def get_next_batch(batch_size=128):
images = []
to_images = []
max_width_image = 0
font_min_length = random.randint(10, 20)
for i in range(batch_size):
font_name = random.choice(AllFontNames)
font_length = random.randint(font_min_length-5, font_min_length+5)
font_size = random.randint(9, 64)
font_mode = random.choice([0,1,2,4])
text = getRedomText(CHARS, eng_world_list, font_length)
image= getImage(text, font_name, font_length, font_size, noise = True, fontmode = font_mode)
image=utils.resize(image, height=image_height)
images.append(image)
to_image=getImage(text, font_name, font_length, image_height, noise = False, fontmode = font_mode, fonthint = 0)
to_image=utils.resize(to_image, height=image_height)
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)
labels = 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)
labels[i,:] = np.transpose(image_vec)
return inputs, labels
def get_next_batch(batch_size=128):
images = []
to_images = []
max_width_image = 0
font_min_length = random.randint(10, 20)
for i in range(batch_size):
font_name = random.choice(AllFontNames)
# font_length = random.randint(font_min_length-5, font_min_length+5)
font_length = random.randint(3, 5)
font_size = 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)
image = utils_font.get_font_image_from_url(text,
font_name,
font_size,
fontmode=font_mode,
fonthint=font_hint)
to_image = image.copy()
image = utils_font.add_noise(image)
image = utils_pil.convert_to_gray(image)
rate = random.randint(8, 17) / font_size
image = utils_pil.resize(image, rate)
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)
labels = 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)
labels[i, :] = np.transpose(image_vec)
return inputs, labels
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 scan(file):
img = Image.open(file.stream)
image = np.array(img)
image = utils.img2gray(image)
image = utils.clearImgGray(image)
utils.save(image, os.path.join(curr_dir,"test","src.png"))
split_images = utils.splitImg(image)
ocr_texts = []
for i, split_image in enumerate(split_images):
# image = utils.img2bwinv(split_image)
image = utils.clearImgGray(split_image)
# image = utils.clearBackgroundColor(image, 255)
image = 255. - image
image = utils.dropZeroEdges(image)
image = utils.resize(image, ocr.image_height)
utils.save(image,os.path.join(curr_dir,"test","%s.png"%i))
image = image / 255.
maxImageWidth = image.shape[1]
maxImageWidth = maxImageWidth + (ocr.POOL_SIZE - maxImageWidth % ocr.POOL_SIZE)
image_vec = utils.img2vec(image,ocr.image_height,maxImageWidth)
ocr_inputs = np.zeros([1, maxImageWidth, ocr.image_height])
ocr_inputs[0,:] = np.transpose(image_vec.reshape((ocr.image_height,maxImageWidth)))
ocr_seq_len = np.ones(ocr_inputs.shape[0]) * (maxImageWidth * ocr.image_height) // (ocr.POOL_SIZE * ocr.POOL_SIZE)
feed = {inputs: ocr_inputs, seq_len: ocr_seq_len, input_keep_prob: 1.0}
start = time.time()
decoded_list = session.run(decoded[0], feed)
seconds = round(time.time() - start,2)
print("filished ocr %s , paid %s seconds" % (i,seconds))
detected_list = ocr.decode_sparse_tensor(decoded_list)
for detect_number in detected_list:
ocr_texts.append(ocr.list_to_chars(detect_number))
return ocr_texts
def crack(file):
img = Image.open(file.stream)
image = np.array(img)
image = img2vec(img2gray(image))
y_ = sess.run([prediction], feed_dict={x: [image]})
result = "".join([str(s) for s in y_[0][0]])
return result
def scan(file):
img_array = np.asarray(bytearray(file.stream.read()), dtype=np.uint8)
image = cv2.imdecode(img_array,0)
split_images = utils.splitImg(image)
ocr_texts = []
for i, split_image in enumerate(split_images):
image =utils.img2bwinv(split_image)
image = utils.dropZeroEdges(image)
image = utils.resize(image, ocr.image_height)
utils.save(image,os.path.join(curr_dir,"test","%s.png"%i))
maxImageWidth = image.shape[1]+5
image_vec = utils.img2vec(image,ocr.image_height,maxImageWidth)
ocr_inputs = np.zeros([1, maxImageWidth, ocr.image_height])
ocr_inputs[0,:] = np.transpose(image_vec.reshape((ocr.image_height,maxImageWidth)))
ocr_seq_len = np.ones(ocr_inputs.shape[0]) * maxImageWidth
feed = {inputs: ocr_inputs, seq_len: ocr_seq_len, input_keep_prob: 1.0}
start = time.time()
decoded_list = session.run(decoded[0], feed)
seconds = round(time.time() - start,2)
print("filished ocr %s , paid %s seconds" % (i,seconds))
detected_list = ocr.decode_sparse_tensor(decoded_list)
for detect_number in detected_list:
ocr_texts.append(ocr.list_to_chars(detect_number))
return ocr_texts
def get_next_batch(batch_size=128):
inputs = np.zeros([batch_size, image_size[1] * image_size[0]])
labels = np.zeros([batch_size, label_size], dtype=int)
batch = random.sample(train_files, batch_size)
for i, line in enumerate(batch):
lines = line.split(" ")
imageFileName = lines[0] + ".png"
text = line[line.index(' '):].strip()
# 在宋体9号字体下，O和0完全一致，因此全部按0处理
# text = text.replace('O','0')
# 文本需要补齐空格
text = text + "".join([' ' for x in range(label_size - len(text))])
if imageFileName in images:
imgvec = images[imageFileName]
else:
# 输出图片为反色黑白
image = readImgFile(os.path.join(curr_dir, DATA_DIR,
imageFileName))
image = img2bwinv(image)
image = dropZeroEdges(image)
image = resize(image, image_size[0])
imgvec = img2vec(image, image_size[0], image_size[1])
# images[imageFileName] = imgvec
inputs[i, :] = imgvec
label_list = []
for c in text:
if c in CHARS:
label_list.append(CHARS.index(c))
else:
label_list.append(CHARS.index(UNKOWN_CHAR))
labels[i, :] = label_list
return inputs, labels
def get_next_batch(batch_size=128):
batch = random.sample(train_files, batch_size)
codes = []
images = []
max_width_image = 0
for line in batch:
lines = line.split(" ")
imageFileName = lines[0]+".png"
text = line[line.index(' '):].strip()
# 输入的图片为反色黑白
image = readImgFile(os.path.join(curr_dir, "data" ,"dataset", imageFileName))
images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
text_list = [CHARS.index(char) for char in text]
codes.append(text_list)
inputs = np.zeros([batch_size, max_width_image, image_height])
for i in range(len(images)):
image_vec = 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]
#labels转成稀疏矩阵
sparse_labels = sparse_tuple_from(labels)
seq_len = np.ones(batch_size) * max_width_image
return inputs, sparse_labels, seq_len
def get_next_batch(batch_size=128):
codes = []
images = []
max_width_image = 0
font_min_length = random.randint(10, 80)
for i in range(batch_size):
font_name = random.choice(FontNames)
font_length = random.randint(font_min_length-5, font_min_length+5)
font_size = random.randint(14, 64)
text, image= getImage(CHARS, font_name, image_height, font_length, font_size, eng_world_list)
images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
text_list = [CHARS.index(char) for char in text]
codes.append(text_list)
# 凑成4的整数倍
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 = 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]
#labels转成稀疏矩阵
sparse_labels = sparse_tuple_from(labels)
#因为模型做了2次pool，所以 seq_len 也需要除以4
seq_len = np.ones(batch_size) * (max_width_image * image_height) // (POOL_SIZE * POOL_SIZE)
return inputs, sparse_labels, seq_len
def get_next_batch(batch_size=128):
codes = []
images = []
max_width_image = 0
for i in range(batch_size):
font_name = random.choice(FontNames)
font_length = random.randint(50, 100)
font_size = random.randint(9, 20)
text, image= getImage(CHARS, font_name, image_height, font_length, font_size)
images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
text_list = [CHARS.index(char) for char in text]
codes.append(text_list)
inputs = np.zeros([batch_size, max_width_image, image_height])
for i in range(len(images)):
image_vec = 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]
#labels转成稀疏矩阵
sparse_labels = sparse_tuple_from(labels)
seq_len = np.ones(batch_size) * max_width_image
return inputs, sparse_labels, seq_len
def get_next_batch(batch_size=128):
images = []
to_images = []
max_width_image = 0
font_min_length = random.randint(10, 20)
for i in range(batch_size):
font_name = random.choice(FontNames)
font_length = random.randint(font_min_length - 5, font_min_length + 5)
font_size = random.randint(9, 64)
text, image = utils.getImage(CHARS, font_name, image_height,
font_length, font_size, eng_world_list)
images.append(image)
if image.shape[1] > max_width_image:
max_width_image = image.shape[1]
to_image = utils.renderNormalFontByPIL(ConsolasFont, 64, text)
to_image = utils.trim(to_image)
w, h = to_image.size
_w = round(w * image_height / h)
_h = image_height
if _w > max_width_image:
_w = max_width_image
_h = round(h * max_width_image / w)
to_image = to_image.resize((_w, _h), Image.ANTIALIAS)
to_image = np.asarray(to_image)
#to_image=utils.resize(to_image, height=image_height)
to_image = utils.img2gray(to_image)
to_image = to_image / 255
to_images.append(to_image)
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.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)
labels[i, :] = np.transpose(image_vec)
return inputs, labels
def get_batch(batch_size=128):
batch_x = np.zeros([batch_size, image_size])
batch_y = np.zeros([batch_size, captcha_size])
for i in range(batch_size):
text, image = captcha(
char_set=char_set, captcha_size=captcha_size, width=image_w, height=image_h)
batch_x[i, :] = img2vec(img2gray(image))
batch_y[i, :] = list(text) # 注意 这里的 lable 不能 one hot
return batch_x, batch_y
def get_batch(batch_size=128):
batch_x = np.zeros([batch_size, image_size])
batch_y = np.zeros([batch_size, char_size])
for i in range(batch_size):
text, image = captcha(char_set=char_set,
captcha_size=1,
width=image_w,
height=image_h)
batch_x[i, :] = img2vec(img2gray(image))
batch_y[i, :] = text2vec(char_set, text)
return batch_x, batch_y
def get_batch(batch_size=128):
batch_files = random.sample(files, batch_size)
batch_x = np.zeros([batch_size, image_size])
batch_y = np.zeros([batch_size, captcha_size * char_size])
for i in range(batch_size):
image = loadimg(os.path.join(fonts_dir, batch_files[i]))
text = batch_files[i][:4]
batch_x[i, :] = img2vec(img2gray(image))
batch_y[i, :] = text2vec(char_set, text)
return batch_x, batch_y
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 scan():
session, inputs, seq_len, input_keep_prob, decoded, log_prob = init()
need_ocr_images = utils.loadImage(os.path.join(curr_dir, 'test', '1.jpg'),
0)
ocr_text_groups = []
for idx, images_group in enumerate(need_ocr_images):
# if idx != 1: continue
ocr_texts = []
# 取最大宽度做为本组的统一输入长度
widths = [image.shape[1] for image in images_group]
maxImageWidth = max(widths) + 5
ocr_inputs = np.zeros(
[len(images_group), maxImageWidth, ocr.image_height])
for i, image in enumerate(images_group):
# image = utils.dropZeroEdges(image)
# utils.show(utils.dropZeroEdges(image))
utils.save(image,
os.path.join(curr_dir, "test", "%s-%s.png" % (idx, i)))
image_vec = utils.img2vec(image, ocr.image_height, maxImageWidth)
ocr_inputs[i, :] = np.transpose(
image_vec.reshape((ocr.image_height, maxImageWidth)))
# utils.show(image)
# return
ocr_seq_len = np.ones(ocr_inputs.shape[0]) * maxImageWidth
feed = {inputs: ocr_inputs, seq_len: ocr_seq_len, input_keep_prob: 1.0}
print("starting ocr inputs %s ..." % idx)
start = time.time()
decoded_list = session.run(decoded[0], feed)
seconds = round(time.time() - start, 2)
print("filished ocr inputs %s, paid %s seconds" % (idx, seconds))
detected_list = ocr.decode_sparse_tensor(decoded_list)
for detect_number in detected_list:
ocr_texts.append(ocr.list_to_chars(detect_number))
ocr_text_groups.append(ocr_texts)
# break
return ocr_text_groups
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):
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
checkpoint_prefix = os.path.join(log_dir, "model.ckpt")
# 找到最新的运算模型文件
metaFile= sorted(
[
(x, os.path.getctime(os.path.join(log_dir,x)))
for x in os.listdir(log_dir) if x.endswith('.meta')
],
key=lambda i: i[1])[-1][0]
sess = tf.Session()
saver = tf.train.import_meta_graph(os.path.join(log_dir,metaFile))
ckpt = tf.train.get_checkpoint_state(log_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print("error: can't load checkpoint data")
exit()
text, image = captcha(char_set="0123456789", captcha_size=4, width=200, height=80)
image = img2vec(img2gray(image))
x = tf.get_default_graph().get_tensor_by_name('x:0')
prediction = tf.get_default_graph().get_tensor_by_name('prediction:0')
y_ = sess.run([prediction], feed_dict={x: [image]})
result = "".join([str(x) for x in y_[0][0]])
print("input: %s , get: %s"%(text,result))
