def create_data_pick_csv_train_val(train_dir, train_ratio=0.92):
list_files = get_list_file_in_folder(os.path.join(train_dir, 'images'))
num_total = len(list_files)
num_train = int(num_total * train_ratio)
num_val = num_total - num_train
random.shuffle(list_files)
list_train = list_files[:num_train]
list_val = list_files[num_train + 1:]
train_txt_list = []
for idx, f in enumerate(list_train):
line = ','.join([str(idx + 1), 'receipts', f])
train_txt_list.append(line + '\n')
with open(os.path.join(train_dir, 'train_list.csv'), mode='w', encoding='utf-8') as f:
f.writelines(train_txt_list)
val_txt_list = []
for idx, f in enumerate(list_val):
line = ','.join([str(idx + 1), 'receipts', f])
val_txt_list.append(line + '\n')
with open(os.path.join(train_dir, 'val_list.csv'), mode='w', encoding='utf-8') as f:
f.writelines(val_txt_list)
print('Done')
def main():
begin_init = time.time()
global anno_path
if not os.path.exists(output_txt_dir):
os.makedirs(output_txt_dir)
if not os.path.exists(output_viz_dir):
os.makedirs(output_viz_dir)
if not os.path.exists(output_rotated_img_dir):
os.makedirs(output_rotated_img_dir)
box_rectify = init_box_rectify_model(weight_path)
end_init = time.time()
print('Init models time:', end_init - begin_init, 'seconds')
begin = time.time()
list_img_path = get_list_file_in_folder(img_dir)
list_img_path = sorted(list_img_path)
for idx, img_name in enumerate(list_img_path):
print('\n', idx, 'Inference', img_name)
test_img = cv2.imread(os.path.join(img_dir, img_name))
begin_detector = time.time()
anno_path = os.path.join(anno_dir, img_name.replace('.jpg', '.txt'))
boxes_list = get_list_boxes_from_icdar(anno_path)
boxes_list = drop_box(boxes_list, drop_gap=rot_drop_thresh)
rotation = get_mean_horizontal_angle(boxes_list, False)
img_rotated, boxes_list = rotate_image_bbox_angle(
test_img, boxes_list, rotation)
degre = calculate_page_orient(box_rectify, img_rotated, boxes_list)
img_rotated, boxes_list = rotate_image_bbox_angle(
img_rotated, boxes_list, degre)
boxes_list = filter_90_box(boxes_list)
end_detector = time.time()
print('get boxes from icdar time:', end_detector - begin_detector,
'seconds')
output_txt_path = os.path.join(
output_txt_dir,
os.path.basename(img_name).split('.')[0] + '.txt')
output_viz_path = os.path.join(output_viz_dir,
os.path.basename(img_name))
output_rotated_img_path = os.path.join(output_rotated_img_dir,
os.path.basename(img_name))
if write_rotated_img:
cv2.imwrite(output_rotated_img_path, img_rotated)
if write_file:
write_output(boxes_list, output_txt_path)
if visualize:
viz_icdar(img_rotated, output_txt_path, output_viz_path)
end_visualize = time.time()
print('Visualize time:', end_visualize - end_detector, 'seconds')
end = time.time()
speed = (end - begin) / len(list_img_path)
print('Processing time:', end - begin, 'seconds. Speed:', round(speed, 4),
'second/image')
print('Done')
def create_data_pick_boxes_and_transcripts(icdar_dir, output_dir):
list_file = get_list_file_in_folder(icdar_dir, ext=['txt'])
for idx, anno in enumerate(list_file):
print(idx, anno)
with open(os.path.join(icdar_dir, anno), mode='r', encoding='utf-8') as f:
list_bboxes = f.readlines()
for idx, line in enumerate(list_bboxes):
list_bboxes[idx] = str(idx + 1) + ',' + line
with open(os.path.join(output_dir, anno.replace('.txt', '.tsv')), mode='wt', encoding='utf-8') as f:
f.writelines(list_bboxes)
def viz_icdar_multi(img_dir, anno_dir, save_viz_dir, extract_kie_type=False, ignor_type=[1]):
list_files = get_list_file_in_folder(img_dir)
for idx, file in enumerate(list_files):
if idx < 0:
continue
# if 'mcocr_public_145014smasw' not in file:
# continue
print(idx, file)
img_path = os.path.join(img_dir, file)
anno_path = os.path.join(anno_dir, file.replace('.jpg', '.txt'))
save_img_path = os.path.join(save_viz_dir, file)
viz_icdar(img_path, anno_path, save_img_path, extract_kie_type, ignor_type)
def filter_space_OCR(ocr_dir):
list_files = get_list_file_in_folder(ocr_dir, ext=['txt'])
for idx, file in enumerate(list_files):
#print(idx, file)
with open(os.path.join(ocr_dir, file), 'r', encoding='utf-8') as f:
anno_txt = f.readlines()
final_res = ''
for anno in anno_txt:
fix_anno = anno.rstrip('\n').rstrip(' ')
if fix_anno != anno.rstrip('\n'):
print(idx, file, '. Fix space')
final_res += fix_anno + '\n'
final_res = final_res.rstrip('\n')
def compare_result(first_res_dir, second_res_dir, ext='txt'):
first_res_img_dir = first_res_dir.replace('/txt', '/imgs')
second_res_img_dir = second_res_dir.replace('/txt', '/imgs')
first_res_viz_dir = first_res_dir.replace('/txt', '/viz_imgs')
second_res_viz_dir = second_res_dir.replace('/txt', '/viz_imgs')
list_files = get_list_file_in_folder(first_res_dir, ext=[ext])
for idx, file in enumerate(list_files):
print(idx, file)
with open(os.path.join(first_res_dir, file),
mode='r',
encoding='utf-8') as f:
first_txt = f.readlines()
with open(os.path.join(second_res_dir, file),
mode='r',
encoding='utf-8') as f:
second_txt = f.readlines()
diff = False
first_img = cv2.imread(
os.path.join(first_res_img_dir, file.replace('.txt', '.jpg')))
second_img = cv2.imread(
os.path.join(second_res_img_dir, file.replace('.txt', '.jpg')))
first_viz_img = cv2.imread(
os.path.join(first_res_viz_dir, file.replace('.txt', '.jpg')))
second_viz_img = cv2.imread(
os.path.join(second_res_viz_dir, file.replace('.txt', '.jpg')))
for line in first_txt:
if line not in second_txt:
print('-------------------- not in second', line)
line = line.rstrip(',\n')
# coors = line.split('\t')[0]
# segment_pts = [int(f) for f in coors.split(',')]
# box = np.array(segment_pts).astype(np.int32).reshape(-1, 2)
# cv2.polylines(first_img, [box], True, color=(0, 255, 0), thickness=2)
diff = True
for line in second_txt:
if line not in first_txt:
print('-------------------- not in first', line)
line = line.rstrip(',\n')
# coors = line.split('\t')[0]
# segment_pts = [int(f) for f in coors.split(',')]
# box = np.array(segment_pts).astype(np.int32).reshape(-1, 2)
# cv2.polylines(second_img, [box], True, color=(255, 0, 0), thickness=2)
diff = True
if diff:
# cv2.imshow('first img', first_img)
# cv2.imshow('second img', second_img)
# cv2.imshow('first viz img', first_viz_img)
# cv2.imshow('second viz img', second_viz_img)
cv2.waitKey(0)
def viz_output_of_pick(img_dir, output_txt_dir, output_viz_dir):
list_output_txt = get_list_file_in_folder(output_txt_dir, ext=['txt'])
list_output_txt = sorted(list_output_txt)
for n, file in enumerate(list_output_txt):
print(n, file)
# if n <60:
# continue
with open(os.path.join(output_txt_dir, file), mode='r', encoding='utf-8') as f:
output_txt = f.readlines()
list_poly = []
for line in output_txt:
coordinates, type, text = line.replace('\n', '').split('\t')
find_poly = poly(coordinates, type=inv_type_map[type], value=text)
list_poly.append(find_poly)
img_name = file.replace('.txt', '.jpg')
image = cv2.imread(os.path.join(img_dir, img_name))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
viz_poly(img=image,
list_poly=list_poly,
save_viz_path=os.path.join(output_viz_dir, img_name))
def viz_same_img_in_different_dirs(first_dir, second_dir, list_path=None, resize_ratio=1.0):
list_files = get_list_file_in_folder(first_dir)
list_err_images = None
if list_path is not None:
list_err_images = get_list_error_imgs(list_path)
for n, file in enumerate(list_files):
if list_err_images is not None:
if file.replace('.jpg', '') not in list_err_images:
continue
print(n, file)
# if n<160:
# continue
first_img_path = os.path.join(first_dir, file)
first_img = cv2.imread(first_img_path)
first_img_res = cv2.resize(first_img,
(int(resize_ratio * first_img.shape[1]), int(resize_ratio * first_img.shape[0])))
cv2.imshow('first_img', first_img_res)
second_img_path = os.path.join(second_dir, file.replace('.jpg', '') + '_ model_epoch_639_minibatch_243000_ 0.1.jpg')
second_img = cv2.imread(second_img_path)
second_img_res = cv2.resize(second_img,
(int(resize_ratio * second_img.shape[1]), int(resize_ratio * second_img.shape[0])))
cv2.imshow('second_img', second_img_res)
cv2.waitKey(0)
def main():
begin_init = time.time()
global anno_path
classifier = init_models(gpu=gpu)
end_init = time.time()
print('Init models time:', end_init - begin_init, 'seconds')
begin = time.time()
list_img_path = []
if img_path != '':
list_img_path.append(img_path)
else:
list_img_path = get_list_file_in_folder(img_dir)
list_img_path = sorted(list_img_path)
for idx, img_name in enumerate(list_img_path):
if idx < 0:
continue
print('\n', idx, 'Inference', img_name)
test_img = cv2.imread(os.path.join(img_dir, img_name))
begin_detector = time.time()
if img_path == '':
anno_path = os.path.join(anno_dir,
img_name.replace('.jpg', '.txt'))
boxes_list = get_list_boxes_from_icdar(anno_path)
end_detector = time.time()
print('get boxes from icdar time:', end_detector - begin_detector,
'seconds')
# multiscale ocr
list_values = []
list_probs = []
total_boxes = len(boxes_list)
# 1 Extend x, no extend y
boxes_data = get_boxes_data(test_img,
boxes_list,
extend_box=True,
min_extend_y=0,
extend_y_ratio=0)
values, probs = classifier.inference(boxes_data, debug=False)
list_values.append(values)
list_probs.append(probs)
# 2 extend y by 10%
boxes_data = get_boxes_data(test_img,
boxes_list,
extend_box=True,
min_extend_y=2,
extend_y_ratio=0.1)
values, probs = classifier.inference(boxes_data, debug=False)
list_values.append(values)
list_probs.append(probs)
# 3 extend y by 20%
boxes_data = get_boxes_data(test_img,
boxes_list,
extend_box=True,
min_extend_y=4,
extend_y_ratio=0.2)
values, probs = classifier.inference(boxes_data, debug=False)
list_values.append(values)
list_probs.append(probs)
# combine final values and probs
final_values = []
final_probs = []
for idx in range(total_boxes):
max_prob = list_probs[0][idx]
max_value = list_values[0][idx]
for n in range(1, len(list_values)):
if list_probs[n][idx] > max_prob:
max_prob = list_probs[n][idx]
max_value = list_values[n][idx]
final_values.append(max_value)
final_probs.append(max_prob)
end_classifier = time.time()
print('Multiscale OCR time:', end_classifier - end_detector, 'seconds')
print('Total predict time:', end_classifier - begin_detector,
'seconds')
output_txt_path = os.path.join(
cls_out_txt_dir,
os.path.basename(img_name).split('.')[0] + '.txt')
output_viz_path = os.path.join(cls_out_viz_dir,
os.path.basename(img_name))
if write_file:
write_output(boxes_list,
final_values,
final_probs,
output_txt_path,
prob_thres=cls_ocr_thres)
if cls_visualize:
viz_icdar(os.path.join(img_dir, img_name),
output_txt_path,
output_viz_path,
ignor_type=[])
end_visualize = time.time()
print('Visualize time:', end_visualize - end_classifier, 'seconds')
end = time.time()
speed = (end - begin) / len(list_img_path)
print('\nTotal processing time:', end - begin, 'seconds. Speed:',
round(speed, 4), 'second/image')
def modify_kie_training_data_by_rules(txt_dir, json_data_path, debug=False):
list_files = get_list_file_in_folder(txt_dir, ext=['.txt'])
with open(json_data_path) as json_file:
data = json.load(json_file)
list_seller = data['seller']
list_address = data['address']
for idx, file in enumerate(list_files):
# with open(os.path.join(txt_dir, file), mode='r', encoding='utf-8') as f:
# anno_list= f.readlines()
#
# if 'mcocr_public_145014smasw' not in file:
# continue
# print(idx, file)
list_icdar_poly = get_list_icdar_poly(os.path.join(txt_dir, file), ignore_kie_type=True)
modify = False
has_TOTALCOST_keys = False
has_TOTALCOST_val = False
for icdar_pol in list_icdar_poly:
# fix wrong SELLER
if icdar_pol.type != 15 and validate_SELLER(list_seller, icdar_pol.value):
icdar_pol.type = 15
modify = True
# fix wrong ADDRESS
if icdar_pol.type != 16 and validate_ADDRESS(list_address, icdar_pol.value):
icdar_pol.type = 16
modify = True
# fix wrong num ber in ADDRESS or SELLER
if icdar_pol.type in [15, 16] and validate_TOTAL_COST_amount(icdar_pol.value):
icdar_pol.type = 1
modify = True
# print(idx, file, icdar_pol.value)
# Fix TIMESTAMP
if icdar_pol.type != 17 and validate_TIMESTAMP(icdar_pol.value):
icdar_pol.type = 17
modify = True
# Fix TOTALCOST
# if icdar_pol.type == 18:
# kk=1
if icdar_pol.type == 18 and validate_TOTAL_COST_keys(icdar_pol.value, cer_thres=0.2):
has_TOTALCOST_keys = True
for icdar_pol2 in list_icdar_poly:
if icdar_pol2.type == 18:
if validate_TOTAL_COST_amount(icdar_pol2.value, thres=0.7) or icdar_pol2.value == '':
has_TOTALCOST_val = True
if not has_TOTALCOST_val:
modify = True
find_TOTALCOST_val_poly(keys_poly=icdar_pol,
list_poly=list_icdar_poly)
# img=cv2.imread(os.path.join(img_dir, file.replace('.txt','.jpg')))
# cv2.imshow('img', img)
# cv2.waitKey(0)
if modify:
modify_icdar = ''
for icdar_pol in list_icdar_poly:
line = icdar_pol.to_icdar_line(type_map)
modify_icdar += line + '\n'
modify_icdar = modify_icdar.rstrip('\n')
print(idx, file)
with open(os.path.join(txt_dir, file), mode='w', encoding='utf-8') as f:
f.write(modify_icdar)
