mirror of
https://github.com/qurator-spk/eynollah.git
synced 2026-02-20 16:32:03 +01:00
training: simplify image preprocessing…
- `utils.provide_patches`: split up loop into * `utils.preprocess_img` (single img function) * `utils.preprocess_imgs` (top-level loop) - capture exceptions for all cases (not just some) at top level and with informative logging - avoid repeating / delegating config keys in several places: only as kwargs to `preprocess_img()` - read files into memory only once, then re-use - improve readability (avoiding long lines, repeated code)
This commit is contained in:
Robert Sachunsky
parent
e69b35b49c
commit
29a0f19cee
2 changed files with 510 additions and 370 deletions
|
|
@ -26,7 +26,7 @@ from eynollah.training.utils import (
|
|||
generate_data_from_folder_evaluation,
|
||||
generate_data_from_folder_training,
|
||||
get_one_hot,
|
||||
provide_patches,
|
||||
preprocess_imgs,
|
||||
return_number_of_total_training_data
|
||||
)
|
||||
|
||||
|
|
@ -240,9 +240,9 @@ def run(_config,
|
|||
os.mkdir(dir_flow_eval_imgs)
|
||||
os.mkdir(dir_flow_eval_labels)
|
||||
|
||||
# set the gpu configuration
|
||||
configuration()
|
||||
|
||||
dir_img, dir_seg = get_dirs_or_files(dir_train)
|
||||
dir_img_val, dir_seg_val = get_dirs_or_files(dir_eval)
|
||||
|
||||
imgs_list=np.array(os.listdir(dir_img))
|
||||
segs_list=np.array(os.listdir(dir_seg))
|
||||
|
||||
|
|
@ -250,50 +250,21 @@ def run(_config,
|
|||
segs_list_test=np.array(os.listdir(dir_seg_val))
|
||||
|
||||
# writing patches into a sub-folder in order to be flowed from directory.
|
||||
common_args = [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, shifting, degrading, brightening,
|
||||
scales, degrade_scales, brightness,
|
||||
flip_index, shuffle_indexes,
|
||||
scaling_bluring, scaling_brightness, scaling_binarization,
|
||||
rotation, rotation_not_90, thetha,
|
||||
scaling_flip, task,
|
||||
]
|
||||
common_kwargs = dict(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, segs_list,
|
||||
dir_img, dir_seg,
|
||||
preprocess_imgs(_config,
|
||||
imgs_list,
|
||||
segs_list,
|
||||
dir_img,
|
||||
dir_seg,
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
*common_args,
|
||||
augmentation=augmentation,
|
||||
**common_kwargs)
|
||||
provide_patches(imgs_list_test, segs_list_test,
|
||||
dir_img_val, dir_seg_val,
|
||||
dir_flow_train_labels)
|
||||
preprocess_imgs(_config,
|
||||
imgs_list_test,
|
||||
segs_list_test,
|
||||
dir_img_val,
|
||||
dir_seg_val,
|
||||
dir_flow_eval_imgs,
|
||||
dir_flow_eval_labels,
|
||||
*common_args,
|
||||
augmentation=False,
|
||||
**common_kwargs)
|
||||
augmentation=False)
|
||||
|
||||
if weighted_loss:
|
||||
weights = np.zeros(n_classes)
|
||||
|
|
@ -307,8 +278,8 @@ def run(_config,
|
|||
label_obj = cv2.imread(label_file)
|
||||
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 Exception as e:
|
||||
print("error reading data file '%s': %s" % (label_file, e), file=sys.stderr)
|
||||
except Exception:
|
||||
_log.exception("error reading data file '%s'", label_file)
|
||||
|
||||
weights = 1.00 / weights
|
||||
weights = weights / float(np.sum(weights))
|
||||
|
|
@ -340,7 +311,6 @@ def run(_config,
|
|||
custom_objects = {"PatchEncoder": PatchEncoder,
|
||||
"Patches": Patches})
|
||||
else:
|
||||
index_start = 0
|
||||
if backbone_type == 'nontransformer':
|
||||
model = resnet50_unet(n_classes,
|
||||
input_height,
|
||||
|
|
@ -391,7 +361,7 @@ def run(_config,
|
|||
pretraining)
|
||||
|
||||
#if you want to see the model structure just uncomment model summary.
|
||||
model.summary()
|
||||
#model.summary()
|
||||
|
||||
if task in ["segmentation", "binarization"]:
|
||||
if is_loss_soft_dice:
|
||||
|
|
@ -423,7 +393,12 @@ def run(_config,
|
|||
SaveWeightsAfterSteps(0, dir_output, _config)]
|
||||
if save_interval:
|
||||
callbacks.append(SaveWeightsAfterSteps(save_interval, dir_output, _config))
|
||||
|
||||
|
||||
_log.info("training on %d batches in %d epochs",
|
||||
len(os.listdir(dir_flow_train_imgs)) // n_batch - 1,
|
||||
n_epochs)
|
||||
_log.info("validating on %d batches",
|
||||
len(os.listdir(dir_flow_eval_imgs)) // n_batch - 1)
|
||||
model.fit(
|
||||
train_gen,
|
||||
steps_per_epoch=len(os.listdir(dir_flow_train_imgs)) // n_batch - 1,
|
||||
|
|
@ -439,7 +414,6 @@ def run(_config,
|
|||
#model.save(dir_output+'/'+'model'+'.h5')
|
||||
|
||||
elif task=='classification':
|
||||
configuration()
|
||||
model = resnet50_classifier(n_classes,
|
||||
input_height,
|
||||
input_width,
|
||||
|
|
@ -474,7 +448,7 @@ def run(_config,
|
|||
|
||||
usable_checkpoints = np.flatnonzero(np.array(history['val_f1']) > f1_threshold_classification)
|
||||
if len(usable_checkpoints) >= 1:
|
||||
print("averaging over usable checkpoints", usable_checkpoints)
|
||||
_log.info("averaging over usable checkpoints: %s", str(usable_checkpoints))
|
||||
all_weights = []
|
||||
for epoch in usable_checkpoints:
|
||||
cp_path = os.path.join(dir_output, 'model_{epoch:02d}'.format(epoch=epoch))
|
||||
|
|
@ -495,10 +469,9 @@ def run(_config,
|
|||
model.save(cp_path)
|
||||
with open(os.path.join(cp_path, "config.json"), "w") as fp:
|
||||
json.dump(_config, fp) # encode dict into JSON
|
||||
print("ensemble model saved under", cp_path)
|
||||
_log.info("ensemble model saved under '%s'", cp_path)
|
||||
|
||||
elif task=='reading_order':
|
||||
configuration()
|
||||
model = machine_based_reading_order_model(
|
||||
n_classes, input_height, input_width, weight_decay, pretraining)
|
||||
|
||||
|
|
|
|||
|
|
@ -1,6 +1,7 @@
|
|||
import os
|
||||
import math
|
||||
import random
|
||||
from logging import getLogger
|
||||
|
||||
import cv2
|
||||
import numpy as np
|
||||
|
|
@ -266,8 +267,9 @@ def generate_data_from_folder_training(path_classes, batchsize, height, width, n
|
|||
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)
|
||||
def do_brightening(img, factor):
|
||||
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
||||
im = Image.fromarray(img_rgb)
|
||||
enhancer = ImageEnhance.Brightness(im)
|
||||
out_img = enhancer.enhance(factor)
|
||||
out_img = out_img.convert('RGB')
|
||||
|
|
@ -737,321 +739,486 @@ def get_patches_num_scale_new(dir_img_f, dir_seg_f, img, label, height, width, i
|
|||
return indexer
|
||||
|
||||
|
||||
def provide_patches(imgs_list_train, segs_list_train, 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, shifting, 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=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):
|
||||
|
||||
def preprocess_imgs(config,
|
||||
imgs_list,
|
||||
segs_list,
|
||||
dir_img,
|
||||
dir_seg,
|
||||
dir_flow_imgs,
|
||||
dir_flow_labels,
|
||||
logger=None,
|
||||
**kwargs,
|
||||
):
|
||||
if logger is None:
|
||||
logger = getLogger('')
|
||||
|
||||
# make a copy for this run
|
||||
config = dict(config)
|
||||
# add derived keys not part of config
|
||||
if config.get('dir_rgb_backgrounds', None):
|
||||
config['list_all_possible_background_images'] = \
|
||||
os.listdir(config['dir_rgb_backgrounds'])
|
||||
if config.get('dir_rgb_foregrounds', None):
|
||||
config['list_all_possible_foreground_rgbs'] = \
|
||||
os.listdir(config['dir_rgb_foregrounds'])
|
||||
# override keys from call
|
||||
config.update(kwargs)
|
||||
|
||||
indexer = 0
|
||||
for im, seg_i in tqdm(zip(imgs_list_train, segs_list_train)):
|
||||
for im, seg_i in tqdm(zip(imgs_list, segs_list)):
|
||||
img = cv2.imread(os.path.join(dir_img, im))
|
||||
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', 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))
|
||||
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,input_width) )
|
||||
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(cv2.flip(cv2.imread(dir_of_label_file), f_i), 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, 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 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:
|
||||
|
||||
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_of_label_file), input_height, input_width))
|
||||
indexer += 1
|
||||
|
||||
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 shifting:
|
||||
shift_types = ['xpos', 'xmin', 'ypos', 'ymin', 'xypos', 'xymin']
|
||||
for st_ind in shift_types:
|
||||
img_shifted, label_shifted = shift_image_and_label(cv2.imread(dir_img + '/'+im),
|
||||
cv2.imread(dir_of_label_file), st_ind)
|
||||
|
||||
cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png', resize_image(img_shifted, input_height, input_width))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png', resize_image(label_shifted, 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_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_of_label_file)),
|
||||
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),
|
||||
cv2.imread(dir_of_label_file), 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_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_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_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:
|
||||
if dir_img_bin:
|
||||
img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
|
||||
|
||||
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 config['task'] in ["segmentation", "binarization"]:
|
||||
lab = cv2.imread(os.path.join(dir_seg, img_name + '.png'))
|
||||
elif config['task'] == "enhancement":
|
||||
lab = cv2.imread(os.path.join(dir_seg, im))
|
||||
else:
|
||||
lab = None
|
||||
|
||||
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_of_label_file),
|
||||
input_height, input_width, indexer=indexer, scaler=sc_ind)
|
||||
try:
|
||||
indexer = preprocess_img(indexer, img, img_name, lab,
|
||||
dir_flow_imgs,
|
||||
dir_flow_labels,
|
||||
**config)
|
||||
|
||||
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_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_of_label_file), f_i),
|
||||
input_height, input_width, indexer=indexer, scaler=sc_ind)
|
||||
except:
|
||||
logger.exception("skipping image %s", img_name)
|
||||
|
||||
def preprocess_img(indexer,
|
||||
img,
|
||||
img_name,
|
||||
lab,
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
input_height=None,
|
||||
input_width=None,
|
||||
augmentation=False,
|
||||
flip_aug=False,
|
||||
flip_index=None,
|
||||
blur_aug=False,
|
||||
blur_k=None,
|
||||
padding_white=False,
|
||||
padding_black=False,
|
||||
scaling=False,
|
||||
scaling_bluring=False,
|
||||
scaling_brightness=False,
|
||||
scaling_binarization=False,
|
||||
scaling_flip=False,
|
||||
scales=None,
|
||||
shifting=False,
|
||||
degrading=False,
|
||||
degrade_scales=None,
|
||||
brightening=False,
|
||||
brightness=None,
|
||||
binarization=False,
|
||||
dir_img_bin=None,
|
||||
add_red_textlines=False,
|
||||
adding_rgb_background=False,
|
||||
dir_rgb_backgrounds=None,
|
||||
adding_rgb_foreground=False,
|
||||
dir_rgb_foregrounds=None,
|
||||
number_of_backgrounds_per_image=None,
|
||||
channels_shuffling=False,
|
||||
shuffle_indexes=None,
|
||||
rotation=False,
|
||||
rotation_not_90=False,
|
||||
thetha=None,
|
||||
patches=False,
|
||||
list_all_possible_background_images=None,
|
||||
list_all_possible_foreground_rgbs=None,
|
||||
**kwargs,
|
||||
):
|
||||
if not patches:
|
||||
cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
|
||||
resize_image(img,
|
||||
input_height,
|
||||
input_width))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
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(img, f_i),
|
||||
input_height,
|
||||
input_width))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(cv2.flip(lab, f_i),
|
||||
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(img, blur_i),
|
||||
input_height,
|
||||
input_width)))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
input_height,
|
||||
input_width))
|
||||
indexer += 1
|
||||
if brightening:
|
||||
for factor in brightness:
|
||||
cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
|
||||
(resize_image(do_brightening(img, factor),
|
||||
input_height,
|
||||
input_width)))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
input_height,
|
||||
input_width))
|
||||
indexer += 1
|
||||
if binarization:
|
||||
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(img),
|
||||
input_height,
|
||||
input_width))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
input_height,
|
||||
input_width))
|
||||
indexer += 1
|
||||
if degrading:
|
||||
for degrade_scale_ind in degrade_scales:
|
||||
cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
|
||||
(resize_image(do_degrading(img, degrade_scale_ind),
|
||||
input_height,
|
||||
input_width)))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
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(img, lab, 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(img, shuffle_index),
|
||||
input_height,
|
||||
input_width)))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(lab,
|
||||
input_height,
|
||||
input_width))
|
||||
indexer += 1
|
||||
if scaling:
|
||||
for sc_ind in scales:
|
||||
img_scaled, label_scaled = \
|
||||
scale_image_for_no_patch(img, lab, 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 shifting:
|
||||
shift_types = ['xpos', 'xmin', 'ypos', 'ymin', 'xypos', 'xymin']
|
||||
for st_ind in shift_types:
|
||||
img_shifted, label_shifted = \
|
||||
shift_image_and_label(img, lab, st_ind)
|
||||
cv2.imwrite(dir_flow_train_imgs + '/img_' + str(indexer) + '.png',
|
||||
resize_image(img_shifted,
|
||||
input_height,
|
||||
input_width))
|
||||
cv2.imwrite(dir_flow_train_labels + '/img_' + str(indexer) + '.png',
|
||||
resize_image(label_shifted,
|
||||
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(lab,
|
||||
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(lab,
|
||||
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(img, 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(lab,
|
||||
input_height,
|
||||
input_width))
|
||||
indexer += 1
|
||||
else:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if augmentation:
|
||||
if rotation:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
rotation_90(img),
|
||||
rotation_90(lab),
|
||||
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(img, lab, 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:
|
||||
img_shuffled = \
|
||||
return_shuffled_channels(img, shuffle_index),
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_shuffled,
|
||||
lab,
|
||||
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,
|
||||
lab,
|
||||
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,
|
||||
lab,
|
||||
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(img, img_bin_corr)
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_red_context,
|
||||
lab,
|
||||
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(img, f_i),
|
||||
cv2.flip(lab, 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(img, blur_i),
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if padding_black:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
do_padding_black(img),
|
||||
do_padding_label(lab),
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if padding_white:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
do_padding_white(img),
|
||||
do_padding_label(lab),
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if brightening:
|
||||
for factor in brightness:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
do_brightening(img, factor),
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if scaling:
|
||||
for sc_ind in scales:
|
||||
indexer = get_patches_num_scale_new(
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img ,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer,
|
||||
scaler=sc_ind)
|
||||
if degrading:
|
||||
for degrade_scale_ind in degrade_scales:
|
||||
img_deg = \
|
||||
do_degrading(img, degrade_scale_ind),
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_deg,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if binarization:
|
||||
if dir_img_bin:
|
||||
img_bin_corr = cv2.imread(dir_img_bin + '/' + img_name+'.png')
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_bin_corr,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
else:
|
||||
indexer = get_patches(dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
otsu_copy(img),
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer)
|
||||
if scaling_brightness:
|
||||
for sc_ind in scales:
|
||||
for factor in brightness:
|
||||
img_bright = do_brightening(img, factor)
|
||||
indexer = get_patches_num_scale_new(
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_bright,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer,
|
||||
scaler=sc_ind)
|
||||
if scaling_bluring:
|
||||
for sc_ind in scales:
|
||||
for blur_i in blur_k:
|
||||
img_blur = bluring(img, blur_i),
|
||||
indexer = get_patches_num_scale_new(
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_blur,
|
||||
lab,
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer,
|
||||
scaler=sc_ind)
|
||||
if scaling_binarization:
|
||||
for sc_ind in scales:
|
||||
img_bin = otsu_copy(img),
|
||||
indexer = get_patches_num_scale_new(
|
||||
dir_flow_train_imgs,
|
||||
dir_flow_train_labels,
|
||||
img_bin,
|
||||
lab,
|
||||
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(img, f_i),
|
||||
cv2.flip(lab, f_i),
|
||||
input_height,
|
||||
input_width,
|
||||
indexer=indexer,
|
||||
scaler=sc_ind)
|
||||
return indexer
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue