def generator(
input_images_dir,
input_gt_dir,
input_size=512,
batch_size=12,
random_scale=np.array([0.8, 0.85, 0.9, 0.95, 1.0, 1.1, 1.2]),
):
#data_loader = SynthTextLoader()
data_loader = ICDARLoader(edition='13', shuffle=True)
#image_list = np.array(data_loader.get_images(FLAGS.training_data_dir))
image_list = np.array(data_loader.get_images(input_images_dir))
print('{} training images in {} '.format(image_list.shape[0],
input_images_dir))
index = np.arange(0, image_list.shape[0])
while True:
np.random.shuffle(index)
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
for i in index:
try:
im_fn = image_list[i]
#print(im_fn)
# if im_fn.split(".")[0][-1] == '0' or im_fn.split(".")[0][-1] == '2':
# continue
im = cv2.imread(im_fn)
h, w, _ = im.shape
#file_name = im_fn.replace(os.path.basename(im_fn).split('.')[1], 'txt').split('/')[-1]
file_name = os.path.basename(im_fn) + '.txt'
# file_name = im_fn.replace(im_fn.split('.')[1], 'txt') # using for synthtext
# txt_fn = os.path.join(FLAGS.training_gt_data_dir, file_name)
txt_fn = os.path.join(input_gt_dir, file_name)
if not os.path.exists(txt_fn):
print('text file {} does not exists'.format(txt_fn))
continue
#print(txt_fn)
text_polys, text_tags, text_labels = data_loader.load_annotation(
txt_fn) # Change for load text transiption
if text_polys.shape[0] == 0:
continue
text_polys, text_tags, text_labels = check_and_validate_polys(
text_polys, text_tags, text_labels, (h, w))
############################# Data Augmentation ##############################
# input image：1920*1080 > 960*540
im = cv2.resize(im, dsize=(960, 540))
text_polys *= 0.5
# color aug
def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5):
x = (np.random.uniform(-1, 1, 3) *
np.array([hgain, sgain, vgain]) + 1).astype(
np.float32) # random gains
img_hsv = (cv2.cvtColor(img, cv2.COLOR_BGR2HSV) *
x.reshape((1, 1, 3))).clip(None,
255).astype(np.uint8)
cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR,
dst=img) # no return needed
#augment_hsv(im)
# random scale this image
rd_scale = np.random.choice(random_scale)
im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale)
#im = cv2.resize(im, dsize=(960, 540), fx=rd_scale, fy=rd_scale) #头两个resize版本（这样的话做不了数据增强
text_polys *= rd_scale
# rotate image from [-10, 10],因为nba都是水平框，数据增强旋转没必要
#angle = random.randint(-10, 10)
#im, text_polys = rotate_image(im, text_polys, angle)
# 600×600 random samples are cropped.
im, text_polys, text_tags, selected_poly = crop_area(
im, text_polys, text_tags, crop_background=False)
# im, text_polys, text_tags, selected_poly = crop_area_fix(im, text_polys, text_tags, crop_size=(600, 600))
text_labels = [text_labels[i] for i in selected_poly]
if text_polys.shape[0] == 0 or len(text_labels) == 0:
continue
# pad the image to the training input size or the longer side of image
new_h, new_w, _ = im.shape
max_h_w_i = np.max([new_h, new_w, input_size])
im_padded = np.zeros((max_h_w_i, max_h_w_i, 3), dtype=np.uint8)
im_padded[:new_h, :new_w, :] = im.copy()
im = im_padded
# resize the image to input size
new_h, new_w, _ = im.shape
resize_h = input_size
resize_w = input_size
im = cv2.resize(im, dsize=(resize_w, resize_h))
resize_ratio_3_x = resize_w / float(new_w)
resize_ratio_3_y = resize_h / float(new_h)
text_polys[:, :, 0] *= resize_ratio_3_x
text_polys[:, :, 1] *= resize_ratio_3_y
new_h, new_w, _ = im.shape
score_map, geo_map, training_mask, rectangles = generate_rbox(
(new_h, new_w), text_polys, text_tags)
mask = [not (word == [-1]) for word in text_labels]
text_labels = list(compress(text_labels, mask))
rectangles = list(compress(rectangles, mask))
assert len(text_labels) == len(
rectangles
), "rotate rectangles' num is not equal to text label"
if len(text_labels) == 0:
continue
boxes_mask = np.array([count] * len(rectangles))
count += 1
batch_images.append(im[:, :, ::-1].astype(np.float32))
batch_image_fns.append(im_fn)
batch_score_maps.append(score_map[::4, ::4, np.newaxis].astype(
np.float32))
batch_geo_maps.append(geo_map[::4, ::4, :].astype(np.float32))
batch_training_masks.append(
training_mask[::4, ::4, np.newaxis].astype(np.float32))
batch_text_polyses.append(rectangles)
batch_boxes_masks.append(boxes_mask)
batch_text_labels.extend(text_labels)
batch_text_tagses.append(text_tags)
if len(batch_images) == batch_size:
batch_text_polyses = np.concatenate(batch_text_polyses)
batch_text_tagses = np.concatenate(batch_text_tagses)
batch_transform_matrixes, batch_box_widths = get_project_matrix_and_width(
batch_text_polyses, batch_text_tagses)
# TODO limit the batch size of recognition
batch_text_labels_sparse = sparse_tuple_from(
np.array(batch_text_labels))
# yield images, image_fns, score_maps, geo_maps, training_masks
yield batch_images, batch_image_fns, batch_score_maps, batch_geo_maps, batch_training_masks, batch_transform_matrixes, batch_boxes_masks, batch_box_widths, batch_text_labels_sparse, batch_text_polyses, batch_text_labels
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
except Exception as e:
import traceback
#print(im_fn)
traceback.print_exc()
continue
def generator_all(input_images_dir, input_gt_dir, output_geo_dir, output_score_dir, input_size=512, class_num=230):
geo_dir = "./"
# data_loader = SynthTextLoader()
data_loader = ICDARLoader(edition='13', shuffle=True)
# image_list = np.array(data_loader.get_images(FLAGS.training_data_dir))
image_list = np.array(data_loader.get_images(input_images_dir))
# print('{} training images in {} '.format(image_list.shape[0], FLAGS.training_data_dir))
index = np.arange(0, image_list.shape[0])
# while True:
# np.random.shuffle(index)
for i in index:
try:
im_fn = image_list[i]
# print(im_fn)
# if im_fn.split(".")[0][-1] == '0' or im_fn.split(".")[0][-1] == '2':
# continue
im = cv2.imread(os.path.join(input_images_dir, im_fn))
h, w, _ = im.shape
file_name = "gt_" + im_fn.replace(os.path.basename(im_fn).split('.')[1], 'txt').split('\\')[-1]
# file_name = im_fn.replace(im_fn.split('.')[1], 'txt') # using for synthtext
# txt_fn = os.path.join(FLAGS.training_gt_data_dir, file_name)
txt_fn = os.path.join(input_gt_dir, file_name)
if not os.path.exists(txt_fn):
print('text file {} does not exists'.format(txt_fn))
continue
# print(txt_fn)
text_polys, text_tags, text_labels = data_loader.load_annotation(
txt_fn) # Change for load text transiption
if text_polys.shape[0] == 0:
continue
text_polys, text_tags, text_labels = check_and_validate_polys(text_polys, text_tags, text_labels,
(h, w))
# prepare_one_img_time = time.time()
# pad the image to the training input size or the longer side of image
new_h, new_w, _ = im.shape
max_h_w_i = np.max([new_h, new_w, input_size])
im_padded = np.zeros((max_h_w_i, max_h_w_i, 3), dtype=np.uint8)
im_padded[:new_h, :new_w, :] = im.copy()
im = im_padded
# resize the image to input size
new_h, new_w, _ = im.shape
resize_h = input_size
resize_w = input_size
im = cv2.resize(im, dsize=(resize_w, resize_h))
resize_ratio_3_x = resize_w / float(new_w)
resize_ratio_3_y = resize_h / float(new_h)
text_polys[:, :, 0] *= resize_ratio_3_x
text_polys[:, :, 1] *= resize_ratio_3_y
new_h, new_w, _ = im.shape
score_map, geo_map, training_mask, rectangles = generate_maps((new_h, new_w), text_polys, text_tags)
# rectangles_list = []
# text_polys = text_polys.astype(np.float32)
# for i in range(2):
# rectangles_list.append(text_polys[i].flatten())
im_name = (im_fn.split("\\")[-1]).split(".")[0]
print(im_name)
with open("{}/{}.pickle".format(output_geo_dir, im_name), "wb") as f_out:
pickle.dump(geo_map, f_out)
f_out.close()
with open("{}/{}.pickle".format(output_score_dir, im_name), "wb") as f_out:
pickle.dump(score_map, f_out)
f_out.close()
# # turn the label to one hot representation
# text_labels = [str(i) for i in text_labels[0]]
# label = ''.join(text_labels)
# one_hot_label = dense_to_one_hot(int(label), n_classes=class_num)
# print(label)
# start_time = time.time()
# with open("{}/{}.pickle".format(output_geo_dir, im_name), "rb") as f_in:
# in_geo_map = pickle.load(f_in)
#
# print("\nprepare_one_img_cost:" + str(time.time() - start_time))
# print(in_geo_map.shape)
except Exception as e:
import traceback
print(im_fn)
traceback.print_exc()
continue
def generator(input_images_dir,
input_gt_dir,
input_size=512,
batch_size=12,
random_scale=np.array([0.8, 0.85, 0.9, 0.95, 1.0, 1.1, 1.2])):
data_loader = ICDARLoader()
image_list = np.array(data_loader.get_images(input_images_dir))
print('{} training images in {} '.format(image_list.shape[0],
input_images_dir))
index = np.arange(0, image_list.shape[0])
while True:
np.random.shuffle(index)
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
for i in index:
try:
im_fn = image_list[i]
# print(im_fn)
# if im_fn.split(".")[0][-1] == '0' or im_fn.split(".")[0][-1] == '2':
# continue
im = cv2.imread(im_fn)
h, w, _ = im.shape
file_name = "gt_" + im_fn.replace(
os.path.basename(im_fn).split('.')[1],
'txt').split('/')[-1]
txt_fn = os.path.join(input_gt_dir, file_name)
if not os.path.exists(txt_fn):
print('text file {} does not exists'.format(txt_fn))
continue
#print(txt_fn)
# Change for load text transiption
text_polys, text_tags, text_labels = data_loader.load_annotation(
txt_fn)
if text_polys.shape[0] == 0:
continue
text_polys, text_tags, text_labels = check_and_validate_polys(
text_polys, text_tags, text_labels, (h, w))
############################# Data Augmentation ##############################
# random scale this image
rd_scale = np.random.choice(random_scale)
im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale)
text_polys *= rd_scale
# rotate image from [-10, 10]
angle = random.randint(-10, 10)
im, text_polys = rotate_image(im, text_polys, angle)
# 600×600 random samples are cropped.
im, text_polys, text_tags, selected_poly = crop_area(
im, text_polys, text_tags, crop_background=False)
# im, text_polys, text_tags, selected_poly = crop_area_fix(im,
# text_polys,
# text_tags,
# crop_size=(600, 600))
text_labels = [text_labels[i] for i in selected_poly]
if text_polys.shape[0] == 0 or len(text_labels) == 0:
continue
# pad the image to the training input size or the longer side of image
new_h, new_w, _ = im.shape
max_h_w_i = np.max([new_h, new_w, input_size])
im_padded = np.zeros((max_h_w_i, max_h_w_i, 3), dtype=np.uint8)
im_padded[:new_h, :new_w, :] = im.copy()
im = im_padded
# resize the image to input size
new_h, new_w, _ = im.shape
resize_h = input_size
resize_w = input_size
im = cv2.resize(im, dsize=(resize_w, resize_h))
resize_ratio_3_x = resize_w / float(new_w)
resize_ratio_3_y = resize_h / float(new_h)
text_polys[:, :, 0] *= resize_ratio_3_x
text_polys[:, :, 1] *= resize_ratio_3_y
new_h, new_w, _ = im.shape
score_map, geo_map, training_mask, rectangles = generate_rbox(
(new_h, new_w), text_polys, text_tags)
mask = [not (word == [-1]) for word in text_labels]
text_labels = list(compress(text_labels, mask))
rectangles = list(compress(rectangles, mask))
assert len(text_labels) == len(
rectangles
), "rotate rectangles' num is not equal to text label"
if len(text_labels) == 0:
continue
boxes_mask = np.array([count] * len(rectangles))
count += 1
batch_images.append(im[:, :, ::-1].astype(np.float32))
batch_image_fns.append(im_fn)
batch_score_maps.append(score_map[::4, ::4, np.newaxis].astype(
np.float32))
batch_geo_maps.append(geo_map[::4, ::4, :].astype(np.float32))
batch_training_masks.append(
training_mask[::4, ::4, np.newaxis].astype(np.float32))
batch_text_polyses.append(rectangles)
batch_boxes_masks.append(boxes_mask)
batch_text_labels.extend(text_labels)
batch_text_tagses.append(text_tags)
if len(batch_images) == batch_size:
batch_text_polyses = np.concatenate(batch_text_polyses)
batch_text_tagses = np.concatenate(batch_text_tagses)
batch_transform_matrixes, batch_box_widths = get_project_matrix_and_width(
batch_text_polyses, batch_text_tagses)
# TODO limit the batch size of recognition
batch_text_labels_sparse = sparse_tuple_from(
np.array(batch_text_labels))
# yield images, image_fns, score_maps, geo_maps, training_masks
yield (batch_images, batch_image_fns, batch_score_maps,
batch_geo_maps, batch_training_masks,
batch_transform_matrixes, batch_boxes_masks,
batch_box_widths, batch_text_labels_sparse,
batch_text_polyses, batch_text_labels)
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
except Exception as e:
import traceback
print(im_fn)
traceback.print_exc()
continue
def generator(input_images_dir, input_gt_dir, input_size=512, batch_size=12, class_num=230,
random_scale=np.array([0.8, 0.85, 0.9, 0.95, 1.0, 1.1, 1.2]), back_side=0):
# data_loader = SynthTextLoader()
data_loader = ICDARLoader(edition='13', shuffle=True)
# image_list = np.array(data_loader.get_images(FLAGS.training_data_dir))
image_list = np.array(data_loader.get_images(input_images_dir))
# print('{} training images in {} '.format(image_list.shape[0], FLAGS.training_data_dir))
index = np.arange(0, image_list.shape[0])
while True:
np.random.shuffle(index)
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
for i in index:
try:
start_time = time.time()
im_fn = image_list[i]
# print(im_fn)
# if im_fn.split(".")[0][-1] == '0' or im_fn.split(".")[0][-1] == '2':
# continue
im = cv2.imread(os.path.join(input_images_dir, im_fn))
h, w, _ = im.shape
file_name = "gt_" + im_fn.replace(os.path.basename(im_fn).split('.')[1], 'txt').split('/')[-1]
# file_name = im_fn.replace(im_fn.split('.')[1], 'txt') # using for synthtext
# txt_fn = os.path.join(FLAGS.training_gt_data_dir, file_name)
txt_fn = os.path.join(input_gt_dir, file_name)
if not os.path.exists(txt_fn):
print('text file {} does not exists'.format(txt_fn))
continue
# print(txt_fn)
text_polys, text_tags, text_labels = data_loader.load_annotation(
txt_fn) # Change for load text transiption
if text_polys.shape[0] == 0:
continue
text_polys, text_tags, text_labels = check_and_validate_polys(text_polys, text_tags, text_labels,
(h, w), back_side)
############################# Data Augmentation ##############################
'''
# random scale this image
rd_scale = np.random.choice(random_scale)
im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale)
text_polys *= rd_scale
# rotate image from [-10, 10]
angle = random.randint(-10, 10)
im, text_polys = rotate_image(im, text_polys, angle)
# 600×600 random samples are cropped.
im, text_polys, text_tags, selected_poly = crop_area(im, text_polys, text_tags, crop_background=False)
# im, text_polys, text_tags, selected_poly = crop_area_fix(im, text_polys, text_tags, crop_size=(600, 600))
text_labels = [text_labels[i] for i in selected_poly]
if text_polys.shape[0] == 0 or len(text_labels) == 0:
continue
'''
################################################################################
# prepare_one_img_time = time.time()
# pad the image to the training input size or the longer side of image
new_h, new_w, _ = im.shape
max_h_w_i = np.max([new_h, new_w, input_size])
im_padded = np.zeros((max_h_w_i, max_h_w_i, 3), dtype=np.uint8)
im_padded[:new_h, :new_w, :] = im.copy()
im = im_padded
# resize the image to input size
new_h, new_w, _ = im.shape
resize_h = input_size
resize_w = input_size
im = cv2.resize(im, dsize=(resize_w, resize_h))
resize_ratio_3_x = resize_w / float(new_w)
resize_ratio_3_y = resize_h / float(new_h)
text_polys[:, :, 0] *= resize_ratio_3_x
text_polys[:, :, 1] *= resize_ratio_3_y
new_h, new_w, _ = im.shape
training_mask, rectangles = generate_rbox((new_h, new_w), text_polys, text_tags)
rectangles_list = []
text_polys = text_polys.astype(np.float32)
# for i in range(2):
rectangles_list.append(text_polys.flatten())
im_name = (im_fn.split("/")[-1]).split(".")[0]
geo_dir = "{}/geo_map".format(im_fn.split("JPEGImages")[0])
with open("{}/{}.pickle".format(geo_dir, im_name), "rb") as f_in:
geo_map = pickle.load(f_in)
f_in.close()
score_dir = "{}/score_map".format(im_fn.split("JPEGImages")[0])
with open("{}/{}.pickle".format(score_dir, im_name), "rb") as f_in:
score_map = pickle.load(f_in)
f_in.close()
# print("\nprepare_one_img_cost:" + str(time.time() - start_time))
mask = [not (word == [-1]) for word in text_labels]
# remove the unreadable text
text_labels = list(compress(text_labels, mask))
rectangles = list(compress(rectangles_list, mask))
assert len(text_labels) == len(rectangles), "rotate rectangles' num is not equal to text label"
if len(text_labels) == 0:
continue
# turn the label to one hot representation
text_labels = [str(i) for i in text_labels[0]]
label = ''.join(text_labels)
one_hot_label = dense_to_one_hot(int(label), n_classes=class_num)
boxes_mask = np.array([count] * len(rectangles))
count += 1
batch_images.append(im[:, :, ::-1].astype(np.float32))
batch_image_fns.append(im_fn)
batch_score_maps.append(score_map[::4, ::4, np.newaxis].astype(np.float32))
batch_geo_maps.append(geo_map[::4, ::4, :].astype(np.float32))
batch_training_masks.append(training_mask[::4, ::4, np.newaxis].astype(np.float32))
batch_text_polyses.append(rectangles)
batch_boxes_masks.append(boxes_mask)
batch_text_labels.extend(one_hot_label)
batch_text_tagses.append(text_tags)
if len(batch_images) == batch_size:
batch_text_polyses = np.concatenate(batch_text_polyses)
batch_text_tagses = np.concatenate(batch_text_tagses)
batch_transform_matrixes, batch_box_widths = get_project_matrix(batch_text_polyses,
batch_text_tagses)
# TODO limit the batch size of recognition
batch_text_labels_sparse = sparse_tuple_from(np.array(batch_text_labels))
# yield images, image_fns, score_maps, geo_maps, training_masks
yield batch_images, batch_image_fns, batch_score_maps, batch_geo_maps, batch_training_masks, batch_transform_matrixes, batch_boxes_masks, batch_box_widths, batch_text_labels_sparse, batch_text_polyses, batch_text_labels
batch_images = []
batch_image_fns = []
batch_score_maps = []
batch_geo_maps = []
batch_training_masks = []
batch_text_polyses = []
batch_text_tagses = []
batch_boxes_masks = []
batch_text_labels = []
count = 0
except Exception as e:
import traceback
print(im_fn)
traceback.print_exc()
continue