tensorflow 实现端到端的OCR：二代身份证号识别
最近在研究OCR识别相关的东西，最终目标是能识别身份证上的所有中文汉字+数字，不过本文先设定一个小目标，先识别定长为18的身份证号，当然本文的思路也是可以复用来识别定长的验证码识别的。
本文实现思路主要来源于Xlvector的博客，采用基于CNN实现端到端的OCR，下面引用博文介绍目前基于深度学习的两种OCR识别方法：
- 把OCR的问题当做一个多标签学习的问题。4个数字组成的验证码就相当于有4个标签的图片识别问题（这里的标签还是有序的），用CNN来解决。
- 把OCR的问题当做一个语音识别的问题，语音识别是把连续的音频转化为文本，验证码识别就是把连续的图片转化为文本，用CNN+LSTM+CTC来解决。
这里方法1主要用来解决固定长度标签的图片识别问题，而方法2主要用来解决不定长度标签的图片识别问题，本文实现方法1识别固定18个数字字符的身份证号
环境依赖
- 本文基于tensorflow框架实现,依赖于tensorflow环境，建议使用anaconda进行python包管理及环境管理
- 本文使用freetype-py 进行训练集图片的实时生成，同时后续也可扩展为能生成中文字符图片的训练集，建议使用pip安装
pip install freetype-py
- 同时本文还依赖于numpy和opencv等常用库
pip install numpy cv2
知识准备
- 本文不具体介绍CNN (卷积神经网络)具体实现原理，不熟悉的建议参看集智博文卷积：如何成为一个很厉害的神经网络，这篇文章写得很?
- 本文实现思路很容易理解，就是把一个有序排列18个数字组成的图片当做一个多标签学习的问题，标签的长度可以任意改变，只要是固定长度的，这个训练方法都是适用的，当然现实中很多情况是需要识别不定长度的标签的，这部分就需要使用方法2(CNN+lSTM+CTC)来解决了。
正文
训练数据集生成
首先先完成训练数据集图片的生成，主要依赖于freetype-py库生成数字/中文的图片。其中要注意的一点是就是生成图片的大小，本文经过多次尝试后，生成的图片是32 x 256大小的，如果图片太大，则可能导致训练不收敛
生成出来的示例图片如下：
image.png
gen_image()方法返回
image_data：图片像素数据 (32,256)
label： 图片标签 18位数字字符 477081933151463759
vec : 图片标签转成向量表示 (180,) 代表每个数字所处的列，总长度 18 * 10
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
身份证文字+数字生成类
@author: pengyuanjie
"""
import numpy as np
import freetype
import copy
import random
import cv2
class put_chinese_text(object):
def __init__(self, ttf):
self._face = freetype.Face(ttf)
def draw_text(self, image, pos, text, text_size, text_color):
'''
draw chinese(or not) text with ttf
:param image: image(numpy.ndarray) to draw text
:param pos: where to draw text
:param text: the context, for chinese should be unicode type
:param text_size: text size
:param text_color:text color
:return: image
'''
self._face.set_char_size(text_size * 64)
metrics = self._face.size
ascender = metrics.ascender/64.0
#descender = metrics.descender/64.0
#height = metrics.height/64.0
#linegap = height - ascender + descender
ypos = int(ascender)
if not isinstance(text, unicode):
text = text.decode('utf-8')
img = self.draw_string(image, pos[0], pos[1]+ypos, text, text_color)
return img
def draw_string(self, img, x_pos, y_pos, text, color):
'''
draw string
:param x_pos: text x-postion on img
:param y_pos: text y-postion on img
:param text: text (unicode)
:param color: text color
:return: image
'''
prev_char = 0
pen = freetype.Vector()
pen.x = x_pos << 6 # div 64
pen.y = y_pos << 6
hscale = 1.0
matrix = freetype.Matrix(int(hscale)*0x10000L, int(0.2*0x10000L),\
int(0.0*0x10000L), int(1.1*0x10000L))
cur_pen = freetype.Vector()
pen_translate = freetype.Vector()
image = copy.deepcopy(img)
for cur_char in text:
self._face.set_transform(matrix, pen_translate)
self._face.load_char(cur_char)
kerning = self._face.get_kerning(prev_char, cur_char)
pen.x += kerning.x
slot = self._face.glyph
bitmap = slot.bitmap
cur_pen.x = pen.x
cur_pen.y = pen.y - slot.bitmap_top * 64
self.draw_ft_bitmap(image, bitmap, cur_pen, color)
pen.x += slot.advance.x
prev_char = cur_char
return image
def draw_ft_bitmap(self, img, bitmap, pen, color):
'''
draw each char
:param bitmap: bitmap
:param pen: pen
:param color: pen color e.g.(0,0,255) - red
:return: image
'''
x_pos = pen.x >> 6
y_pos = pen.y >> 6
cols = bitmap.width
rows = bitmap.rows
glyph_pixels = bitmap.buffer
for row in range(rows):
for col in range(cols):
if glyph_pixels[row*cols + col] != 0:
img[y_pos + row][x_pos + col][0] = color[0]
img[y_pos + row][x_pos + col][1] = color[1]
img[y_pos + row][x_pos + col][2] = color[2]
class gen_id_card(object):
def __init__(self):
#self.words = open('AllWords.txt', 'r').read().split(' ')
self.number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
self.char_set = self.number
#self.char_set = self.words + self.number
self.len = len(self.char_set)
self.max_size = 18
self.ft = put_chinese_text('fonts/OCR-B.ttf')
#随机生成字串，长度固定
#返回text,及对应的向量
def random_text(self):
text = ''
vecs = np.zeros((self.max_size * self.len))
#size = random.randint(1, self.max_size)
size = self.max_size
for i in range(size):
c = random.choice(self.char_set)
vec = self.char2vec(c)
text = text + c
vecs[i*self.len:(i+1)*self.len] = np.copy(vec)
return text,vecs
#根据生成的text，生成image,返回标签和图片元素数据
def gen_image(self):
text,vec = self.random_text()
img = np.zeros([32,256,3])
color_ = (255,255,255) # Write
pos = (0, 0)
text_size = 21
image = self.ft.draw_text(img, pos, text, text_size, color_)
#仅返回单通道值，颜色对于汉字识别没有什么意义
return image[:,:,2],text,vec
#单字转向量
def char2vec(self, c):
vec = np.zeros((self.len))
for j in range(self.len):
if self.char_set[j] == c:
vec[j] = 1
return vec
#向量转文本
def vec2text(self, vecs):
text = ''
v_len = len(vecs)
for i in range(v_len):
if(vecs[i] == 1):
text = text + self.char_set[i % self.len]
return text
if __name__ == '__main__':
genObj = gen_id_card()
image_data,label,vec = genObj.gen_image()
cv2.imshow('image', image_data)
cv2.waitKey(0)
构建网络，开始训练
首先定义生成一个batch的方法：
# 生成一个训练batch
def get_next_batch(batch_size=128):
obj = gen_id_card()
batch_x = np.zeros([batch_size, IMAGE_HEIGHT*IMAGE_WIDTH])
batch_y = np.zeros([batch_size, MAX_CAPTCHA*CHAR_SET_LEN])
for i in range(batch_size):
image, text, vec = obj.gen_image()
batch_x[i,:] = image.reshape((IMAGE_HEIGHT*IMAGE_WIDTH))
batch_y[i,:] = vec
return batch_x, batch_y
用了Batch Normalization，个人还不是很理解，读者可自行百度，代码来源于参考博文
#Batch Normalization? 有空再理解,tflearn or slim都有封装
## http://stackoverflow.com/a/34634291/2267819
def batch_norm(x, beta, gamma, phase_train, scope='bn', decay=0.9, eps=1e-5):
with tf.variable_scope(scope):
#beta = tf.get_variable(name='beta', shape=[n_out], initializer=tf.constant_initializer(0.0), trainable=True)
#gamma = tf.get_variable(name='gamma', shape=[n_out], initializer=tf.random_normal_initializer(1.0, stddev), trainable=True)
batch_mean, batch_var = tf.nn.moments(x, [0, 1, 2], name='moments')
ema = tf.train.ExponentialMovingAverage(decay=decay)
def mean_var_with_update():
ema_apply_op = ema.apply([batch_mean, batch_var])
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
mean, var = tf.cond(phase_train, mean_var_with_update, lambda: (ema.average(batch_mean), ema.average(batch_var)))
normed = tf.nn.batch_normalization(x, mean, var, beta, gamma, eps)
return normed
定义4层CNN和一层全连接层，卷积核分别是2层5x5、2层3x3，每层均使用tf.nn.relu非线性化,并使用max_pool，网络结构读者可自行调参优化
# 定义CNN
def crack_captcha_cnn(w_alpha=0.01, b_alpha=0.1):
x = tf.reshape(X, shape=[-1, IMAGE_HEIGHT, IMAGE_WIDTH, 1])
# 4 conv layer
w_c1 = tf.Variable(w_alpha*tf.random_normal([5, 5, 1, 32]))
b_c1 = tf.Variable(b_alpha*tf.random_normal([32]))
conv1 = tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1)
conv1 = batch_norm(conv1, tf.constant(0.0, shape=[32]), tf.random_normal(shape=[32], mean=1.0, stddev=0.02), train_phase, scope='bn_1')
conv1 = tf.nn.relu(conv1)
conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv1 = tf.nn.dropout(conv1, keep_prob)
w_c2 = tf.Variable(w_alpha*tf.random_normal([5, 5, 32, 64]))
b_c2 = tf.Variable(b_alpha*tf.random_normal([64]))
conv2 = tf.nn.bias_add(tf.nn.conv2d(conv1, w_c2, strides=[1, 1, 1, 1], padding='SAME'), b_c2)
conv2 = batch_norm(conv2, tf.constant(0.0, shape=[64]), tf.random_normal(shape=[64], mean=1.0, stddev=0.02), train_phase, scope='bn_2')
conv2 = tf.nn.relu(conv2)
conv2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv2 = tf.nn.dropout(conv2, keep_prob)
w_c3 = tf.Variable(w_alpha*tf.random_normal([3, 3, 64, 64]))
b_c3 = tf.Variable(b_alpha*tf.random_normal([64]))
conv3 = tf.nn.bias_add(tf.nn.conv2d(conv2, w_c3, strides=[1, 1, 1, 1], padding='SAME'), b_c3)
conv3 = batch_norm(conv3, tf.constant(0.0, shape=[64]), tf.random_normal(shape=[64], mean=1.0, stddev=0.02), train_phase, scope='bn_3')
conv3 = tf.nn.relu(conv3)
conv3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv3 = tf.nn.dropout(conv3, keep_prob)
w_c4 = tf.Variable(w_alpha*tf.random_normal([3, 3, 64, 64]))
b_c4 = tf.Variable(b_alpha*tf.random_normal([64]))
conv4 = tf.nn.bias_add(tf.nn.conv2d(conv3, w_c4, strides=[1, 1, 1, 1], padding='SAME'), b_c4)
conv4 = batch_norm(conv4, tf.constant(0.0, shape=[64]), tf.random_normal(shape=[64], mean=1.0, stddev=0.02), train_phase, scope='bn_4')
conv4 = tf.nn.relu(conv4)
conv4 = tf.nn.max_pool(conv4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv4 = tf.nn.dropout(conv4, keep_prob)
# Fully connected layer
w_d = tf.Variable(w_alpha*tf.random_normal([2*16*64, 1024]))
b_d = tf.Variable(b_alpha*tf.random_normal([1024]))
dense = tf.reshape(conv4, [-1, w_d.get_shape().as_list()[0]])
dense = tf.nn.relu(tf.add(tf.matmul(dense, w_d), b_d))
dense = tf.nn.dropout(dense, keep_prob)
w_out = tf.Variable(w_alpha*tf.random_normal([1024, MAX_CAPTCHA*CHAR_SET_LEN]))
b_out = tf.Variable(b_alpha*tf.random_normal([MAX_CAPTCHA*CHAR_SET_LEN]))
out = tf.add(tf.matmul(dense, w_out), b_out)
return out
最后执行训练，使用sigmoid分类，每100次计算一次准确率，如果准确率超过80%，则保存模型并结束训练
# 训练
def train_crack_captcha_cnn():
output = crack_captcha_cnn()
# loss
#loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=output, labels=Y))
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=output, labels=Y))
# 最后一层用来分类的softmax和sigmoid有什么不同？
# optimizer 为了加快训练 learning_rate应该开始大，然后慢慢衰
optimizer = tf.train.AdamOptimizer(learning_rate=0.002).minimize(loss)
predict = tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN])
max_idx_p = tf.argmax(predict, 2)
max_idx_l = tf.argmax(tf.reshape(Y, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2)
correct_pred = tf.equal(max_idx_p, max_idx_l)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
step = 0
while True:
batch_x, batch_y = get_next_batch(64)
_, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75, train_phase:True})
print(step, loss_)
# 每100 step计算一次准确率
if step % 100 == 0 and step != 0:
batch_x_test, batch_y_test = get_next_batch(100)
acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1., train_phase:False})
print "第%s步，训练准确率为：%s" % (step, acc)
# 如果准确率大80%,保存模型,完成训练
if acc > 0.8:
saver.save(sess, "crack_capcha.model", global_step=step)
break
step += 1
执行结果，笔者在大概500次训练后，得到准确率84.3%的结果
image.png
大概训练1500~2200次左右，准确率就能达到98%，打印前5条测试样本可以看出，输出结果基本与label一致了
image.png
后记
最后所有代码和字体资源文件托管在我的Github下
笔者在一开始训练的时候图片大小是64 x 512的，训练的时候发现训练速度很慢，而且训练的loss不收敛一直保持在0.33左右，缩小图片为32 x 256后解决，不知道为啥，猜测要么是网络层级不够，或者特征层数不够吧。
小目标完成后，为了最终目标的完成，后续可能尝试方法2，去识别不定长的中文字符图片，不过要先去理解LSTM网络和 CTC模型了。
参考链接
TensorFlow练习20: 使用深度学习破解字符验证码
Python2.x上使用freetype实现OpenCV2.x的中文输出
端到端的OCR：基于CNN的实现