def get_next_batch_for_srgan(batch_size=128):
inputs_images = []
targets_images = []
max_width_image = 0
for i in range(batch_size):
font_name = random.choice(AllFontNames)
font_length = random.randint(3, 70)
font_size = 36 #random.randint(image_height, 64)
font_mode = random.choice([0, 1, 2, 4])
font_hint = random.choice([0, 1, 2, 3, 4, 5]) #删除了2
text = utils_font.get_random_text(CHARS, eng_world_list, font_length)
image = utils_font.get_font_image_from_url(text, font_name, font_size,
font_mode, font_hint)
image = utils_pil.resize_by_height(image, image_height)
image = utils_pil.convert_to_gray(image)
targets_image = image.copy()
targets_image = np.asarray(targets_image)
targets_image = (255. - targets_image) / 255.
# targets_image = np.reshape(targets_image,[-1])
# targets_image = np.pad(targets_image,(0, image_size*image_size-np.size(targets_image)),"constant")
# targets_image = np.reshape(targets_image, [image_size,image_size])
targets_images.append(targets_image)
_h = random.randint(9, image_height // random.choice([1, 1.5, 2, 2.5]))
image = utils_pil.resize_by_height(image, _h)
image = utils_pil.resize_by_height(image, image_height,
random.random() > 0.5)
# image,_,_,_,_ = utils_pil.random_space(image)
image = utils_font.add_noise(image)
image = np.asarray(image)
# image = utils.resize(image, height=image_height)
image = image * random.uniform(0.3, 1)
if random.random() > 0.5:
image = (255. - image) / 255.
else:
image = image / 255.
# image = np.reshape(image,[-1])
# image = np.pad(image,(0, image_size*image_size-np.size(image)),"constant")
# image = np.reshape(image, [image_size,image_size])
inputs_images.append(image)
inputs = np.zeros([batch_size, image_size, image_size])
for i in range(batch_size):
inputs[i, :] = utils.square_img(inputs_images[i],
np.zeros([image_size, image_size]))
targets = np.zeros([batch_size, image_size, image_size])
for i in range(batch_size):
targets[i, :] = utils.square_img(targets_images[i],
np.zeros([image_size, image_size]))
return inputs, targets
def get_next_batch_for_srgan(batch_size=128):
inputs_images = []
targets_images = []
max_width_image = 0
for i in range(batch_size):
font_name = random.choice(AllFontNames)
font_length = random.randint(4, 5)
font_size = 36 #random.randint(image_height, 64)
font_mode = random.choice([0,1,2,4])
font_hint = random.choice([0,1,2,3,4,5]) #删除了2
text = utils_font.get_random_text(CHARS, eng_world_list, font_length)
image = utils_font.get_font_image_from_url(text, font_name, font_size, font_mode, font_hint)
image = utils_pil.resize_by_height(image, image_height)
image = utils_pil.convert_to_gray(image)
targets_image = image.copy()
_h = random.randint(9, image_height // random.choice([1,1.5,2,2.5]))
image = utils_pil.resize_by_height(image, _h)
image = utils_pil.resize_by_height(image, image_height, random.random()>0.5)
targets_image = np.asarray(targets_image)
# targets_image = utils.resize(targets_image, height=image_height)
# targets_image = utils_pil.convert_to_bw(targets_image)
targets_images.append((255. - targets_image) / 255.)
image = utils_font.add_noise(image)
image = np.asarray(image)
# image = utils.resize(image, height=image_height)
image = image * random.uniform(0.3, 1)
if random.random()>0.5:
image = (255. - image) / 255.
else:
image = image / 255.
inputs_images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
if targets_image.shape[1] > max_width_image:
max_width_image = targets_image.shape[1]
# max_width_image = max_width_image + (POOL_SIZE - max_width_image % POOL_SIZE)
inputs = np.zeros([batch_size, max_width_image, image_height])
for i in range(batch_size):
image_vec = utils.img2vec(inputs_images[i], height=image_height, width=max_width_image, flatten=False)
inputs[i,:] = np.transpose(image_vec)
targets = np.zeros([batch_size, max_width_image, image_height])
for i in range(batch_size):
image_vec = utils.img2vec(targets_images[i], height=image_height, width=max_width_image, flatten=False)
targets[i,:] = np.transpose(image_vec)
return inputs, targets
def get_next_batch_for_res(batch_size=128):
images = []
codes = []
max_width_image = 0
info = ""
for i in range(batch_size):
font_name = random.choice(AllFontNames)
font_length = random.randint(25, 30)
if random.random()>0.5:
font_size = random.randint(8, 49)
else:
font_size = random.randint(8, 15)
font_mode = random.choice([0,1,2,4])
font_hint = random.choice([0,1,2,3,4,5]) #删除了2
text = random.sample(CHARS, 12)
text = text+text+[" "," "]
random.shuffle(text)
text = "".join(text).strip()
codes.append([CHARS.index(char) for char in text])
image = utils_font.get_font_image_from_url(text, font_name, font_size, font_mode, font_hint )
image = utils_pil.resize_by_height(image, image_height, random.random()>0.5)
image = utils_font.add_noise(image)
image = utils_pil.convert_to_gray(image)
image = np.asarray(image)
image = utils.resize(image, height=image_height)
if random.random()>0.5:
image = (255. - image) / 255.
else:
image = image / 255.
images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
info = info+"%s\n\r" % utils_font.get_font_url(text, font_name, font_size, font_mode, font_hint)
max_width_image = max_width_image + (POOL_SIZE - max_width_image % POOL_SIZE)
inputs = np.zeros([batch_size, max_width_image, image_height])
for i in range(len(images)):
image_vec = utils.img2vec(images[i], height=image_height, width=max_width_image, flatten=False)
inputs[i,:] = np.transpose(image_vec)
labels = [np.asarray(i) for i in codes]
sparse_labels = utils.sparse_tuple_from(labels)
seq_len = np.ones(batch_size) * (max_width_image * image_height ) // (POOL_SIZE * POOL_SIZE)
return inputs, sparse_labels, seq_len, info
def get_next_batch_for_res(batch_size=128):
inputs_images = []
codes = []
max_width_image = 0
info = []
seq_len = np.ones(batch_size)
for i in range(batch_size):
serialized_example = next(dataset, None)
if serialized_example==None:
raise Exception("has finished train one data file, stop")
dataset_example.ParseFromString(serialized_example)
font_name = str(dataset_example.features.feature['font_name'].bytes_list.value[0], encoding="utf-8")
font_size = dataset_example.features.feature['font_size'].int64_list.value[0]
font_mode = dataset_example.features.feature['font_mode'].int64_list.value[0]
font_hint = dataset_example.features.feature['font_mode'].int64_list.value[0]
text = str(dataset_example.features.feature['label'].bytes_list.value[0], encoding="utf-8")
size = dataset_example.features.feature['size'].int64_list.value
image = dataset_example.features.feature['image'].bytes_list.value[0]
image = utils_pil.frombytes(tuple(size), image)
image = utils_pil.convert_to_gray(image)
w, h = size
if h > image_height:
image = utils_pil.resize_by_height(image, image_height)
image = utils_pil.resize_by_height(image, image_height-random.randint(1,5))
image, _ = utils_pil.random_space2(image, image, image_height)
image = utils_font.add_noise(image)
image = np.asarray(image)
image = utils.resize(image, image_height, MAX_IMAGE_WIDTH)
if random.random()>0.5:
image = image / 255.
else:
image = (255. - image) / 255.
if max_width_image < image.shape[1]:
max_width_image = image.shape[1]
inputs_images.append(image)
codes.append([CHARS.index(char) for char in text])
info.append([font_name, str(font_size), str(font_mode), str(font_hint), str(len(text))])
seq_len[i]=len(text)+1
# 凑成4的整数倍
# if max_width_image % 4 > 0:
# max_width_image = max_width_image + 4 - max_width_image % 4
# 如果图片超过最大宽度
if max_width_image < MAX_IMAGE_WIDTH:
max_width_image = MAX_IMAGE_WIDTH
# raise Exception("img width must %s <= %s " % (max_width_image, MAX_IMAGE_WIDTH))
inputs = np.zeros([batch_size, image_height, max_width_image, 1])
for i in range(batch_size):
image_vec = utils.img2vec(inputs_images[i], height=image_height, width=max_width_image, flatten=False)
inputs[i,:] = np.reshape(image_vec,(image_height, max_width_image, 1))
# print(inputs.shape, len(codes))
labels = [np.asarray(i) for i in codes]
sparse_labels = utils.sparse_tuple_from(labels)
# max_width_image = math.ceil((max_width_image-3+1.)/2.)
# max_width_image = math.ceil((max_width_image-3+1.)/1.)
# max_width_image = math.ceil((max_width_image-3+1.)/2.)
# max_width_image = math.ceil((max_width_image-3+1.)/1.)
# max_width_image = math.ceil((max_width_image-3+1.)/2.)
seq_len = np.ones(batch_size) * SEQ_LENGTH
# print(inputs.shape, seq_len.shape, [len(l) for l in labels])
return inputs, sparse_labels, seq_len, info
def get_next_batch_for_res(batch_size=128, has_sparse=True, has_onehot=True, \
max_width=4096, height=32, need_pad_width_to_max_width=False):
inputs_images = []
codes = []
# 当前这一批图片中的最大宽度
max_width_image = 0
info = []
seq_len = np.ones(batch_size)
for i in range(batch_size):
serialized_example = next(dataset, None)
if serialized_example == None:
raise Exception("has finished train one data file, stop")
dataset_example.ParseFromString(serialized_example)
font_name = str(
dataset_example.features.feature['font_name'].bytes_list.value[0],
encoding="utf-8")
font_size = dataset_example.features.feature[
'font_size'].int64_list.value[0]
font_mode = dataset_example.features.feature[
'font_mode'].int64_list.value[0]
font_hint = dataset_example.features.feature[
'font_hint'].int64_list.value[0]
text = str(
dataset_example.features.feature['label'].bytes_list.value[0],
encoding="utf-8")
size = dataset_example.features.feature['size'].int64_list.value
image = dataset_example.features.feature['image'].bytes_list.value[0]
image = utils_pil.frombytes(tuple(size), image)
# 图旋转灰度
image = utils_pil.convert_to_gray(image)
w, h = size
if h > height:
image = utils_pil.resize_by_height(image, height)
# 随机移动图片位置
image = utils_pil.resize_by_height(image,
height - random.randint(1, 5))
image, _ = utils_pil.random_space2(image, image, height)
# 增加噪点
image = utils_font.add_noise(image)
# 转为 opencv 格式
image = np.asarray(image)
# 默认按高度缩放，如果宽度超过了最大宽度，就按宽度缩放
image = utils.resize(image, height, max_width)
# 随机反色并归一化
if random.random() > 0.5:
image = image / 255.
else:
image = (255. - image) / 255.
# 记下当前的最大图片宽度
if max_width_image < image.shape[1]:
max_width_image = image.shape[1]
inputs_images.append(image)
codes.append([CHARS.index(char) for char in text])
info.append([
font_name,
str(font_size),
str(font_mode),
str(font_hint),
str(len(text))
])
# 凑成4的整数倍
if max_width_image % 4 > 0:
max_width_image = max_width_image + 4 - max_width_image % 4
# 如果图片超过最大宽度，懒得去缩放，直接报异常
if max_width_image > max_width:
raise Exception("img width must %s <= %s " %
(max_width_image, max_width))
if need_pad_width_to_max_width:
max_width_image = max_width
inputs = np.zeros([batch_size, image_height, max_width_image, 1])
for i in range(batch_size):
image_vec = utils.img2vec(inputs_images[i],
height=image_height,
width=max_width_image,
flatten=False)
inputs[i, :] = np.reshape(image_vec,
(image_height, max_width_image, 1))
labels = [np.asarray(i) for i in codes]
sparse_labels = None
onehot_labels = None
if has_sparse:
sparse_labels = utils.sparse_tuple_from(labels)
sparse_labels = np.array(sparse_labels)
if has_onehot:
onehot_labels = []
for label in labels:
label_one_hot = np.eye(CLASSES_NUMBER)[label]
onehot_labels.append(label_one_hot)
onehot_labels = np.array(onehot_labels)
return inputs, np.array(labels), sparse_labels, onehot_labels, info
def get_next_batch(batch_size=128):
images = []
to_images = []
codes = []
max_width_image = 0
for i in range(batch_size):
font_name = random.choice(AllFontNames)
font_length = random.randint(25, 30)
font_size = 36 #random.randint(image_height, 64)
font_mode = random.choice([0, 1, 2, 4])
font_hint = random.choice([0, 1, 2, 3, 4, 5])
text = utils_font.get_random_text(CHARS, eng_world_list, font_length)
# text = random.sample(CHARS, 12)
# text = text+text
# random.shuffle(text)
# text = "".join(text).strip()
codes.append([CHARS.index(char) for char in text])
image = utils_font.get_font_image_from_url(text,
font_name,
font_size,
fontmode=font_mode,
fonthint=font_hint)
image = utils_pil.resize_by_height(image, image_height)
to_image = image.copy()
image = utils_font.add_noise(image)
image = utils_pil.convert_to_gray(image)
_h = random.randint(9, image_height // random.choice([1, 1.5, 2, 2.5]))
image = utils_pil.resize_by_height(image, _h, random.random() > 0.5)
image = utils_pil.resize_by_height(image, image_height,
random.random() > 0.5)
image = np.asarray(image)
image = utils.resize(image, height=image_height)
image = (255. - image) / 255.
images.append(image)
# to_image = utils_font.get_font_image_from_url(text, font_name ,image_height, fontmode = font_mode, fonthint = font_hint)
to_image = utils_pil.convert_to_gray(to_image)
to_image = np.asarray(to_image)
to_image = utils.resize(to_image, height=image_height)
to_image = utils.img2bwinv(to_image)
to_image = to_image / 255.
to_images.append(to_image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
if to_image.shape[1] > max_width_image:
max_width_image = to_image.shape[1]
max_width_image = max_width_image + (POOL_SIZE -
max_width_image % POOL_SIZE)
inputs = np.zeros([batch_size, max_width_image, image_height])
for i in range(len(images)):
image_vec = utils.img2vec(images[i],
height=image_height,
width=max_width_image,
flatten=False)
inputs[i, :] = np.transpose(image_vec)
targets = np.zeros([batch_size, max_width_image, image_height])
for i in range(len(to_images)):
image_vec = utils.img2vec(to_images[i],
height=image_height,
width=max_width_image,
flatten=False)
targets[i, :] = np.transpose(image_vec)
labels = [np.asarray(i) for i in codes]
sparse_labels = utils.sparse_tuple_from(labels)
seq_len = np.ones(batch_size) * (max_width_image *
image_height) // (POOL_SIZE * POOL_SIZE)
return inputs, targets, sparse_labels, seq_len
def get_next_batch_for_res(batch_size=128,
add_noise=True,
_font_name=None,
_font_size=None,
_font_mode=None,
_font_hint=None):
inputs_images = []
codes = []
max_width_image = 0
info = []
for i in range(batch_size):
font_name = _font_name
font_size = _font_size
font_mode = _font_mode
font_hint = _font_hint
if font_name == None:
font_name = random.choice(AllFontNames)
if font_size == None:
if random.random() > 0.5:
font_size = random.randint(9, 49)
else:
font_size = random.randint(9, 15)
if font_mode == None:
font_mode = random.choice([0, 1, 2, 4])
if font_hint == None:
font_hint = random.choice([0, 1, 2, 3, 4, 5])
while True:
font_length = random.randint(5, 400)
# text = random.sample(CHARS, font_length)
# text = text+text+[" "," "]
# random.shuffle(text)
# text = "".join(text).strip()
text = utils_font.get_random_text(CHARS, eng_world_list,
font_length)
image = utils_font.get_font_image_from_url(text, font_name,
font_size, font_mode,
font_hint)
temp_image = utils_pil.resize_by_height(image, image_height)
w, h = temp_image.size
if w * h < image_size * image_size: break
image = utils_pil.convert_to_gray(image)
w, h = image.size
if h > image_height:
image = utils_pil.resize_by_height(image, image_height)
if add_noise and random.random() > 0.5:
_h = random.randint(9, image_height + 1)
image = utils_pil.resize_by_height(image, _h)
image = utils_pil.random_space2(image, image_height)
if add_noise:
image = utils_font.add_noise(image)
image = np.asarray(image)
# image = utils.resize(image, height=image_height)
if add_noise:
image = image * random.uniform(0.3, 1)
if add_noise and random.random() > 0.5:
image = image / 255.
else:
image = (255. - image) / 255.
inputs_images.append(image)
codes.append([CHARS.index(char) for char in text])
info.append(
[font_name,
str(font_size),
str(font_mode),
str(font_hint)])
inputs = np.zeros([batch_size, image_size, image_size])
for i in range(batch_size):
inputs[i, :] = utils.square_img(inputs_images[i],
np.zeros([image_size, image_size]))
labels = [np.asarray(i) for i in codes]
sparse_labels = utils.sparse_tuple_from(labels)
seq_len = np.ones(batch_size) * (image_size * image_size) // (POOL_SIZE *
POOL_SIZE)
return inputs, sparse_labels, seq_len, info
def get_next_batch_for_gan(batch_size=128):
input_images = []
trim_images = []
for i in range(batch_size):
font_name = random.choice(AllFontNames)
# font_size = image_height #random.randint(image_height, 64)
# if font_size==None:
if random.random() > 0.5:
font_size = random.randint(9, 49)
else:
font_size = random.randint(9, 15)
font_mode = random.choice([0, 1, 2, 4])
font_hint = random.choice([0, 1, 2, 3, 4, 5]) #删除了2
while True:
font_length = random.randint(3, 400)
# text = utils_font.get_random_text(CHARS, eng_world_list, font_length)
text = utils_font.get_words_text(CHARS, eng_world_list,
font_length)
image = utils_font.get_font_image_from_url(text,
font_name,
font_size,
font_mode,
font_hint,
trim=False)
if font_hint in (0, 1, 3, 5):
clear_trim_image = utils_font.get_font_image_from_url(
text, font_name, font_size, font_mode, 0, trim=False)
else:
clear_trim_image = utils_font.get_font_image_from_url(
text, font_name, font_size, font_mode, 4, trim=False)
bbox = utils_pil.get_trim_box(clear_trim_image)
image = image.crop(bbox)
clear_trim_image = clear_trim_image.crop(bbox)
temp_image = utils_pil.resize_by_height(image, image_height)
if clear_trim_image.size[0] != image.size[
0] or clear_trim_image.size[1] != image.size[1]:
print("get size not same,", image.size, clear_trim_image.size,
font_name, font_size, font_mode, font_hint)
continue
w, h = temp_image.size
if w * h <= image_size * image_size:
break
image = utils_pil.convert_to_gray(image) #原始图片
clear_trim_image = utils_pil.convert_to_gray(clear_trim_image)
w, h = image.size
if h > image_height:
image = utils_pil.resize_by_height(image, image_height)
clear_trim_image = utils_pil.resize_by_size(
clear_trim_image, image.size)
source_image = image.copy()
# 随机缩放下图片
w, h = image.size
if random.random() > 0.5 and h < image_height:
_h = random.randint(h + 1, image_height + 1)
image = utils_pil.resize_by_height(image, _h,
random.random() > 0.5)
clear_trim_image = utils_pil.resize_by_size(
clear_trim_image, image.size)
if clear_trim_image.size[0] != image.size[
0] or clear_trim_image.size[1] != image.size[1]:
print("random resize get size not same,", image.size,
clear_trim_image.size, font_name, font_size, font_mode,
font_hint)
# image = utils_pil.resize_by_height(image, image_height, random.random()>0.5)
# 随机移动位置 trims_image 为字体实际位置标识
image, clear_trim_image = utils_pil.random_space2(
image, clear_trim_image, image_height)
if clear_trim_image.size[0] != image.size[0] or clear_trim_image.size[
1] != image.size[1]:
print("random move space get size not same,", image.size,
clear_trim_image.size, font_name, font_size, font_mode,
font_hint)
trims_image = np.asarray(clear_trim_image)
# 转黑白二值化，降低维度
trims_image = (255. - trims_image) / 255.
trim_images.append(trims_image)
if random.random() > 0.9:
image = utils_font.add_noise(image)
image = np.asarray(image)
image.flags.writeable = True
if random.random() > 0.9:
image[image > 200] = 255 * random.uniform(0.5, 1)
if random.random() > 0.5:
image = (255. - image) / 255.
else:
image = image / 255.
input_images.append(image)
inputs = np.zeros([batch_size, image_size, image_size])
for i in range(batch_size):
inputs[i, :] = utils.square_img(input_images[i],
np.zeros([image_size, image_size]),
image_height)
trims = np.zeros([batch_size, image_size, image_size])
for i in range(batch_size):
trims[i, :] = utils.square_img(trim_images[i],
np.zeros([image_size, image_size]),
image_height)
trims[trims == 0] == -1
return inputs, trims