new (hybrid cnn+transformer) textline model which can accelerate to extract contour textlines faster

qurator/eynollah/cli.py

@@ -76,7 +76,13 @@ from qurator.eynollah.eynollah import Eynollah
     "--curved-line/--no-curvedline",
     "-cl/-nocl",
     is_flag=True,
-    help="if this parameter set to true, this tool will try to return contoure of textlines instead of rectabgle bounding box of textline. This should be taken into account that with this option the tool need more time to do process.",
+    help="if this parameter set to true, this tool will try to return contoure of textlines instead of rectangle bounding box of textline. This should be taken into account that with this option the tool need more time to do process.",
+)
+@click.option(
+    "--textline_light/--no-textline_light",
+    "-tll/-notll",
+    is_flag=True,
+    help="if this parameter set to true, this tool will try to return contoure of textlines instead of rectangle bounding box of textline with a faster method.",
 )
 @click.option(
     "--full-layout/--no-full-layout",
@@ -139,6 +145,7 @@ def main(
     enable_plotting,
     allow_enhancement,
     curved_line,
+    textline_light,
     full_layout,
     tables,
     input_binary,
@@ -170,6 +177,7 @@ def main(
         enable_plotting=enable_plotting,
         allow_enhancement=allow_enhancement,
         curved_line=curved_line,
+        textline_light=textline_light,
         full_layout=full_layout,
         tables=tables,
         input_binary=input_binary,

qurator/eynollah/eynollah.py

@@ -30,6 +30,7 @@ from scipy.signal import find_peaks
 import matplotlib.pyplot as plt
 from scipy.ndimage import gaussian_filter1d
 from keras.backend import set_session
+from tensorflow.keras import layers
 
 from .utils.contour import (
     filter_contours_area_of_image,
@@ -83,6 +84,60 @@ DPI_THRESHOLD = 298
 MAX_SLOPE = 999
 KERNEL = np.ones((5, 5), np.uint8)
 
+projection_dim = 64
+patch_size = 1
+num_patches =21*21#14*14#28*28#14*14#28*28
+
+
+class Patches(layers.Layer):
+    def __init__(self, **kwargs):
+        super(Patches, self).__init__()
+        self.patch_size = patch_size
+
+    def call(self, images):
+        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]
+        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, **kwargs):
+        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
+
 class Eynollah:
     def __init__(
         self,
@@ -100,6 +155,7 @@ class Eynollah:
         enable_plotting=False,
         allow_enhancement=False,
         curved_line=False,
+        textline_light=False,
         full_layout=False,
         tables=False,
         input_binary=False,
@@ -130,6 +186,7 @@ class Eynollah:
         self.enable_plotting = enable_plotting
         self.allow_enhancement = allow_enhancement
         self.curved_line = curved_line
+        self.textline_light = textline_light
         self.full_layout = full_layout
         self.tables = tables
         self.input_binary = input_binary
@@ -151,6 +208,7 @@ class Eynollah:
                 dir_out=self.dir_out,
                 image_filename=self.image_filename,
                 curved_line=self.curved_line,
+                textline_light = self.textline_light,
                 pcgts=pcgts)
         self.logger = logger if logger else getLogger('eynollah')
         self.dir_models = dir_models
@@ -165,6 +223,9 @@ class Eynollah:
         self.model_page_dir = dir_models + "/eynollah-page-extraction_20210425.h5"
         self.model_region_dir_p_ens = dir_models + "/eynollah-main-regions-ensembled_20210425.h5"
         self.model_region_dir_p_ens_light = dir_models + "/eynollah-main-regions_20220314.h5"
+        if self.textline_light:
+            self.model_textline_dir = dir_models + "/model_17.h5"
+        else:
             self.model_textline_dir = dir_models + "/eynollah-textline_20210425.h5"
         self.model_tables = dir_models + "/eynollah-tables_20210319.h5"
         
@@ -603,7 +664,10 @@ class Eynollah:
         gpu_options = tf.compat.v1.GPUOptions(allow_growth=True)
         #gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=7.7, allow_growth=True)
         session = tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(gpu_options=gpu_options))
+        try:
             model = load_model(model_dir, compile=False)
+        except:
+            model = load_model(model_dir , compile=False,custom_objects = {"PatchEncoder": PatchEncoder, "Patches": Patches})
 
         return model, session
 
@@ -1368,10 +1432,15 @@ class Eynollah:
 
             # plt.imshow(mask_only_con_region)
             # plt.show()
+            
+            if self.textline_light:
+                all_text_region_raw = np.copy(textline_mask_tot_ea)
+                all_text_region_raw[mask_only_con_region == 0] = 0
+                cnt_clean_rot_raw, hir_on_cnt_clean_rot = return_contours_of_image(all_text_region_raw)
+                cnt_clean_rot = filter_contours_area_of_image(all_text_region_raw, cnt_clean_rot_raw, hir_on_cnt_clean_rot, max_area=1, min_area=0.00001)
+            else:
                 all_text_region_raw = np.copy(textline_mask_tot_ea[boxes_text[mv][1] : boxes_text[mv][1] + boxes_text[mv][3], boxes_text[mv][0] : boxes_text[mv][0] + boxes_text[mv][2]])
                 mask_only_con_region = mask_only_con_region[boxes_text[mv][1] : boxes_text[mv][1] + boxes_text[mv][3], boxes_text[mv][0] : boxes_text[mv][0] + boxes_text[mv][2]]
-
-
                 all_text_region_raw[mask_only_con_region == 0] = 0
                 cnt_clean_rot = textline_contours_postprocessing(all_text_region_raw, [slope_deskew][0], contours_par_per_process[mv], boxes_text[mv])
 
@@ -1481,7 +1550,9 @@ class Eynollah:
         if not self.dir_in:
             session_textline.close()
 
-
+        if self.textline_light:
+            return (prediction_textline[:, :, 0]==1)*1, (prediction_textline_longshot_true_size[:, :, 0]==1)*1
+        else:
             return prediction_textline[:, :, 0], prediction_textline_longshot_true_size[:, :, 0]
 
     def do_work_of_slopes(self, q, poly, box_sub, boxes_per_process, textline_mask_tot, contours_per_process):
@@ -2562,6 +2633,8 @@ class Eynollah:
         scaler_h_textline = 1  # 1.2#1.2
         scaler_w_textline = 1  # 0.9#1
         textline_mask_tot_ea, _ = self.textline_contours(image_page, True, scaler_h_textline, scaler_w_textline)
+        if self.textline_light:
+            textline_mask_tot_ea = textline_mask_tot_ea.astype(np.int16)
         if not self.dir_in:
             K.clear_session()
         if self.plotter:
@@ -2870,7 +2943,6 @@ class Eynollah:
             self.ls_imgs = [1]
         
         for img_name in self.ls_imgs:
-            print(img_name,'img_name')
             t0 = time.time()
             if self.dir_in:
                 self.reset_file_name_dir(os.path.join(self.dir_in,img_name))
@@ -2887,6 +2959,7 @@ class Eynollah:
                 num_col, num_col_classifier, img_only_regions, page_coord, image_page, mask_images, mask_lines, text_regions_p_1, cont_page, table_prediction, textline_mask_tot_ea = \
                         self.run_graphics_and_columns_light(text_regions_p_1, textline_mask_tot_ea, num_col_classifier, num_column_is_classified, erosion_hurts)
                 #self.logger.info("run graphics %.1fs ", time.time() - t1t)
+                textline_mask_tot_ea_org = np.copy(textline_mask_tot_ea)
             else:
                 text_regions_p_1 ,erosion_hurts, polygons_lines_xml = self.get_regions_from_xy_2models(img_res, is_image_enhanced, num_col_classifier)
                 self.logger.info("Textregion detection took %.1fs ", time.time() - t1)
@@ -3043,6 +3116,10 @@ class Eynollah:
             
             if not self.curved_line:
                 if self.light_version:
+                    if self.textline_light:
+                        slopes, all_found_texline_polygons, boxes_text, txt_con_org, contours_only_text_parent, all_box_coord, index_by_text_par_con = self.get_slopes_and_deskew_new_light(txt_con_org, contours_only_text_parent, textline_mask_tot_ea_org, image_page_rotated, boxes_text, slope_deskew)
+                        slopes_marginals, all_found_texline_polygons_marginals, boxes_marginals, _, polygons_of_marginals, all_box_coord_marginals, _ = self.get_slopes_and_deskew_new_light(polygons_of_marginals, polygons_of_marginals, textline_mask_tot_ea_org, image_page_rotated, boxes_marginals, slope_deskew)
+                    else:
                         slopes, all_found_texline_polygons, boxes_text, txt_con_org, contours_only_text_parent, all_box_coord, index_by_text_par_con = self.get_slopes_and_deskew_new_light(txt_con_org, contours_only_text_parent, textline_mask_tot_ea, image_page_rotated, boxes_text, slope_deskew)
                         slopes_marginals, all_found_texline_polygons_marginals, boxes_marginals, _, polygons_of_marginals, all_box_coord_marginals, _ = self.get_slopes_and_deskew_new_light(polygons_of_marginals, polygons_of_marginals, textline_mask_tot_ea, image_page_rotated, boxes_marginals, slope_deskew)
                 else:

qurator/eynollah/writer.py

@@ -22,12 +22,13 @@ import numpy as np
 
 class EynollahXmlWriter():
 
-    def __init__(self, *, dir_out, image_filename, curved_line, pcgts=None):
+    def __init__(self, *, dir_out, image_filename, curved_line,textline_light, pcgts=None):
         self.logger = getLogger('eynollah.writer')
         self.counter = EynollahIdCounter()
         self.dir_out = dir_out
         self.image_filename = image_filename
         self.curved_line = curved_line
+        self.textline_light = textline_light
         self.pcgts = pcgts
         self.scale_x = None # XXX set outside __init__
         self.scale_y = None # XXX set outside __init__
@@ -60,7 +61,7 @@ class EynollahXmlWriter():
             marginal_region.add_TextLine(textline)
             points_co = ''
             for l in range(len(all_found_texline_polygons_marginals[marginal_idx][j])):
-                if not self.curved_line:
+                if not (self.curved_line or self.textline_light):
                     if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
                         textline_x_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x) )
                         textline_y_coord = max(0, int((all_found_texline_polygons_marginals[marginal_idx][j][l][1] + all_box_coord_marginals[marginal_idx][0] + page_coord[0]) / self.scale_y) )
@@ -70,7 +71,7 @@ class EynollahXmlWriter():
                     points_co += str(textline_x_coord)
                     points_co += ','
                     points_co += str(textline_y_coord)
-                if self.curved_line and np.abs(slopes_marginals[marginal_idx]) <= 45:
+                if (self.curved_line or self.textline_light) and np.abs(slopes_marginals[marginal_idx]) <= 45:
                     if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
                         points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + page_coord[2]) / self.scale_x))
                         points_co += ','
@@ -80,7 +81,7 @@ class EynollahXmlWriter():
                         points_co += ','
                         points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0][1] + page_coord[0]) / self.scale_y))
 
-                elif self.curved_line and np.abs(slopes_marginals[marginal_idx]) > 45:
+                elif (self.curved_line or self.textline_light) and np.abs(slopes_marginals[marginal_idx]) > 45:
                     if len(all_found_texline_polygons_marginals[marginal_idx][j][l]) == 2:
                         points_co += str(int((all_found_texline_polygons_marginals[marginal_idx][j][l][0] + all_box_coord_marginals[marginal_idx][2] + page_coord[2]) / self.scale_x))
                         points_co += ','
@@ -101,7 +102,7 @@ class EynollahXmlWriter():
             region_bboxes = all_box_coord[region_idx]
             points_co = ''
             for idx_contour_textline, contour_textline in enumerate(all_found_texline_polygons[region_idx][j]):
-                if not self.curved_line:
+                if not (self.curved_line or self.textline_light):
                     if len(contour_textline) == 2:
                         textline_x_coord = max(0, int((contour_textline[0] + region_bboxes[2] + page_coord[2]) / self.scale_x))
                         textline_y_coord = max(0, int((contour_textline[1] + region_bboxes[0] + page_coord[0]) / self.scale_y))
@@ -112,7 +113,7 @@ class EynollahXmlWriter():
                     points_co += ','
                     points_co += str(textline_y_coord)
 
-                if self.curved_line and np.abs(slopes[region_idx]) <= 45:
+                if (self.curved_line or self.textline_light) and np.abs(slopes[region_idx]) <= 45:
                     if len(contour_textline) == 2:
                         points_co += str(int((contour_textline[0] + page_coord[2]) / self.scale_x))
                         points_co += ','
@@ -121,7 +122,7 @@ class EynollahXmlWriter():
                         points_co += str(int((contour_textline[0][0] + page_coord[2]) / self.scale_x))
                         points_co += ','
                         points_co += str(int((contour_textline[0][1] + page_coord[0])/self.scale_y))
-                elif self.curved_line and np.abs(slopes[region_idx]) > 45:
+                elif (self.curved_line or self.textline_light) and np.abs(slopes[region_idx]) > 45:
                     if len(contour_textline)==2:
                         points_co += str(int((contour_textline[0] + region_bboxes[2] + page_coord[2])/self.scale_x))
                         points_co += ','

requirements.txt

@@ -1,8 +1,8 @@
 # ocrd includes opencv, numpy, shapely, click
-ocrd >= 2.23.3
+ocrd
-keras >= 2.3.1, < 2.4
+keras == 2.6.0
-scikit-learn >= 0.23.2
+scikit-learn
-tensorflow-gpu >= 1.15, < 2
+tensorflow-gpu == 2.6.0
-imutils >= 0.5.3
+imutils
 matplotlib
-setuptools >= 50
+setuptools