoration":6 ,"stamp":7}
+"printspace_as_class_in_layout" : 8
+}
+```
+For the layout use case, it is beneficial to first understand the structure of the page XML file and its elements. In a given image, the annotations of elements are recorded in a page XML file, including their contours and classes. For an image document, the known regions are 'textregion', 'separatorregion', 'imageregion', 'graphicregion', 'noiseregion', and 'tableregion'.
+
+Text regions and graphic regions also have their own specific types. The known types for us for text regions are 'paragraph', 'header', 'heading', 'marginalia', 'drop-capital', 'footnote', 'footnote-continued', 'signature-mark', 'page-number', and 'catch-word'. The known types for graphic regions are 'handwritten-annotation', 'decoration', 'stamp', and 'signature'.
+Since we don't know all types of text and graphic regions, unknown cases can arise. To handle these, we have defined two additional types: "rest_as_paragraph" and "rest_as_decoration" to ensure that no unknown types are missed. This way, users can extract all known types from the labels and be confident that no unknown types are overlooked.
+
+In the custom JSON file shown above, "header" and "heading" are extracted as the same class, while "marginalia" is shown as a different class. All other text region types, including "drop-capital," are grouped into the same class. For the graphic region, "stamp" has its own class, while all other types are classified together. "Image region" and "separator region" are also present in the label. However, other regions like "noise region" and "table region" will not be included in the label PNG file, even if they have information in the page XML files, as we chose not to include them.
+
+`python generate_gt_for_training.py pagexml2label -dx "dir of GT xml files" -do "dir where output label png files will be written" -cfg "custom config json file" -to "output type which has 2d and 3d. 2d is used for training and 3d is just to visualise the labels" "`
+
+We have also defined an artificial class that can be added to the boundary of text region types or text lines. This key is called "artificial_class_on_boundary." If users want to apply this to certain text regions in the layout use case, the example JSON config file should look like this:
+
+```yaml
+{
+    "use_case": "layout",
+    "textregions": {
+        "paragraph": 1,
+        "drop-capital": 1,
+        "header": 2,
+        "heading": 2,
+        "marginalia": 3
+    },
+    "imageregion": 4,
+    "separatorregion": 5,
+    "graphicregions": {
+        "rest_as_decoration": 6
+    },
+    "artificial_class_on_boundary": ["paragraph", "header", "heading", "marginalia"],
+    "artificial_class_label": 7
+}
+```
+
+This implies that the artificial class label, denoted by 7, will be present on PNG files and will only be added to the elements labeled as "paragraph," "header," "heading," and "marginalia."
+
+For "textline," "word," and "glyph," the artificial class on the boundaries will be activated only if the "artificial_class_label" key is specified in the config file. Its value should be set as 2 since these elements represent binary cases. For example, if the background and textline are denoted as 0 and 1 respectively, then the artificial class should be assigned the value 2. The example JSON config file should look like this for "textline" use case:
+
+```yaml
+{
+    "use_case": "textline",
+    "artificial_class_label": 2
+}
+```
+
+If the coordinates of "PrintSpace" or "Border" are present in the page XML ground truth files, and the user wishes to crop only the print space area, this can be achieved by activating the "-ps" argument. However, it should be noted that in this scenario, since cropping will be applied to the label files, the directory of the original images must be provided to ensure that they are cropped in sync with the labels. This ensures that the correct images and labels required for training are obtained. The command should resemble the following:
+
+`python generate_gt_for_training.py pagexml2label -dx "dir of GT xml files" -do "dir where output label png files will be written" -cfg "custom config json file" -to "output type which has 2d and 3d. 2d is used for training and 3d is just to visualise the labels" -ps -di "dir where the org images are located" -doi "dir where the cropped output images will be written" `
+
+## Train a model
+### classification
+
+For the classification use case, we haven't provided a ground truth generator, as it's unnecessary. For classification, all we require is a training directory with subdirectories, each containing images of its respective classes. We need separate directories for training and evaluation, and the class names (subdirectories) must be consistent across both directories. Additionally, the class names should be specified in the config JSON file, as shown in the following example. If, for instance, we aim to classify "apple" and "orange," with a total of 2 classes, the "classification_classes_name" key in the config file should appear as follows:
+
+```yaml
+{
+    "backbone_type" : "nontransformer",
+    "task": "classification",
+    "n_classes" : 2,
+    "n_epochs" : 10,
+    "input_height" : 448,
+    "input_width" : 448,
+    "weight_decay" : 1e-6,
+    "n_batch" : 4,
+    "learning_rate": 1e-4,
+    "f1_threshold_classification": 0.8,
+    "pretraining" : true,
+    "classification_classes_name" : {"0":"apple",  "1":"orange"},
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+The "dir_train" should be like this:
+
+``` 
+.
+└── train             # train directory
+   ├── apple          # directory of images for apple class
+   └── orange         # directory of images for orange class
+```
+
+And the "dir_eval" the same structure as train directory:
+
+``` 
+.
+└── eval              # evaluation directory
+   ├── apple          # directory of images for apple class
+   └── orange         # directory of images for orange class
+
+```
+
+The classification model can be trained using the following command line:
+
+`python train.py with config_classification.json`
+
+
+As evident in the example JSON file above, for classification, we utilize a "f1_threshold_classification" parameter. This parameter is employed to gather all models with an evaluation f1 score surpassing this threshold. Subsequently, an ensemble of these model weights is executed, and a model is saved in the output directory as "model_ens_avg". Additionally, the weight of the best model based on the evaluation f1 score is saved as "model_best".
+
+### reading order
+An example config json file for machine based reading order should be like this:
+
+```yaml
+{
+    "backbone_type" : "nontransformer",
+    "task": "reading_order",
+    "n_classes" : 1,
+    "n_epochs" : 5,
+    "input_height" : 672,
+    "input_width" : 448,
+    "weight_decay" : 1e-6,
+    "n_batch" : 4,
+    "learning_rate": 1e-4,
+    "pretraining" : true,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+The "dir_train" should be like this:
+
+```
+.
+└── train             # train directory
+   ├── images          # directory of images
+   └── labels         # directory of labels
+```
+
+And the "dir_eval" the same structure as train directory:
+
+```
+.
+└── eval             # evaluation directory
+   ├── images          # directory of images
+   └── labels         # directory of labels
+```
+
+The classification model can be trained like the classification case command line.
+
+### Segmentation (Textline, Binarization, Page extraction and layout) and enhancement
+
+#### Parameter configuration for segmentation or enhancement usecases
+
+The following parameter configuration can be applied to all segmentation use cases and enhancements. The augmentation, its sub-parameters, and continued training are defined only for segmentation use cases and enhancements, not for classification and machine-based reading order, as you can see in their example config files.
+
+* backbone_type: For segmentation tasks (such as text line, binarization, and layout detection) and enhancement, we offer two backbone options: a "nontransformer" and a "transformer" backbone. For the "transformer" backbone, we first apply a CNN followed by a transformer. In contrast, the "nontransformer" backbone utilizes only a CNN ResNet-50.
+* task : The task parameter can have values such as "segmentation", "enhancement", "classification", and "reading_order".
+* patches: If you want to break input images into smaller patches (input size of the model) you need to set this parameter to ``true``. In the case that the model should see the image once, like page extraction, patches should be set to ``false``.
+* n_batch: Number of batches at each iteration.
+* n_classes: Number of classes. In the case of binary classification this should be 2. In the case of reading_order it should set to 1. And for the case of layout detection just the unique number of classes should be given.
+* n_epochs: Number of epochs.
+* input_height: This indicates the height of model's input.
+* input_width: This indicates the width of model's input.
+* weight_decay: Weight decay of l2 regularization of model layers.
+* pretraining: Set to ``true`` to load pretrained weights of ResNet50 encoder. The downloaded weights should be saved in a folder named "pretrained_model" in the same directory of "train.py" script.
+* augmentation: If you want to apply any kind of augmentation this parameter should first set to ``true``.
+* flip_aug: If ``true``, different types of filp will be applied on image. Type of flips is given with "flip_index" parameter.
+* blur_aug: If ``true``, different types of blurring will be applied on image. Type of blurrings is given with "blur_k" parameter.
+* scaling: If ``true``, scaling will be applied on image. Scale of scaling is given with "scales" parameter.
+* degrading: If ``true``, degrading will be applied to the image. The amount of degrading is defined with "degrade_scales" parameter.
+* brightening: If ``true``, brightening will be applied to the image. The amount of brightening is defined with "brightness" parameter.
+* rotation_not_90: If ``true``, rotation (not 90 degree) will be applied on image. Rotation angles are given with "thetha" parameter.
+* rotation: If ``true``, 90 degree rotation will be applied on image.
+* binarization: If ``true``,Otsu thresholding will be applied to augment the input data with binarized images.
+* scaling_bluring: If ``true``, combination of scaling and blurring will be applied on image.
+* scaling_binarization: If ``true``, combination of scaling and binarization will be applied on image.
+* scaling_flip: If ``true``, combination of scaling and flip will be applied on image.
+* flip_index: Type of flips.
+* blur_k: Type of blurrings.
+* scales: Scales of scaling.
+* brightness: The amount of brightenings.
+* thetha: Rotation angles.
+* degrade_scales: The amount of degradings.
+* continue_training: If ``true``, it means that you have already trained a model and you would like to continue the training. So it is needed to provide the dir of trained model with "dir_of_start_model" and index for naming the models. For example if you have already trained for 3 epochs then your last index is 2 and if you want to continue from model_1.h5, you can set ``index_start`` to 3 to start naming model with index 3. 
+* weighted_loss: If ``true``, this means that you want to apply weighted categorical_crossentropy as loss fucntion. Be carefull if you set to ``true``the parameter "is_loss_soft_dice" should be ``false``
+* data_is_provided: If you have already provided the input data you can set this to ``true``. Be sure that the train and eval data are in "dir_output". Since when once we provide training data we resize and augment them and then we write them in sub-directories train and eval in "dir_output". 
+* dir_train: This is the directory of "images" and "labels" (dir_train should include two subdirectories with names of images and labels ) for raw images and labels. Namely they are not prepared (not resized and not augmented) yet for training the model. When we run this tool these raw data will be transformed to suitable size needed for the model and they will be written in "dir_output" in train and eval directories. Each of train and eval include "images" and "labels" sub-directories.
+* index_start: Starting index for saved models in the case that "continue_training"  is ``true``.
+* dir_of_start_model: Directory containing pretrained model to continue training the model in the case that "continue_training"  is ``true``.
+* transformer_num_patches_xy: Number of patches for vision transformer in x and y direction respectively.
+* transformer_patchsize_x: Patch size of vision transformer patches in x direction.
+* transformer_patchsize_y: Patch size of vision transformer patches in y direction.
+* transformer_projection_dim: Transformer projection dimension. Default value is 64.
+* transformer_mlp_head_units: Transformer Multilayer Perceptron (MLP) head units. Default value is [128, 64].
+* transformer_layers: transformer layers. Default value is 8.
+* transformer_num_heads: Transformer number of heads. Default value is 4.
+* transformer_cnn_first: We have two types of vision transformers. In one type, a CNN is applied first, followed by a transformer. In the other type, this order is reversed. If transformer_cnn_first is true, it means the CNN will be applied before the transformer. Default value is true.
+
+In the case of segmentation and enhancement the train and evaluation directory should be as following.
+
+The "dir_train" should be like this:
+
+```
+.
+└── train             # train directory
+   ├── images          # directory of images
+   └── labels         # directory of labels
+```
+
+And the "dir_eval" the same structure as train directory:
+
+```
+.
+└── eval             # evaluation directory
+   ├── images          # directory of images
+   └── labels         # directory of labels
+```
+
+After configuring the JSON file for segmentation or enhancement, training can be initiated by running the following command, similar to the process for classification and reading order:
+
+`python train.py with config_classification.json`
+
+#### Binarization
+
+An example config json file for binarization can be like this:
+
+```yaml
+{
+    "backbone_type" : "transformer",
+    "task": "binarization",
+    "n_classes" : 2,
+    "n_epochs" : 4,
+    "input_height" : 224,
+    "input_width" : 672,
+    "weight_decay" : 1e-6,
+    "n_batch" : 1,
+    "learning_rate": 1e-4,
+    "patches" : true,
+    "pretraining" : true,
+    "augmentation" : true,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "transformer_num_patches_xy": [7, 7],
+    "transformer_patchsize_x": 3,
+    "transformer_patchsize_y": 1,
+    "transformer_projection_dim": 192,
+    "transformer_mlp_head_units": [128, 64],
+    "transformer_layers": 8,
+    "transformer_num_heads": 4,
+    "transformer_cnn_first": true,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ", 
+    "weighted_loss": false,
+    "is_loss_soft_dice": false,
+    "data_is_provided": false,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+#### Textline
+
+```yaml
+{
+    "backbone_type" : "nontransformer",
+    "task": "segmentation",
+    "n_classes" : 2,
+    "n_epochs" : 4,
+    "input_height" : 448,
+    "input_width" : 224,
+    "weight_decay" : 1e-6,
+    "n_batch" : 1,
+    "learning_rate": 1e-4,
+    "patches" : true,
+    "pretraining" : true,
+    "augmentation" : true,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ", 
+    "weighted_loss": false,
+    "is_loss_soft_dice": false,
+    "data_is_provided": false,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+#### Enhancement
+
+```yaml
+{
+    "backbone_type" : "nontransformer",
+    "task": "enhancement",
+    "n_classes" : 3,
+    "n_epochs" : 4,
+    "input_height" : 448,
+    "input_width" : 224,
+    "weight_decay" : 1e-6,
+    "n_batch" : 4,
+    "learning_rate": 1e-4,
+    "patches" : true,
+    "pretraining" : true,
+    "augmentation" : true,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ", 
+    "weighted_loss": false,
+    "is_loss_soft_dice": false,
+    "data_is_provided": false,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+It's important to mention that the value of n_classes for enhancement should be 3, as the model's output is a 3-channel image.
+
+#### Page extraction
+
+```yaml
+{
+    "backbone_type" : "nontransformer",
+    "task": "segmentation",
+    "n_classes" : 2,
+    "n_epochs" : 4,
+    "input_height" : 448,
+    "input_width" : 224,
+    "weight_decay" : 1e-6,
+    "n_batch" : 1,
+    "learning_rate": 1e-4,
+    "patches" : false,
+    "pretraining" : true,
+    "augmentation" : false,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ", 
+    "weighted_loss": false,
+    "is_loss_soft_dice": false,
+    "data_is_provided": false,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+
+For page segmentation (or print space or border segmentation), the model needs to view the input image in its entirety, hence the patches parameter should be set to false.
+
+#### layout segmentation
+
+An example config json file for layout segmentation with 5 classes (including background) can be like this:
+
+```yaml
+{
+    "backbone_type" : "transformer",
+    "task": "segmentation",
+    "n_classes" : 5,
+    "n_epochs" : 4,
+    "input_height" : 448,
+    "input_width" : 224,
+    "weight_decay" : 1e-6,
+    "n_batch" : 1,
+    "learning_rate": 1e-4,
+    "patches" : true,
+    "pretraining" : true,
+    "augmentation" : true,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": false,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "transformer_num_patches_xy": [7, 14],
+    "transformer_patchsize_x": 1,
+    "transformer_patchsize_y": 1,
+    "transformer_projection_dim": 64,
+    "transformer_mlp_head_units": [128, 64],
+    "transformer_layers": 8,
+    "transformer_num_heads": 4,
+    "transformer_cnn_first": true,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ", 
+    "weighted_loss": false,
+    "is_loss_soft_dice": false,
+    "data_is_provided": false,
+    "dir_train": "./train",
+    "dir_eval": "./eval",
+    "dir_output": "./output"
+}
+```
+## Inference with the trained model
+### classification
+
+For conducting inference with a trained model, you simply need to execute the following command line, specifying the directory of the model and the image on which to perform inference:
+
+
+`python inference.py -m "model dir" -i "image" `
+
+This will straightforwardly return the class of the image.
+
+### machine based reading order
+
+
+To infer the reading order using an reading order model, we need a page XML file containing layout information but without the reading order. We simply need to provide the model directory, the XML file, and the output directory. The new XML file with the added reading order will be written to the output directory with the same name. We need to run:
+
+`python inference.py -m "model dir" -xml "page xml file" -o "output dir to write new xml with reading order" `
+
+
+### Segmentation (Textline, Binarization, Page extraction and layout) and enhancement
+
+For conducting inference with a trained model for segmentation and enhancement you need to run the following command line:
+
+
+`python inference.py -m "model dir" -i "image" -p -s "output image" `
+
+
+Note that in the case of page extraction the -p flag is not needed.
+
+For segmentation or binarization tasks, if a ground truth (GT) label is available, the IOU evaluation metric can be calculated for the output. To do this, you need to provide the GT label using the argument -gt.
+
+
+

train.py

@@ -1,15 +1,33 @@
 import os
 import sys
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 
 import tensorflow as tf
 from tensorflow.compat.v1.keras.backend import set_session
 import warnings
 from tensorflow.keras.optimizers import *
+from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping, Callback, ModelCheckpoint
 from sacred import Experiment
 from models import *
 from utils import *
 from metrics import *
 from tensorflow.keras.models import load_model
 from tqdm import tqdm
+import json
+from sklearn.metrics import f1_score
+
+
+def get_warmup_schedule(start_lr, target_lr, warmup_epochs, steps_per_epoch):
+    initial_learning_rate = start_lr
+    target_learning_rate = target_lr
+    warmup_steps = warmup_epochs * steps_per_epoch
+    
+    lr_schedule = tf.keras.optimizers.schedules.LinearSchedule(
+        initial_learning_rate=initial_learning_rate,
+        final_learning_rate=target_learning_rate,
+        total_steps=warmup_steps
+    )
+    
+    return lr_schedule
 
 
 def configuration():
@@ -42,172 +60,430 @@ def config_params():
     learning_rate = 1e-4  # Set the learning rate.
     patches = False  # Divides input image into smaller patches (input size of the model) when set to true. For the model to see the full image, like page extraction, set this to false.
     augmentation = False  # To apply any kind of augmentation, this parameter must be set to true.
-    flip_aug = False  # If true, different types of flipping will be applied to the image. Types of flips are defined with "flip_index" in train.py.
+    flip_aug = False  # If true, different types of flipping will be applied to the image. Types of flips are defined with "flip_index" in config_params.json.
-    blur_aug = False  # If true, different types of blurring will be applied to the image. Types of blur are defined with "blur_k" in train.py.
+    blur_aug = False  # If true, different types of blurring will be applied to the image. Types of blur are defined with "blur_k" in config_params.json.
-    scaling = False  # If true, scaling will be applied to the image. The amount of scaling is defined with "scales" in train.py.
+    padding_white = False # If true, white padding will be applied to the image.
+    padding_black = False # If true, black padding will be applied to the image.
+    scaling = False  # If true, scaling will be applied to the image. The amount of scaling is defined with "scales" in config_params.json.
+    degrading = False  # If true, degrading will be applied to the image. The amount of degrading is defined with "degrade_scales" in config_params.json.
+    brightening = False  # If true, brightening will be applied to the image. The amount of brightening is defined with "brightness" in config_params.json.
     binarization = False  # If true, Otsu thresholding will be applied to augment the input with binarized images.
+    adding_rgb_background = False
+    adding_rgb_foreground = False
+    add_red_textlines = False
+    channels_shuffling = False
     dir_train = None  # Directory of training dataset with subdirectories having the names "images" and "labels".
     dir_eval = None  # Directory of validation dataset with subdirectories having the names "images" and "labels".
     dir_output = None  # Directory where the output model will be saved.
     pretraining = False  # Set to true to load pretrained weights of ResNet50 encoder.
     scaling_bluring = False  # If true, a combination of scaling and blurring will be applied to the image.
     scaling_binarization = False  # If true, a combination of scaling and binarization will be applied to the image.
+    rotation = False # If true, a 90 degree rotation will be implemeneted.
+    rotation_not_90 = False # If true rotation based on provided angles with thetha will be implemeneted.
+    scaling_brightness = False  # If true, a combination of scaling and brightening will be applied to the image.
     scaling_flip = False  # If true, a combination of scaling and flipping will be applied to the image.
-    thetha = [10, -10]  # Rotate image by these angles for augmentation.
+    thetha = None  # Rotate image by these angles for augmentation.
-    blur_k = ['blur', 'gauss', 'median']  # Blur image for augmentation.
+    shuffle_indexes = None
-    scales = [0.5, 2]  # Scale patches for augmentation.
+    blur_k = None  # Blur image for augmentation.
-    flip_index = [0, 1, -1]  # Flip image for augmentation.
+    scales = None  # Scale patches for augmentation.
+    degrade_scales = None  # Degrade image for augmentation.
+    brightness = None #  Brighten image for augmentation.
+    flip_index = None  #  Flip image for augmentation.
     continue_training = False  # Set to true if you would like to continue training an already trained a model.
-    index_start = 0  # Index of model to continue training from. E.g. if you trained for 3 epochs and last index is 2, to continue from model_1.h5, set "index_start" to 3 to start naming model with index 3.
+    transformer_patchsize_x = None  # Patch size of vision transformer patches in x direction.
+    transformer_patchsize_y = None # Patch size of vision transformer patches in y direction.
+    transformer_num_patches_xy = None  # Number of patches for vision transformer in x and y direction respectively.
+    transformer_projection_dim = 64 # Transformer projection dimension. Default value is 64.
+    transformer_mlp_head_units = [128, 64] # Transformer Multilayer Perceptron (MLP) head units. Default value is [128, 64]
+    transformer_layers = 8 # transformer layers. Default value is 8.
+    transformer_num_heads = 4 # Transformer number of heads. Default value is 4.
+    transformer_cnn_first = True # We have two types of vision transformers. In one type, a CNN is applied first, followed by a transformer. In the other type, this order is reversed. If transformer_cnn_first is true, it means the CNN will be applied before the transformer. Default value is true.
+    index_start = 0  #  Index of model to continue training from. E.g. if you trained for 3 epochs and last index is 2, to continue from model_1.h5, set "index_start" to 3 to start naming model with index 3.
     dir_of_start_model = ''  # Directory containing pretrained encoder to continue training the model.
     is_loss_soft_dice = False  # Use soft dice as loss function. When set to true, "weighted_loss" must be false.
     weighted_loss = False  # Use weighted categorical cross entropy as loss fucntion. When set to true, "is_loss_soft_dice" must be false.
     data_is_provided = False  # Only set this to true when you have already provided the input data and the train and eval data are in "dir_output".
+    task = "segmentation" # This parameter defines task of model which can be segmentation, enhancement or classification.
+    f1_threshold_classification = None # This threshold is used to consider models with an evaluation f1 scores bigger than it. The selected model weights undergo a weights ensembling. And avreage ensembled model will be written to output.
+    classification_classes_name = None # Dictionary of classification classes names.
+    backbone_type = None # As backbone we have 2 types of backbones. A vision transformer alongside a CNN and we call it "transformer" and only CNN called "nontransformer"
+    
+    dir_img_bin = None
+    number_of_backgrounds_per_image = 1
+    dir_rgb_backgrounds = None
+    dir_rgb_foregrounds = None
+    reduce_lr_enabled = False  # Whether to use ReduceLROnPlateau callback
+    reduce_lr_monitor = 'val_loss'  # Metric to monitor for reducing learning rate
+    reduce_lr_factor = 0.5  # Factor to reduce learning rate by
+    reduce_lr_patience = 3  # Number of epochs to wait before reducing learning rate
+    reduce_lr_min_lr = 1e-6  # Minimum learning rate
+    reduce_lr_min_delta = 1e-5  # Minimum change in monitored value to be considered as improvement
+    early_stopping_enabled = False  # Whether to use EarlyStopping callback
+    early_stopping_monitor = 'val_loss'  # Metric to monitor for early stopping
+    early_stopping_patience = 10  # Number of epochs to wait before stopping
+    early_stopping_restore_best_weights = True  # Whether to restore best weights when stopping
+    warmup_enabled = False  # Whether to use learning rate warmup
+    warmup_epochs = 5  # Number of epochs for warmup
+    warmup_start_lr = 1e-6  # Starting learning rate for warmup
 
 
 @ex.automain
-def run(n_classes, n_epochs, input_height,
+def run(_config, n_classes, n_epochs, input_height,
         input_width, weight_decay, weighted_loss,
         index_start, dir_of_start_model, is_loss_soft_dice,
         n_batch, patches, augmentation, flip_aug,
-        blur_aug, scaling, binarization,
+        blur_aug, padding_white, padding_black, scaling, degrading,channels_shuffling,
-        blur_k, scales, dir_train, data_is_provided,
+        brightening, binarization, adding_rgb_background, adding_rgb_foreground, add_red_textlines, blur_k, scales, degrade_scales,shuffle_indexes,
-        scaling_bluring, scaling_binarization, rotation,
+        brightness, dir_train, data_is_provided, scaling_bluring,
-        rotation_not_90, thetha, scaling_flip, continue_training,
+        scaling_brightness, scaling_binarization, rotation, rotation_not_90,
-        flip_index, dir_eval, dir_output, pretraining, learning_rate):
+        thetha, scaling_flip, continue_training, transformer_projection_dim,
-    if data_is_provided:
+        transformer_mlp_head_units, transformer_layers, transformer_num_heads, transformer_cnn_first,
-        dir_train_flowing = os.path.join(dir_output, 'train')
+        transformer_patchsize_x, transformer_patchsize_y,
-        dir_eval_flowing = os.path.join(dir_output, 'eval')
+        transformer_num_patches_xy, backbone_type, flip_index, dir_eval, dir_output,
-
+        pretraining, learning_rate, task, f1_threshold_classification, classification_classes_name, dir_img_bin, number_of_backgrounds_per_image,dir_rgb_backgrounds, dir_rgb_foregrounds,
-        dir_flow_train_imgs = os.path.join(dir_train_flowing, 'images')
+        reduce_lr_enabled, reduce_lr_monitor, reduce_lr_factor, reduce_lr_patience, reduce_lr_min_lr, reduce_lr_min_delta,
-        dir_flow_train_labels = os.path.join(dir_train_flowing, 'labels')
+        early_stopping_enabled, early_stopping_monitor, early_stopping_patience, early_stopping_restore_best_weights,
+        warmup_enabled, warmup_epochs, warmup_start_lr):
+    
+    if dir_rgb_backgrounds:
+        list_all_possible_background_images = os.listdir(dir_rgb_backgrounds)
+    else:
+        list_all_possible_background_images = None
     
-        dir_flow_eval_imgs = os.path.join(dir_eval_flowing, 'images')
+    if dir_rgb_foregrounds:
-        dir_flow_eval_labels = os.path.join(dir_eval_flowing, 'labels')
+        list_all_possible_foreground_rgbs = os.listdir(dir_rgb_foregrounds)
+    else:
+        list_all_possible_foreground_rgbs = None
         
-        configuration()
+    if task == "segmentation" or task == "enhancement" or task == "binarization":
+        if data_is_provided:
+            dir_train_flowing = os.path.join(dir_output, 'train')
+            dir_eval_flowing = os.path.join(dir_output, 'eval')
                 
-    else:
-        dir_img, dir_seg = get_dirs_or_files(dir_train)
-        dir_img_val, dir_seg_val = get_dirs_or_files(dir_eval)
 
-        # make first a directory in output for both training and evaluations in order to flow data from these directories.
+            dir_flow_train_imgs = os.path.join(dir_train_flowing, 'images')
-        dir_train_flowing = os.path.join(dir_output, 'train')
+            dir_flow_train_labels = os.path.join(dir_train_flowing, 'labels')
-        dir_eval_flowing = os.path.join(dir_output, 'eval')
 
-        dir_flow_train_imgs = os.path.join(dir_train_flowing, 'images/')
+            dir_flow_eval_imgs = os.path.join(dir_eval_flowing, 'images')
-        dir_flow_train_labels = os.path.join(dir_train_flowing, 'labels/')
+            dir_flow_eval_labels = os.path.join(dir_eval_flowing, 'labels')
 
-        dir_flow_eval_imgs = os.path.join(dir_eval_flowing, 'images/')
+            configuration()
-        dir_flow_eval_labels = os.path.join(dir_eval_flowing, 'labels/')
 
-        if os.path.isdir(dir_train_flowing):
-            os.system('rm -rf ' + dir_train_flowing)
-            os.makedirs(dir_train_flowing)
         else:
-            os.makedirs(dir_train_flowing)
+            dir_img, dir_seg = get_dirs_or_files(dir_train)
+            dir_img_val, dir_seg_val = get_dirs_or_files(dir_eval)
+
+            # make first a directory in output for both training and evaluations in order to flow data from these directories.
+            dir_train_flowing = os.path.join(dir_output, 'train')
+            dir_eval_flowing = os.path.join(dir_output, 'eval')
+
+            dir_flow_train_imgs = os.path.join(dir_train_flowing, 'images/')
+            dir_flow_train_labels = os.path.join(dir_train_flowing, 'labels/')
+
+            dir_flow_eval_imgs = os.path.join(dir_eval_flowing, 'images/')
+            dir_flow_eval_labels = os.path.join(dir_eval_flowing, 'labels/')
+
+            if os.path.isdir(dir_train_flowing):
+                os.system('rm -rf ' + dir_train_flowing)
+                os.makedirs(dir_train_flowing)
+            else:
+                os.makedirs(dir_train_flowing)
+
+            if os.path.isdir(dir_eval_flowing):
+                os.system('rm -rf ' + dir_eval_flowing)
+                os.makedirs(dir_eval_flowing)
+            else:
+                os.makedirs(dir_eval_flowing)
+
+            os.mkdir(dir_flow_train_imgs)
+            os.mkdir(dir_flow_train_labels)
+
+            os.mkdir(dir_flow_eval_imgs)
+            os.mkdir(dir_flow_eval_labels)
+
+            # set the gpu configuration
+            configuration()
+            
+            imgs_list=np.array(os.listdir(dir_img))
+            segs_list=np.array(os.listdir(dir_seg))
+            
+            imgs_list_test=np.array(os.listdir(dir_img_val))
+            segs_list_test=np.array(os.listdir(dir_seg_val))
+
+            # writing patches into a sub-folder in order to be flowed from directory.
+            provide_patches(imgs_list, segs_list, dir_img, dir_seg, dir_flow_train_imgs,
+                            dir_flow_train_labels, input_height, input_width, blur_k,
+                            blur_aug, padding_white, padding_black, flip_aug, binarization, adding_rgb_background,adding_rgb_foreground, add_red_textlines, channels_shuffling,
+                            scaling, degrading, brightening, scales, degrade_scales, brightness,
+                            flip_index,shuffle_indexes, scaling_bluring, scaling_brightness, scaling_binarization,
+                            rotation, rotation_not_90, thetha, scaling_flip, task, augmentation=augmentation,
+                            patches=patches, dir_img_bin=dir_img_bin,number_of_backgrounds_per_image=number_of_backgrounds_per_image,list_all_possible_background_images=list_all_possible_background_images, dir_rgb_backgrounds=dir_rgb_backgrounds, dir_rgb_foregrounds=dir_rgb_foregrounds,list_all_possible_foreground_rgbs=list_all_possible_foreground_rgbs)
+            
+            provide_patches(imgs_list_test, segs_list_test, dir_img_val, dir_seg_val,
+                            dir_flow_eval_imgs, dir_flow_eval_labels, input_height, input_width,
+                            blur_k, blur_aug, padding_white, padding_black, flip_aug, binarization, adding_rgb_background, adding_rgb_foreground, add_red_textlines, channels_shuffling,
+                            scaling, degrading, brightening, scales, degrade_scales, brightness,
+                            flip_index, shuffle_indexes, scaling_bluring, scaling_brightness, scaling_binarization,
+                            rotation, rotation_not_90, thetha, scaling_flip, task, augmentation=False, patches=patches,dir_img_bin=dir_img_bin,number_of_backgrounds_per_image=number_of_backgrounds_per_image,list_all_possible_background_images=list_all_possible_background_images, dir_rgb_backgrounds=dir_rgb_backgrounds,dir_rgb_foregrounds=dir_rgb_foregrounds,list_all_possible_foreground_rgbs=list_all_possible_foreground_rgbs )
 
-        if os.path.isdir(dir_eval_flowing):
+        if weighted_loss:
-            os.system('rm -rf ' + dir_eval_flowing)
+            weights = np.zeros(n_classes)
-            os.makedirs(dir_eval_flowing)
+            if data_is_provided:
+                for obj in os.listdir(dir_flow_train_labels):
+                    try:
+                        label_obj = cv2.imread(dir_flow_train_labels + '/' + obj)
+                        label_obj_one_hot = get_one_hot(label_obj, label_obj.shape[0], label_obj.shape[1], n_classes)
+                        weights += (label_obj_one_hot.sum(axis=0)).sum(axis=0)
+                    except:
+                        pass
+            else:
+
+                for obj in os.listdir(dir_seg):
+                    try:
+                        label_obj = cv2.imread(dir_seg + '/' + obj)
+                        label_obj_one_hot = get_one_hot(label_obj, label_obj.shape[0], label_obj.shape[1], n_classes)
+                        weights += (label_obj_one_hot.sum(axis=0)).sum(axis=0)
+                    except:
+                        pass
+
+            weights = 1.00 / weights
+
+            weights = weights / float(np.sum(weights))
+            weights = weights / float(np.min(weights))
+            weights = weights / float(np.sum(weights))
+
+        if continue_training:
+            if backbone_type=='nontransformer':
+                if is_loss_soft_dice and (task == "segmentation" or task == "binarization"):
+                    model = load_model(dir_of_start_model, compile=True, custom_objects={'soft_dice_loss': soft_dice_loss})
+                if weighted_loss and (task == "segmentation" or task == "binarization"):
+                    model = load_model(dir_of_start_model, compile=True, custom_objects={'loss': weighted_categorical_crossentropy(weights)})
+                if not is_loss_soft_dice and not weighted_loss:
+                    model = load_model(dir_of_start_model , compile=True)
+            elif backbone_type=='transformer':
+                if is_loss_soft_dice and (task == "segmentation" or task == "binarization"):
+                    model = load_model(dir_of_start_model, compile=True, custom_objects={"PatchEncoder": PatchEncoder, "Patches": Patches,'soft_dice_loss': soft_dice_loss})
+                if weighted_loss and (task == "segmentation" or task == "binarization"):
+                    model = load_model(dir_of_start_model, compile=True, custom_objects={'loss': weighted_categorical_crossentropy(weights)})
+                if not is_loss_soft_dice and not weighted_loss:
+                    model = load_model(dir_of_start_model , compile=True,custom_objects = {"PatchEncoder": PatchEncoder, "Patches": Patches})
+        else:
+            index_start = 0
+            if backbone_type=='nontransformer':
+                model = resnet50_unet(n_classes, input_height, input_width, task, weight_decay, pretraining)
+            elif backbone_type=='transformer':
+                num_patches_x = transformer_num_patches_xy[0]
+                num_patches_y = transformer_num_patches_xy[1]
+                num_patches = num_patches_x * num_patches_y
+                
+                if transformer_cnn_first:
+                    if (input_height != (num_patches_y * transformer_patchsize_y * 32) ):
+                        print("Error: transformer_patchsize_y or transformer_num_patches_xy height value error . input_height should be equal to ( transformer_num_patches_xy height value * transformer_patchsize_y * 32)")
+                        sys.exit(1)
+                    if (input_width != (num_patches_x * transformer_patchsize_x * 32) ):
+                        print("Error: transformer_patchsize_x or transformer_num_patches_xy width value error . input_width should be equal to ( transformer_num_patches_xy width value * transformer_patchsize_x * 32)")
+                        sys.exit(1)
+                    if (transformer_projection_dim % (transformer_patchsize_y * transformer_patchsize_x)) != 0:
+                        print("Error: transformer_projection_dim error. The remainder when parameter transformer_projection_dim is divided by (transformer_patchsize_y*transformer_patchsize_x) should be zero")
+                        sys.exit(1)
+                        
+                    
+                    model = vit_resnet50_unet(n_classes, transformer_patchsize_x, transformer_patchsize_y, num_patches, transformer_mlp_head_units, transformer_layers, transformer_num_heads, transformer_projection_dim, input_height, input_width, task, weight_decay, pretraining)
+                else:
+                    if (input_height != (num_patches_y * transformer_patchsize_y) ):
+                        print("Error: transformer_patchsize_y or transformer_num_patches_xy height value error . input_height should be equal to ( transformer_num_patches_xy height value * transformer_patchsize_y)")
+                        sys.exit(1)
+                    if (input_width != (num_patches_x * transformer_patchsize_x) ):
+                        print("Error: transformer_patchsize_x or transformer_num_patches_xy width value error . input_width should be equal to ( transformer_num_patches_xy width value * transformer_patchsize_x)")
+                        sys.exit(1)
+                    if (transformer_projection_dim % (transformer_patchsize_y * transformer_patchsize_x)) != 0:
+                        print("Error: transformer_projection_dim error. The remainder when parameter transformer_projection_dim is divided by (transformer_patchsize_y*transformer_patchsize_x) should be zero")
+                        sys.exit(1)
+                    model = vit_resnet50_unet_transformer_before_cnn(n_classes, transformer_patchsize_x, transformer_patchsize_y, num_patches, transformer_mlp_head_units, transformer_layers, transformer_num_heads, transformer_projection_dim, input_height, input_width, task, weight_decay, pretraining)
+        
+        #if you want to see the model structure just uncomment model summary.
+        model.summary()
+
+        
+        # Create callbacks list
+        callbacks = []
+        if reduce_lr_enabled:
+            reduce_lr = ReduceLROnPlateau(
+                monitor=reduce_lr_monitor,
+                factor=reduce_lr_factor,
+                patience=reduce_lr_patience,
+                min_lr=reduce_lr_min_lr,
+                min_delta=reduce_lr_min_delta,
+                verbose=1
+            )
+            callbacks.append(reduce_lr)
+            
+        if early_stopping_enabled:
+            early_stopping = EarlyStopping(
+                monitor=early_stopping_monitor,
+                patience=early_stopping_patience,
+                restore_best_weights=early_stopping_restore_best_weights,
+                verbose=1
+            )
+            callbacks.append(early_stopping)
+        
+        # Add checkpoint to save models every epoch
+        class ModelCheckpointWithConfig(ModelCheckpoint):
+            def __init__(self, *args, **kwargs):
+                self._config = _config
+                super().__init__(*args, **kwargs)
+            
+            def on_epoch_end(self, epoch, logs=None):
+                super().on_epoch_end(epoch, logs)
+                model_dir = os.path.join(dir_output, f"model_{epoch+1}")
+                with open(os.path.join(model_dir, "config.json"), "w") as fp:
+                    json.dump(self._config, fp)
+        
+        checkpoint_epoch = ModelCheckpointWithConfig(
+            os.path.join(dir_output, "model_{epoch}"),
+            save_freq='epoch',
+            save_weights_only=False,
+            save_best_only=False,
+            verbose=1
+        )
+        callbacks.append(checkpoint_epoch)
+        
+        # Calculate steps per epoch
+        steps_per_epoch = int(len(os.listdir(dir_flow_train_imgs)) / n_batch) - 1
+        
+        # Create optimizer with or without warmup
+        if warmup_enabled:
+            lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
+                initial_learning_rate=warmup_start_lr,
+                decay_steps=warmup_epochs * steps_per_epoch,
+                end_learning_rate=learning_rate,
+                power=1.0  # Linear decay
+            )
+            optimizer = Adam(learning_rate=lr_schedule)
         else:
-            os.makedirs(dir_eval_flowing)
+            optimizer = Adam(learning_rate=learning_rate)
+        
+        if (task == "segmentation" or task == "binarization"):
+            if not is_loss_soft_dice and not weighted_loss:
+                model.compile(loss='categorical_crossentropy',
+                              optimizer=optimizer, metrics=['accuracy'])
+            if is_loss_soft_dice:                    
+                model.compile(loss=soft_dice_loss,
+                              optimizer=optimizer, metrics=['accuracy'])
+            if weighted_loss:
+                model.compile(loss=weighted_categorical_crossentropy(weights),
+                              optimizer=optimizer, metrics=['accuracy'])
+        elif task == "enhancement":
+            model.compile(loss='mean_squared_error',
+                          optimizer=optimizer, metrics=['accuracy'])
+        
+        # generating train and evaluation data
+        train_gen = data_gen(dir_flow_train_imgs, dir_flow_train_labels, batch_size=n_batch,
+                            input_height=input_height, input_width=input_width, n_classes=n_classes, task=task)
+        val_gen = data_gen(dir_flow_eval_imgs, dir_flow_eval_labels, batch_size=n_batch,
+                        input_height=input_height, input_width=input_width, n_classes=n_classes, task=task)
+        
+        # Single fit call with all epochs
+        history = model.fit(
+            train_gen,
+            steps_per_epoch=steps_per_epoch,
+            validation_data=val_gen,
+            validation_steps=1,
+            epochs=n_epochs,
+            callbacks=callbacks
+        )
         
-        os.mkdir(dir_flow_train_imgs)
+        # Save the best model (either from early stopping or final model)
-        os.mkdir(dir_flow_train_labels)
+        model.save(os.path.join(dir_output, 'model_best'))
         
-        os.mkdir(dir_flow_eval_imgs)
+        with open(os.path.join(dir_output, 'model_best', "config.json"), "w") as fp:
-        os.mkdir(dir_flow_eval_labels)
+            json.dump(_config, fp)  # encode dict into JSON
             
-        # set the gpu configuration
+    elif task=='classification':
         configuration()
+        model = resnet50_classifier(n_classes,  input_height, input_width, weight_decay, pretraining)
 
-        # writing patches into a sub-folder in order to be flowed from directory.
+        opt_adam = Adam(learning_rate=0.001)
-        provide_patches(dir_img, dir_seg, dir_flow_train_imgs,
+        model.compile(loss='categorical_crossentropy',
-                        dir_flow_train_labels,
+                            optimizer = opt_adam,metrics=['accuracy'])
-                        input_height, input_width, blur_k, blur_aug,
-                        flip_aug, binarization, scaling, scales, flip_index,
-                        scaling_bluring, scaling_binarization, rotation,
-                        rotation_not_90, thetha, scaling_flip,
-                        augmentation=augmentation, patches=patches)
-
-        provide_patches(dir_img_val, dir_seg_val, dir_flow_eval_imgs,
-                        dir_flow_eval_labels,
-                        input_height, input_width, blur_k, blur_aug,
-                        flip_aug, binarization, scaling, scales, flip_index,
-                        scaling_bluring, scaling_binarization, rotation,
-                        rotation_not_90, thetha, scaling_flip,
-                        augmentation=False, patches=patches)
-
-    if weighted_loss:
-        weights = np.zeros(n_classes)
-        if data_is_provided:
-            for obj in os.listdir(dir_flow_train_labels):
-                try:
-                    label_obj = cv2.imread(dir_flow_train_labels + '/' + obj)
-                    label_obj_one_hot = get_one_hot(label_obj, label_obj.shape[0], label_obj.shape[1], n_classes)
-                    weights += (label_obj_one_hot.sum(axis=0)).sum(axis=0)
-                except:
-                    pass
-        else:
 
-            for obj in os.listdir(dir_seg):
-                try:
-                    label_obj = cv2.imread(dir_seg + '/' + obj)
-                    label_obj_one_hot = get_one_hot(label_obj, label_obj.shape[0], label_obj.shape[1], n_classes)
-                    weights += (label_obj_one_hot.sum(axis=0)).sum(axis=0)
-                except:
-                    pass
         
-        weights = 1.00 / weights
+        list_classes = list(classification_classes_name.values())
+        testX, testY = generate_data_from_folder_evaluation(dir_eval, input_height, input_width, n_classes, list_classes)
 
-        weights = weights / float(np.sum(weights))
+        y_tot=np.zeros((testX.shape[0],n_classes))
-        weights = weights / float(np.min(weights))
-        weights = weights / float(np.sum(weights))
 
-    if continue_training:
+        score_best=[]
-        if is_loss_soft_dice:
+        score_best.append(0)
-            model = load_model(dir_of_start_model, compile=True, custom_objects={'soft_dice_loss': soft_dice_loss})
-        if weighted_loss:
-            model = load_model(dir_of_start_model, compile=True,
-                               custom_objects={'loss': weighted_categorical_crossentropy(weights)})
-        if not is_loss_soft_dice and not weighted_loss:
-            model = load_model(dir_of_start_model, compile=True)
-    else:
-        # get our model.
-        index_start = 0
-        model = resnet50_unet(n_classes, input_height, input_width, weight_decay, pretraining)
 
-    # if you want to see the model structure just uncomment model summary.
+        num_rows = return_number_of_total_training_data(dir_train)
-    # model.summary()
+        weights=[]
         
-    if not is_loss_soft_dice and not weighted_loss:
+        for i in range(n_epochs):
-        model.compile(loss='categorical_crossentropy',
+            history = model.fit( generate_data_from_folder_training(dir_train, n_batch , input_height, input_width, n_classes, list_classes), steps_per_epoch=num_rows / n_batch, verbose=1)#,class_weight=weights)
-                      optimizer=Adam(lr=learning_rate), metrics=['accuracy'])
+            
-    if is_loss_soft_dice:
+            y_pr_class = []
-        model.compile(loss=soft_dice_loss,
+            for jj in range(testY.shape[0]):
-                      optimizer=Adam(lr=learning_rate), metrics=['accuracy'])
+                y_pr=model.predict(testX[jj,:,:,:].reshape(1,input_height,input_width,3), verbose=0)
-
+                y_pr_ind= np.argmax(y_pr,axis=1)
-    if weighted_loss:
+                y_pr_class.append(y_pr_ind)
-        model.compile(loss=weighted_categorical_crossentropy(weights),
+            
-                      optimizer=Adam(lr=learning_rate), metrics=['accuracy'])
+            y_pr_class = np.array(y_pr_class)
-
+            f1score=f1_score(np.argmax(testY,axis=1), y_pr_class, average='macro')
-    # generating train and evaluation data
+            print(i,f1score)
-    train_gen = data_gen(dir_flow_train_imgs, dir_flow_train_labels, batch_size=n_batch,
+            
-                         input_height=input_height, input_width=input_width, n_classes=n_classes)
+            if f1score>score_best[0]:
-    val_gen = data_gen(dir_flow_eval_imgs, dir_flow_eval_labels, batch_size=n_batch,
+                score_best[0]=f1score
-                       input_height=input_height, input_width=input_width, n_classes=n_classes)
+                model.save(os.path.join(dir_output,'model_best'))
-
+                
-    for i in tqdm(range(index_start, n_epochs + index_start)):
+            if f1score > f1_threshold_classification:
-        model.fit_generator(
+                weights.append(model.get_weights() )
-            train_gen,
+                
-            steps_per_epoch=int(len(os.listdir(dir_flow_train_imgs)) / n_batch) - 1,
+
-            validation_data=val_gen,
+        if len(weights) >= 1:
-            validation_steps=1,
+            new_weights=list()
-            epochs=1)
+            for weights_list_tuple in zip(*weights):
-        model.save(dir_output + '/' + 'model_' + str(i))
+                new_weights.append( [np.array(weights_).mean(axis=0) for weights_ in zip(*weights_list_tuple)]  )
+                
+            new_weights = [np.array(x) for x in new_weights]
+            model_weight_averaged=tf.keras.models.clone_model(model)
+            model_weight_averaged.set_weights(new_weights)
+    
+            model_weight_averaged.save(os.path.join(dir_output,'model_ens_avg'))
+            with open(os.path.join( os.path.join(dir_output,'model_ens_avg'), "config.json"), "w") as fp:
+                json.dump(_config, fp)  # encode dict into JSON
+            
+        with open(os.path.join( os.path.join(dir_output,'model_best'), "config.json"), "w") as fp:
+            json.dump(_config, fp)  # encode dict into JSON
+            
+    elif task=='reading_order':
+        configuration()
+        model = machine_based_reading_order_model(n_classes,input_height,input_width,weight_decay,pretraining)
+        
+        dir_flow_train_imgs = os.path.join(dir_train, 'images')
+        dir_flow_train_labels = os.path.join(dir_train, 'labels')
+        
+        classes = os.listdir(dir_flow_train_labels)
+        num_rows =len(classes)
+        #ls_test = os.listdir(dir_flow_train_labels)
+
+        #f1score_tot = [0]
+        indexer_start = 0
+        opt = SGD(learning_rate=0.01, momentum=0.9)
+        opt_adam = tf.keras.optimizers.Adam(learning_rate=0.0001)
+        model.compile(loss="binary_crossentropy",
+                            optimizer = opt_adam,metrics=['accuracy'])
+        for i in range(n_epochs):
+            history = model.fit(generate_arrays_from_folder_reading_order(dir_flow_train_labels, dir_flow_train_imgs, n_batch, input_height, input_width, n_classes), steps_per_epoch=num_rows / n_batch, verbose=1)
+            model.save( os.path.join(dir_output,'model_'+str(i+indexer_start) ))
+            
+            with open(os.path.join(os.path.join(dir_output,'model_'+str(i)),"config.json"), "w") as fp:
+                json.dump(_config, fp)  # encode dict into JSON
+            '''
+            if f1score>f1score_tot[0]:
+                f1score_tot[0] = f1score
+                model_dir = os.path.join(dir_out,'model_best')
+                model.save(model_dir)
+            '''
 
-    # os.system('rm -rf '+dir_train_flowing)
-    # os.system('rm -rf '+dir_eval_flowing)
     
-    # model.save(dir_output+'/'+'model'+'.h5')

train_no_patches_448x448.json

@@ -0,0 +1,53 @@
+{
+    "backbone_type" : "nontransformer",
+    "task": "segmentation",
+    "n_classes" : 3,
+    "n_epochs" : 50,
+    "input_height" : 448,
+    "input_width" : 448,
+    "weight_decay" : 1e-4,
+    "n_batch" : 4,
+    "learning_rate": 2e-5,
+    "patches" : false,
+    "pretraining" : true,
+    "augmentation" : true,
+    "flip_aug" : false,
+    "blur_aug" : false,
+    "scaling" : true,
+    "degrading": true,
+    "brightening": false,
+    "binarization" : false,
+    "scaling_bluring" : false,
+    "scaling_binarization" : false,
+    "scaling_flip" : false,
+    "rotation": false,
+    "rotation_not_90": false,
+    "blur_k" : ["blur","guass","median"],
+    "scales" : [0.6, 0.7, 0.8, 0.9],
+    "brightness" : [1.3, 1.5, 1.7, 2],
+    "degrade_scales" : [0.2, 0.4],
+    "flip_index" : [0, 1, -1],
+    "thetha" : [10, -10],
+    "continue_training": false,
+    "index_start" : 0,
+    "dir_of_start_model" : " ",
+    "weighted_loss": false,
+    "is_loss_soft_dice": true,
+    "data_is_provided": true,
+    "dir_train": "/home/incognito/sbb_pixelwise_segmentation/dataset/sam_41_mss/dir_train/train",
+    "dir_eval": "/home/incognito/sbb_pixelwise_segmentation/dataset/sam_41_mss/dir_train/eval",
+    "dir_output": "runs/sam_41_mss_npt_448x448",
+    "reduce_lr_enabled": true,
+    "reduce_lr_monitor": "val_loss",
+    "reduce_lr_factor": 0.2,
+    "reduce_lr_patience": 3,
+    "reduce_lr_min_delta": 1e-5,
+    "reduce_lr_min_lr": 1e-6,
+    "early_stopping_enabled": true,
+    "early_stopping_monitor": "val_loss",
+    "early_stopping_patience": 6,
+    "early_stopping_restore_best_weights": true,
+    "warmup_enabled": true,
+    "warmup_epochs": 5,
+    "warmup_start_lr": 1e-6
+}

utils.py

@@ -8,6 +8,228 @@ import random
 from tqdm import tqdm
 import imutils
 import math
+from tensorflow.keras.utils import to_categorical
+from PIL import Image, ImageEnhance
+
+
+def return_shuffled_channels(img, channels_order):
+    """
+    channels order in ordinary case is like this [0, 1, 2]. In the case of shuffling the order should be provided.
+    """
+    img_sh = np.copy(img)
+    
+    img_sh[:,:,0]= img[:,:,channels_order[0]]
+    img_sh[:,:,1]= img[:,:,channels_order[1]]
+    img_sh[:,:,2]= img[:,:,channels_order[2]]
+    return img_sh
+
+def return_binary_image_with_red_textlines(img_bin):
+    img_red = np.copy(img_bin)
+    
+    img_red[:,:,0][img_bin[:,:,0] == 0] = 255
+    return img_red
+
+def return_binary_image_with_given_rgb_background(img_bin, img_rgb_background):
+    img_rgb_background = resize_image(img_rgb_background ,img_bin.shape[0], img_bin.shape[1])
+    
+    img_final = np.copy(img_bin)
+    
+    img_final[:,:,0][img_bin[:,:,0] != 0] = img_rgb_background[:,:,0][img_bin[:,:,0] != 0]
+    img_final[:,:,1][img_bin[:,:,1] != 0] = img_rgb_background[:,:,1][img_bin[:,:,1] != 0]
+    img_final[:,:,2][img_bin[:,:,2] != 0] = img_rgb_background[:,:,2][img_bin[:,:,2] != 0]
+    
+    return img_final
+
+def return_binary_image_with_given_rgb_background_and_given_foreground_rgb(img_bin, img_rgb_background, rgb_foreground):
+    img_rgb_background = resize_image(img_rgb_background ,img_bin.shape[0], img_bin.shape[1])
+    
+    img_final = np.copy(img_bin)
+    img_foreground = np.zeros(img_bin.shape)
+    
+    
+    img_foreground[:,:,0][img_bin[:,:,0] == 0] = rgb_foreground[0]
+    img_foreground[:,:,1][img_bin[:,:,0] == 0] = rgb_foreground[1]
+    img_foreground[:,:,2][img_bin[:,:,0] == 0] = rgb_foreground[2]
+    
+    
+    img_final[:,:,0][img_bin[:,:,0] != 0] = img_rgb_background[:,:,0][img_bin[:,:,0] != 0]
+    img_final[:,:,1][img_bin[:,:,1] != 0] = img_rgb_background[:,:,1][img_bin[:,:,1] != 0]
+    img_final[:,:,2][img_bin[:,:,2] != 0] = img_rgb_background[:,:,2][img_bin[:,:,2] != 0]
+    
+    img_final = img_final + img_foreground
+    return img_final
+
+def return_binary_image_with_given_rgb_background_red_textlines(img_bin, img_rgb_background, img_color):
+    img_rgb_background = resize_image(img_rgb_background ,img_bin.shape[0], img_bin.shape[1])
+    
+    img_final = np.copy(img_color)
+    
+    img_final[:,:,0][img_bin[:,:,0] != 0] = img_rgb_background[:,:,0][img_bin[:,:,0] != 0]
+    img_final[:,:,1][img_bin[:,:,1] != 0] = img_rgb_background[:,:,1][img_bin[:,:,1] != 0]
+    img_final[:,:,2][img_bin[:,:,2] != 0] = img_rgb_background[:,:,2][img_bin[:,:,2] != 0]
+    
+    return img_final
+
+def return_image_with_red_elements(img, img_bin):
+    img_final = np.copy(img)
+    
+    img_final[:,:,0][img_bin[:,:,0]==0] = 0
+    img_final[:,:,1][img_bin[:,:,0]==0] = 0
+    img_final[:,:,2][img_bin[:,:,0]==0] = 255
+    return img_final
+    
+    
+
+def scale_image_for_no_patch(img, label, scale):
+    h_n = int(img.shape[0]*scale)
+    w_n = int(img.shape[1]*scale)
+    
+    channel0_avg = int( np.mean(img[:,:,0]) )
+    channel1_avg = int( np.mean(img[:,:,1]) )
+    channel2_avg = int( np.mean(img[:,:,2]) )
+    
+    h_diff = img.shape[0] - h_n
+    w_diff = img.shape[1] - w_n
+    
+    h_start = int(h_diff / 2.)
+    w_start = int(w_diff / 2.)
+    
+    img_res = resize_image(img, h_n, w_n)
+    label_res = resize_image(label, h_n, w_n)
+    
+    img_scaled_padded = np.copy(img)
+    
+    label_scaled_padded = np.zeros(label.shape)
+    
+    img_scaled_padded[:,:,0] = channel0_avg
+    img_scaled_padded[:,:,1] = channel1_avg
+    img_scaled_padded[:,:,2] = channel2_avg
+    
+    img_scaled_padded[h_start:h_start+h_n, w_start:w_start+w_n,:] = img_res[:,:,:]
+    label_scaled_padded[h_start:h_start+h_n, w_start:w_start+w_n,:] = label_res[:,:,:]
+    
+    return img_scaled_padded, label_scaled_padded
+
+
+def return_number_of_total_training_data(path_classes):
+    sub_classes = os.listdir(path_classes)
+    n_tot = 0
+    for sub_c in sub_classes:
+        sub_files =  os.listdir(os.path.join(path_classes,sub_c))
+        n_tot = n_tot + len(sub_files)
+    return n_tot
+        
+    
+    
+def generate_data_from_folder_evaluation(path_classes, height, width, n_classes, list_classes):
+    #sub_classes = os.listdir(path_classes)
+    #n_classes = len(sub_classes)
+    all_imgs = []
+    labels = []
+    #dicts =dict()
+    #indexer= 0
+    for indexer, sub_c in enumerate(list_classes):
+        sub_files =  os.listdir(os.path.join(path_classes,sub_c  )) 
+        sub_files = [os.path.join(path_classes,sub_c  )+'/' + x for x in sub_files]
+        #print(     os.listdir(os.path.join(path_classes,sub_c  ))     )
+        all_imgs = all_imgs + sub_files
+        sub_labels = list( np.zeros( len(sub_files) ) +indexer )
+
+        #print( len(sub_labels) )
+        labels = labels + sub_labels
+        #dicts[sub_c] = indexer
+        #indexer +=1 
+        
+
+    categories =  to_categorical(range(n_classes)).astype(np.int16)#[  [1 , 0, 0 , 0 , 0 , 0]  , [0 , 1, 0 , 0 , 0 , 0]  , [0 , 0, 1 , 0 , 0 , 0] , [0 , 0, 0 , 1 , 0 , 0] , [0 , 0, 0 , 0 , 1 , 0]  , [0 , 0, 0 , 0 , 0 , 1] ]
+    ret_x= np.zeros((len(labels), height,width, 3)).astype(np.int16)
+    ret_y= np.zeros((len(labels), n_classes)).astype(np.int16)
+    
+    #print(all_imgs)
+    for i in range(len(all_imgs)):
+        row = all_imgs[i]
+        #####img = cv2.imread(row, 0)
+        #####img= resize_image (img, height, width)
+        #####img = img.astype(np.uint16)
+        #####ret_x[i, :,:,0] = img[:,:]
+        #####ret_x[i, :,:,1] = img[:,:]
+        #####ret_x[i, :,:,2] = img[:,:]
+        
+        img = cv2.imread(row)
+        img= resize_image (img, height, width)
+        img = img.astype(np.uint16)
+        ret_x[i, :,:] = img[:,:,:]
+        
+        ret_y[i, :] =  categories[ int( labels[i] ) ][:]
+    
+    return ret_x/255., ret_y
+
+def generate_data_from_folder_training(path_classes, batchsize, height, width, n_classes, list_classes):
+    #sub_classes = os.listdir(path_classes)
+    #n_classes = len(sub_classes)
+
+    all_imgs = []
+    labels = []
+    #dicts =dict()
+    #indexer= 0
+    for indexer, sub_c in enumerate(list_classes):
+        sub_files =  os.listdir(os.path.join(path_classes,sub_c  )) 
+        sub_files = [os.path.join(path_classes,sub_c  )+'/' + x for x in sub_files]
+        #print(     os.listdir(os.path.join(path_classes,sub_c  ))     )
+        all_imgs = all_imgs + sub_files
+        sub_labels = list( np.zeros( len(sub_files) ) +indexer )
+
+        #print( len(sub_labels) )
+        labels = labels + sub_labels
+        #dicts[sub_c] = indexer
+        #indexer +=1 
+        
+    ids = np.array(range(len(labels)))
+    random.shuffle(ids)
+    
+    shuffled_labels = np.array(labels)[ids]
+    shuffled_files = np.array(all_imgs)[ids]
+    categories = to_categorical(range(n_classes)).astype(np.int16)#[  [1 , 0, 0 , 0 , 0 , 0]  , [0 , 1, 0 , 0 , 0 , 0]  , [0 , 0, 1 , 0 , 0 , 0] , [0 , 0, 0 , 1 , 0 , 0] , [0 , 0, 0 , 0 , 1 , 0]  , [0 , 0, 0 , 0 , 0 , 1] ]
+    ret_x= np.zeros((batchsize, height,width, 3)).astype(np.int16)
+    ret_y= np.zeros((batchsize, n_classes)).astype(np.int16)
+    batchcount = 0
+    while True:
+        for i in range(len(shuffled_files)):
+            row = shuffled_files[i]
+            #print(row)
+            ###img = cv2.imread(row, 0)
+            ###img= resize_image (img, height, width)
+            ###img = img.astype(np.uint16)
+            ###ret_x[batchcount, :,:,0] = img[:,:]
+            ###ret_x[batchcount, :,:,1] = img[:,:]
+            ###ret_x[batchcount, :,:,2] = img[:,:]
+            
+            img = cv2.imread(row)
+            img= resize_image (img, height, width)
+            img = img.astype(np.uint16)
+            ret_x[batchcount, :,:,:] = img[:,:,:]
+            
+            #print(int(shuffled_labels[i]) )
+            #print( categories[int(shuffled_labels[i])] )
+            ret_y[batchcount, :] =  categories[ int( shuffled_labels[i] ) ][:]
+            
+            batchcount+=1
+            
+            if batchcount>=batchsize:
+                ret_x = ret_x/255.
+                yield (ret_x, ret_y)
+                ret_x= np.zeros((batchsize, height,width, 3)).astype(np.int16)
+                ret_y= np.zeros((batchsize, n_classes)).astype(np.int16)
+                batchcount = 0
+
+def do_brightening(img_in_dir, factor):
+    im = Image.open(img_in_dir)
+    enhancer = ImageEnhance.Brightness(im)
+    out_img = enhancer.enhance(factor)
+    out_img = out_img.convert('RGB')
+    opencv_img = np.array(out_img)
+    opencv_img = opencv_img[:,:,::-1].copy()
+    return opencv_img
 
 
 def bluring(img_in, kind):
@@ -138,15 +360,38 @@ def IoU(Yi, y_predi):
         FP = np.sum((Yi != c) & (y_predi == c))
         FN = np.sum((Yi == c) & (y_predi != c))
         IoU = TP / float(TP + FP + FN)
-        print("class {:02.0f}: #TP={:6.0f}, #FP={:6.0f}, #FN={:5.0f}, IoU={:4.3f}".format(c, TP, FP, FN, IoU))
+        #print("class {:02.0f}: #TP={:6.0f}, #FP={:6.0f}, #FN={:5.0f}, IoU={:4.3f}".format(c, TP, FP, FN, IoU))
         IoUs.append(IoU)
     mIoU = np.mean(IoUs)
-    print("_________________")
+    #print("_________________")
-    print("Mean IoU: {:4.3f}".format(mIoU))
+    #print("Mean IoU: {:4.3f}".format(mIoU))
     return mIoU
 
+def generate_arrays_from_folder_reading_order(classes_file_dir, modal_dir, batchsize, height, width, n_classes):
+    all_labels_files = os.listdir(classes_file_dir)
+    ret_x= np.zeros((batchsize, height, width, 3))#.astype(np.int16)
+    ret_y= np.zeros((batchsize, n_classes)).astype(np.int16)
+    batchcount = 0
+    while True:
+        for i in all_labels_files:
+            file_name = os.path.splitext(i)[0]
+            img = cv2.imread(os.path.join(modal_dir,file_name+'.png'))
 
-def data_gen(img_folder, mask_folder, batch_size, input_height, input_width, n_classes):
+            label_class = int( np.load(os.path.join(classes_file_dir,i)) )
+
+            ret_x[batchcount, :,:,0] = img[:,:,0]/3.0
+            ret_x[batchcount, :,:,2] = img[:,:,2]/3.0
+            ret_x[batchcount, :,:,1] = img[:,:,1]/5.0
+
+            ret_y[batchcount, :] =  label_class
+            batchcount+=1
+            if batchcount>=batchsize:
+                yield (ret_x, ret_y)
+                ret_x= np.zeros((batchsize, height, width, 3))#.astype(np.int16)
+                ret_y= np.zeros((batchsize, n_classes)).astype(np.int16)
+                batchcount = 0
+
+def data_gen(img_folder, mask_folder, batch_size, input_height, input_width, n_classes, task='segmentation'):
     c = 0
     n = [f for f in os.listdir(img_folder) if not f.startswith('.')]  # os.listdir(img_folder) #List of training images
     random.shuffle(n)
@@ -155,21 +400,22 @@ def data_gen(img_folder, mask_folder, batch_size, input_height, input_width, n_c
         mask = np.zeros((batch_size, input_height, input_width, n_classes)).astype('float')
 
         for i in range(c, c + batch_size):  # initially from 0 to 16, c = 0.
-            # print(img_folder+'/'+n[i])
-
             try:
-                filename = n[i].split('.')[0]
+                filename = os.path.splitext(n[i])[0]
 
                 train_img = cv2.imread(img_folder + '/' + n[i]) / 255.
                 train_img = cv2.resize(train_img, (input_width, input_height),
                                        interpolation=cv2.INTER_NEAREST)  # Read an image from folder and resize
 
                 img[i - c] = train_img  # add to array - img[0], img[1], and so on.
-                train_mask = cv2.imread(mask_folder + '/' + filename + '.png')
+                if task == "segmentation" or task=="binarization":
-                # print(mask_folder+'/'+filename+'.png')
+                    train_mask = cv2.imread(mask_folder + '/' + filename + '.png')
-                # print(train_mask.shape)
+                    train_mask = get_one_hot(resize_image(train_mask, input_height, input_width), input_height, input_width,
-                train_mask = get_one_hot(resize_image(train_mask, input_height, input_width), input_height, input_width,
+                                            n_classes)
-                                         n_classes)
+                elif task == "enhancement":
+                    train_mask = cv2.imread(mask_folder + '/' + filename + '.png')/255.
+                    train_mask = resize_image(train_mask, input_height, input_width)
+                    
                 # train_mask = train_mask.reshape(224, 224, 1) # Add extra dimension for parity with train_img size [512 * 512 * 3]
 
                 mask[i - c] = train_mask
@@ -241,9 +487,56 @@ def get_patches(dir_img_f, dir_seg_f, img, label, height, width, indexer):
     return indexer
 
 
+def do_padding_white(img):
+    img_org_h = img.shape[0]
+    img_org_w = img.shape[1]
+    
+    index_start_h = 4
+    index_start_w = 4
+    
+    img_padded = np.zeros((img.shape[0] + 2*index_start_h, img.shape[1]+ 2*index_start_w, img.shape[2])) + 255
+    img_padded[index_start_h: index_start_h + img.shape[0], index_start_w: index_start_w + img.shape[1], :] = img[:, :, :]
+    
+    return img_padded.astype(float)
+
+
+def do_degrading(img, scale):
+    img_org_h = img.shape[0]
+    img_org_w = img.shape[1]
+    
+    img_res = resize_image(img, int(img_org_h * scale), int(img_org_w * scale))
+    
+    return resize_image(img_res, img_org_h, img_org_w)
+    
+    
+def do_padding_black(img):
+    img_org_h = img.shape[0]
+    img_org_w = img.shape[1]
+    
+    index_start_h = 4
+    index_start_w = 4
+    
+    img_padded = np.zeros((img.shape[0] + 2*index_start_h, img.shape[1] + 2*index_start_w, img.shape[2]))
+    img_padded[index_start_h: index_start_h + img.shape[0], index_start_w: index_start_w + img.shape[1], :] = img[:, :, :]
+    
+    return img_padded.astype(float)
+
+
+def do_padding_label(img):
+    img_org_h = img.shape[0]
+    img_org_w = img.shape[1]
+    
+    index_start_h = 4
+    index_start_w = 4
+    
+    img_padded = np.zeros((img.shape[0] + 2*index_start_h, img.shape[1] + 2*index_start_w, img.shape[2]))
+    img_padded[index_start_h: index_start_h + img.shape[0], index_start_w: index_start_w + img.shape[1], :] = img[:, :, :]
+    
+    return img_padded.astype(np.int16)
+
 def do_padding(img, label, height, width):
-    height_new = img.shape[0]
+    height_new=img.shape[0]
-    width_new = img.shape[1]
+    width_new=img.shape[1]
     
     h_start = 0
     w_start = 0
@@ -262,7 +555,7 @@ def do_padding(img, label, height, width):
     img_new[h_start:h_start + img.shape[0], w_start:w_start + img.shape[1], :] = np.copy(img[:, :, :])
     label_new[h_start:h_start + label.shape[0], w_start:w_start + label.shape[1], :] = np.copy(label[:, :, :])
     
-    return img_new, label_new
+    return img_new,label_new
 
 
 def get_patches_num_scale(dir_img_f, dir_seg_f, img, label, height, width, indexer, n_patches, scaler):
@@ -275,6 +568,7 @@ def get_patches_num_scale(dir_img_f, dir_seg_f, img, label, height, width, index
     height_scale = int(height * scaler)
     width_scale = int(width * scaler)
     
+    
     nxf = img_w / float(width_scale)
     nyf = img_h / float(height_scale)
     
@@ -301,6 +595,7 @@ def get_patches_num_scale(dir_img_f, dir_seg_f, img, label, height, width, index
                 index_y_u = img_h
                 index_y_d = img_h - height_scale
                 
+            
             img_patch = img[index_y_d:index_y_u, index_x_d:index_x_u, :]
             label_patch = label[index_y_d:index_y_u, index_x_d:index_x_u, :]
             
@@ -356,9 +651,6 @@ def get_patches_num_scale_new(dir_img_f, dir_seg_f, img, label, height, width, i
             img_patch = img[index_y_d:index_y_u, index_x_d:index_x_u, :]
             label_patch = label[index_y_d:index_y_u, index_x_d:index_x_u, :]
             
-            # img_patch=resize_image(img_patch,height,width)
-            # label_patch=resize_image(label_patch,height,width)
-
             cv2.imwrite(dir_img_f + '/img_' + str(indexer) + '.png', img_patch)
             cv2.imwrite(dir_seg_f + '/img_' + str(indexer) + '.png', label_patch)
             indexer += 1
@@ -366,129 +658,311 @@ def get_patches_num_scale_new(dir_img_f, dir_seg_f, img, label, height, width, i
     return indexer
 
 
-def provide_patches(dir_img, dir_seg, dir_flow_train_imgs,
+def provide_patches(imgs_list_train, segs_list_train, dir_img, dir_seg, dir_flow_train_imgs,
-                    dir_flow_train_labels,
+                    dir_flow_train_labels, input_height, input_width, blur_k, blur_aug,
-                    input_height, input_width, blur_k, blur_aug,
+                    padding_white, padding_black, flip_aug, binarization, adding_rgb_background, adding_rgb_foreground, add_red_textlines, channels_shuffling, scaling, degrading,
-                    flip_aug, binarization, scaling, scales, flip_index,
+                    brightening, scales, degrade_scales, brightness, flip_index, shuffle_indexes,
-                    scaling_bluring, scaling_binarization, rotation,
+                    scaling_bluring, scaling_brightness, scaling_binarization, rotation,
-                    rotation_not_90, thetha, scaling_flip,
+                    rotation_not_90, thetha, scaling_flip, task, augmentation=False, patches=False, dir_img_bin=None,number_of_backgrounds_per_image=None,list_all_possible_background_images=None, dir_rgb_backgrounds=None, dir_rgb_foregrounds=None, list_all_possible_foreground_rgbs=None):
-                    augmentation=False, patches=False):
-    imgs_cv_train = np.array(os.listdir(dir_img))
-    segs_cv_train = np.array(os.listdir(dir_seg))
     
     indexer = 0
-    for im, seg_i in tqdm(zip(imgs_cv_train, segs_cv_train)):
+    for im, seg_i in tqdm(zip(imgs_list_train, segs_list_train)):
-        img_name = im.split('.')[0]
+        img_name = os.path.splitext(im)[0]
+        if task == "segmentation" or task == "binarization":
+            dir_of_label_file = os.path.join(dir_seg, img_name + '.png')
+        elif task=="enhancement":
+            dir_of_label_file = os.path.join(dir_seg, im)
+            
         if not patches:
-            cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+            cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(cv2.imread(dir_img + '/' + im), input_height, input_width))
-                        resize_image(cv2.imread(dir_img + '/' + im), input_height, input_width))
+            cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png', resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
-            cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
-                        resize_image(cv2.imread(dir_seg + '/' + img_name + '.png'), input_height, input_width))
             indexer += 1
             
             if augmentation:
                 if flip_aug:
                     for f_i in flip_index:
                         cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
-                                    resize_image(cv2.flip(cv2.imread(dir_img + '/' + im), f_i), input_height,
+                                    resize_image(cv2.flip(cv2.imread(dir_img+'/'+im),f_i),input_height,input_width) )
-                                                 input_width))
                         
                         cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
-                                    resize_image(cv2.flip(cv2.imread(dir_seg + '/' + img_name + '.png'), f_i),
+                                    resize_image(cv2.flip(cv2.imread(dir_of_label_file), f_i), input_height, input_width)) 
-                                                 input_height, input_width))
                         indexer += 1
                         
                 if blur_aug:   
                     for blur_i in blur_k:
                         cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
-                                    (resize_image(bluring(cv2.imread(dir_img + '/' + im), blur_i), input_height,
+                                    (resize_image(bluring(cv2.imread(dir_img + '/' + im), blur_i), input_height, input_width)))
-                                                  input_width)))
                         
                         cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
-                                    resize_image(cv2.imread(dir_seg + '/' + img_name + '.png'), input_height,
+                                    resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
-                                                 input_width))
                         indexer += 1
+                if brightening:
+                    for factor in brightness:
+                        try:
+                            cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                                        (resize_image(do_brightening(dir_img + '/' +im, factor), input_height, input_width)))
+                        
+                            cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                        resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                            indexer += 1
+                        except:
+                            pass
                     
                 if binarization:
-                    cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                    
-                                resize_image(otsu_copy(cv2.imread(dir_img + '/' + im)), input_height, input_width))
+                    if dir_img_bin:
+                        img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                        
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                                    resize_image(img_bin_corr, input_height, input_width))
+                    else:
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                                    resize_image(otsu_copy(cv2.imread(dir_img + '/' + im)), input_height, input_width))
                     
                     cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
-                                resize_image(cv2.imread(dir_seg + '/' + img_name + '.png'), input_height, input_width))
+                                resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
                     indexer += 1
                     
-        if patches:
+                if degrading:  
+                    for degrade_scale_ind in degrade_scales:
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                                    (resize_image(do_degrading(cv2.imread(dir_img + '/' + im), degrade_scale_ind), input_height, input_width)))
+                        
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                    resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                        indexer += 1
+                        
+                if rotation_not_90:
+                    for thetha_i in thetha:
+                        img_max_rotated, label_max_rotated = rotation_not_90_func(cv2.imread(dir_img + '/'+im),
+                                                                                  cv2.imread(dir_of_label_file), thetha_i)
+                        
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_max_rotated, input_height, input_width))
+                        
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png', resize_image(label_max_rotated, input_height, input_width))
+                        indexer += 1
+                        
+                if channels_shuffling:
+                    for shuffle_index in shuffle_indexes:
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
+                                    (resize_image(return_shuffled_channels(cv2.imread(dir_img + '/' + im), shuffle_index), input_height, input_width)))
+                        
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                    resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                        indexer += 1
+                        
+                if scaling:
+                    for sc_ind in scales:
+                        img_scaled, label_scaled = scale_image_for_no_patch(cv2.imread(dir_img + '/'+im),
+                                                                                  cv2.imread(dir_of_label_file), sc_ind)
+                        
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_scaled, input_height, input_width))
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png', resize_image(label_scaled, input_height, input_width))
+                        indexer += 1
+                        
+                if adding_rgb_background:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    for i_n in range(number_of_backgrounds_per_image):
+                        background_image_chosen_name = random.choice(list_all_possible_background_images)
+                        img_rgb_background_chosen = cv2.imread(dir_rgb_backgrounds + '/' + background_image_chosen_name)
+                        img_with_overlayed_background = return_binary_image_with_given_rgb_background(img_bin_corr, img_rgb_background_chosen)
+                        
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_with_overlayed_background, input_height, input_width))
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                    resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                        
+                        indexer += 1
+                        
+                if adding_rgb_foreground:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    for i_n in range(number_of_backgrounds_per_image):
+                        background_image_chosen_name = random.choice(list_all_possible_background_images)
+                        foreground_rgb_chosen_name = random.choice(list_all_possible_foreground_rgbs)
+                        
+                        img_rgb_background_chosen = cv2.imread(dir_rgb_backgrounds + '/' + background_image_chosen_name)
+                        foreground_rgb_chosen = np.load(dir_rgb_foregrounds + '/' + foreground_rgb_chosen_name)
+                        
+                        img_with_overlayed_background = return_binary_image_with_given_rgb_background_and_given_foreground_rgb(img_bin_corr, img_rgb_background_chosen, foreground_rgb_chosen)
+                        
+                        cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_with_overlayed_background, input_height, input_width))
+                        cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                    resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                        
+                        indexer += 1
+                        
+                if add_red_textlines:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    img_red_context = return_image_with_red_elements(cv2.imread(dir_img + '/'+im), img_bin_corr)
+                    
+                    cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_red_context, input_height, input_width))
+                    cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
+                                resize_image(cv2.imread(dir_of_label_file), input_height, input_width))
+                    
+                    indexer += 1
+                        
+                    
                     
+                    
+        if patches:
             indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
-                                  cv2.imread(dir_img + '/' + im), cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                  cv2.imread(dir_img + '/' + im), cv2.imread(dir_of_label_file),
                                   input_height, input_width, indexer=indexer)
             
             if augmentation:
-
                 if rotation:
                     indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
-                                          rotation_90(cv2.imread(dir_img + '/' + im)),
+                                        rotation_90(cv2.imread(dir_img + '/' + im)),
-                                          rotation_90(cv2.imread(dir_seg + '/' + img_name + '.png')),
+                                        rotation_90(cv2.imread(dir_of_label_file)),
-                                          input_height, input_width, indexer=indexer)
+                                        input_height, input_width, indexer=indexer)
                     
                 if rotation_not_90:
-
                     for thetha_i in thetha:
-                        img_max_rotated, label_max_rotated = rotation_not_90_func(cv2.imread(dir_img + '/' + im),
+                        img_max_rotated, label_max_rotated = rotation_not_90_func(cv2.imread(dir_img + '/'+im),
-                                                                                  cv2.imread(
+                                                                                  cv2.imread(dir_of_label_file), thetha_i)
-                                                                                      dir_seg + '/' + img_name + '.png'),
-                                                                                  thetha_i)
                         indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
                                               img_max_rotated,
                                               label_max_rotated,
                                               input_height, input_width, indexer=indexer)
+                        
+                if channels_shuffling:
+                    for shuffle_index in shuffle_indexes:
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                              return_shuffled_channels(cv2.imread(dir_img + '/' + im), shuffle_index),
+                                              cv2.imread(dir_of_label_file),
+                                              input_height, input_width, indexer=indexer)
+                        
+                if adding_rgb_background:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    for i_n in range(number_of_backgrounds_per_image):
+                        background_image_chosen_name = random.choice(list_all_possible_background_images)
+                        img_rgb_background_chosen = cv2.imread(dir_rgb_backgrounds + '/' + background_image_chosen_name)
+                        img_with_overlayed_background = return_binary_image_with_given_rgb_background(img_bin_corr, img_rgb_background_chosen)
+                        
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                              img_with_overlayed_background,
+                                              cv2.imread(dir_of_label_file),
+                                              input_height, input_width, indexer=indexer)
+                        
+                        
+                if adding_rgb_foreground:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    for i_n in range(number_of_backgrounds_per_image):
+                        background_image_chosen_name = random.choice(list_all_possible_background_images)
+                        foreground_rgb_chosen_name = random.choice(list_all_possible_foreground_rgbs)
+                        
+                        img_rgb_background_chosen = cv2.imread(dir_rgb_backgrounds + '/' + background_image_chosen_name)
+                        foreground_rgb_chosen = np.load(dir_rgb_foregrounds + '/' + foreground_rgb_chosen_name)
+                        
+                        img_with_overlayed_background = return_binary_image_with_given_rgb_background_and_given_foreground_rgb(img_bin_corr, img_rgb_background_chosen, foreground_rgb_chosen)
+                        
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                              img_with_overlayed_background,
+                                              cv2.imread(dir_of_label_file),
+                                              input_height, input_width, indexer=indexer)
+                        
+                        
+                if add_red_textlines:
+                    img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
+                    img_red_context = return_image_with_red_elements(cv2.imread(dir_img + '/'+im), img_bin_corr)
+                    
+                    indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                            img_red_context,
+                                            cv2.imread(dir_of_label_file),
+                                            input_height, input_width, indexer=indexer)
+                
                 if flip_aug:
                     for f_i in flip_index:
                         indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
                                               cv2.flip(cv2.imread(dir_img + '/' + im), f_i),
-                                              cv2.flip(cv2.imread(dir_seg + '/' + img_name + '.png'), f_i),
+                                              cv2.flip(cv2.imread(dir_of_label_file), f_i),
                                               input_height, input_width, indexer=indexer)
                 if blur_aug:   
                     for blur_i in blur_k:
                         indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
                                               bluring(cv2.imread(dir_img + '/' + im), blur_i),
-                                              cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                              cv2.imread(dir_of_label_file),
                                               input_height, input_width, indexer=indexer)          
+                if padding_black:
+                    indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                          do_padding_black(cv2.imread(dir_img + '/' + im)),
+                                          do_padding_label(cv2.imread(dir_of_label_file)),
+                                          input_height, input_width, indexer=indexer)       
         
+                if padding_white:   
+                    indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                          do_padding_white(cv2.imread(dir_img + '/'+im)),
+                                          do_padding_label(cv2.imread(dir_of_label_file)),
+                                          input_height, input_width, indexer=indexer)       
+                    
+                if brightening:
+                    for factor in brightness:
+                        try:
+                            indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                                  do_brightening(dir_img + '/' +im, factor),
+                                                  cv2.imread(dir_of_label_file),
+                                                  input_height, input_width, indexer=indexer)
+                        except:
+                            pass
                 if scaling:  
                     for sc_ind in scales:
                         indexer = get_patches_num_scale_new(dir_flow_train_imgs, dir_flow_train_labels,
-                                                            cv2.imread(dir_img + '/' + im),
+                                                            cv2.imread(dir_img + '/' + im) ,
-                                                            cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                                            cv2.imread(dir_of_label_file),
                                                             input_height, input_width, indexer=indexer, scaler=sc_ind)
+                        
+                if degrading:  
+                    for degrade_scale_ind in degrade_scales:
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                              do_degrading(cv2.imread(dir_img + '/' + im), degrade_scale_ind),
+                                              cv2.imread(dir_of_label_file),
+                                              input_height, input_width, indexer=indexer)
+                        
                 if binarization:
-                    indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                    if dir_img_bin:
-                                          otsu_copy(cv2.imread(dir_img + '/' + im)),
+                        img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
-                                          cv2.imread(dir_seg + '/' + img_name + '.png'),
+                        
-                                          input_height, input_width, indexer=indexer)
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                            img_bin_corr,
+                                            cv2.imread(dir_of_label_file),
+                                            input_height, input_width, indexer=indexer)
+                        
+                    else:
+                        indexer = get_patches(dir_flow_train_imgs, dir_flow_train_labels,
+                                            otsu_copy(cv2.imread(dir_img + '/' + im)),
+                                            cv2.imread(dir_of_label_file),
+                                            input_height, input_width, indexer=indexer)
+
+                if scaling_brightness:
+                    for sc_ind in scales:
+                        for factor in brightness:
+                            try:
+                                indexer = get_patches_num_scale_new(dir_flow_train_imgs,
+                                                                    dir_flow_train_labels,
+                                                                    do_brightening(dir_img + '/' + im, factor)
+                                                                    ,cv2.imread(dir_of_label_file)
+                                                                    ,input_height, input_width, indexer=indexer, scaler=sc_ind)
+                            except:
+                                pass
                         
                 if scaling_bluring:  
                     for sc_ind in scales:
                         for blur_i in blur_k:
                             indexer = get_patches_num_scale_new(dir_flow_train_imgs, dir_flow_train_labels,
                                                                 bluring(cv2.imread(dir_img + '/' + im), blur_i),
-                                                                cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                                                cv2.imread(dir_of_label_file),
-                                                                input_height, input_width, indexer=indexer,
+                                                                input_height, input_width, indexer=indexer, scaler=sc_ind)
-                                                                scaler=sc_ind)
 
                 if scaling_binarization:  
                     for sc_ind in scales:
                         indexer = get_patches_num_scale_new(dir_flow_train_imgs, dir_flow_train_labels,
                                                             otsu_copy(cv2.imread(dir_img + '/' + im)),
-                                                            cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                                            cv2.imread(dir_of_label_file),
                                                             input_height, input_width, indexer=indexer, scaler=sc_ind)
                         
                 if scaling_flip:  
                     for sc_ind in scales:
                         for f_i in flip_index:
                             indexer = get_patches_num_scale_new(dir_flow_train_imgs, dir_flow_train_labels,
-                                                                cv2.flip(cv2.imread(dir_img + '/' + im), f_i),
+                                                                cv2.flip( cv2.imread(dir_img + '/' + im), f_i),
-                                                                cv2.flip(cv2.imread(dir_seg + '/' + img_name + '.png'),
+                                                                cv2.flip(cv2.imread(dir_of_label_file), f_i),
-                                                                         f_i),
+                                                                input_height, input_width, indexer=indexer, scaler=sc_ind)
-                                                                input_height, input_width, indexer=indexer,
-                                                                scaler=sc_ind)