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):
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 scan(file):
img = Image.open(file.stream)
image = np.array(img)
image = utils.img2gray(image)
utils.save(image * 255, os.path.join(curr_dir, "test", "p0.png"))
# image = utils.clearImgGray(image)
# utils.save(image * 255, os.path.join(curr_dir,"test","p1.png"))
split_images = utils.splitImg(image)
ocr_texts = []
for i, split_image in enumerate(split_images):
inv_image = utils.img2bwinv(split_image)
inv_image = utils.clearImg(inv_image)
image = 255. - split_image
image = utils.dropZeroEdges(inv_image, image)
image = utils.resize(image, ocr.image_height)
image = image / 255.
ocr_inputs = np.zeros([1, ocr.image_size, ocr.image_size])
ocr_inputs[0, :] = utils.square_img(
image, np.zeros([ocr.image_size, ocr.image_size]))
ocr_seq_len = np.ones(1) * (ocr.image_size * ocr.image_size) // (
ocr.POOL_SIZE * ocr.POOL_SIZE)
start = time.time()
p_net_g = session.run(net_g, {inputs: ocr_inputs})
p_net_g = np.squeeze(p_net_g, axis=3)
debug_net_g = np.copy(p_net_g)
for j in range(1):
_t_img = utils.unsquare_img(p_net_g[j], ocr.image_height)
_t_img_bin = np.copy(_t_img)
_t_img_bin[_t_img_bin <= 0.2] = 0
_t_img = utils.dropZeroEdges(_t_img_bin, _t_img, min_rate=0.1)
_t_img = utils.resize(_t_img, ocr.image_height)
if _t_img.shape[0] * _t_img.shape[
1] <= ocr.image_size * ocr.image_size:
p_net_g[j] = utils.square_img(
_t_img, np.zeros([ocr.image_size, ocr.image_size]),
ocr.image_height)
_img = np.vstack((ocr_inputs[0], debug_net_g[0], p_net_g[0]))
utils.save(_img * 255, os.path.join(curr_dir, "test", "%s.png" % i))
decoded_list = session.run(res_decoded[0], {
inputs: p_net_g,
seq_len: ocr_seq_len
})
seconds = round(time.time() - start, 2)
print("filished ocr %s , paid %s seconds" % (i, seconds))
detected_list = utils.decode_sparse_tensor(decoded_list)
for detect_number in detected_list:
ocr_texts.append(ocr.list_to_chars(detect_number))
return ocr_texts
def getImage(text, font_name, font_length, font_size, noise=False, fontmode=None, fonthint=None):
params= {}
params['text'] = text
params['fontname'] = font_name
params['fontsize'] = font_size
# params['fontmode'] = random.choice([0,1,2,4,8])
if fontmode == None:
params['fontmode'] = random.choice([0,1,2,4])
else:
params['fontmode'] = fontmode
if fonthint == None:
params['fonthint'] = random.choice([0,1,2,3,4,5])
else:
params['fonthint'] = fonthint
r = http('http://192.168.2.113:8888/',params)
_img = Image.open(io.BytesIO(r))
img=Image.new("RGB",_img.size,(255,255,255))
img.paste(_img,(0,0),_img)
img = utils.trim(img)
if noise:
w,h = img.size
_h = random.randint(9, image_height)
_w = round(w * _h / h)
img = img.resize((_w,_h), Image.ANTIALIAS)
img = np.asarray(img)
img = 1 - utils.img2gray(img)/255.
img = utils.dropZeroEdges(img)
filter = np.random.random(img.shape) - 0.9
filter = np.maximum(filter, 0)
img = img + filter * 5
imin, imax = img.min(), img.max()
img = (img - imin)/(imax - imin)
else:
img = np.asarray(img)
img = utils.img2gray(img)
img = utils.img2bwinv(img)
img = img / 255.
img = utils.dropZeroEdges(img)
return img
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
os.mkdir(os.path.join(curr_dir, "data" ,"dataset"))
with open(os.path.join(curr_dir, "data", "index.txt")) as index_file:
for i, line in enumerate(index_file.readlines()):
if i%10000==0: print("resizing image no: ",i)
lines = line.split(" ")
image_name = lines[0]+".png"
dst_image_name = os.path.join(curr_dir, "data" ,"dataset", image_name)
if os.path.exists(dst_image_name):
train_files.append(line)
continue
if not os.path.exists(os.path.dirname(dst_image_name)):
os.mkdir(os.path.dirname(dst_image_name))
src_image_name = os.path.join(curr_dir,"data",image_name)
try:
image = readImgFile(src_image_name)
image = img2bwinv(image)
image = dropZeroEdges(image)
except:
print(dst_image_name,"error")
continue
resized_image = resize(image,image_height)
save(resized_image,dst_image_name)
train_files.append(line)
def neural_networks():
# 输入：训练的数量，一张图片的宽度，一张图片的高度 [-1,-1,16]
inputs = tf.placeholder(tf.float32, [None, None, image_height], name="inputs")
# 定义 ctc_loss 是稀疏矩阵
labels = tf.sparse_placeholder(tf.int32, name="labels")
# 1维向量 size [batch_size] 等于 np.ones(batch_size)* image_width
