OCR是一个古老的问题。这里我们考虑一类特殊的OCR问题，就是验证码的识别。传统做验证码的识别，需要经过如下步骤：

1. 二值化

2. 字符分割

3. 字符识别

这里最难的就是分割。如果字符之间有粘连，那分割起来就无比痛苦了。

最近研究深度学习，发现有人做端到端的OCR。于是准备尝试一下。一般来说目前做基于深度学习的OCR大概有如下套路：

1. 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。

2. 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。

目前第1种方法可以做到90%多的准确率(4个都猜对了才算对)，第二种方法我目前的实验还只能到20%多，还在研究中。所以这篇文章先介绍第一种方法。

我们以python-captcha验证码的识别为例来做验证码识别。

下图是一些这个验证码的例子：

python-captcha

可以看到这里面有粘连，也有形变，噪音。所以我们可以看看用CNN识别这个验证码的效果。

首先，我们定义一个迭代器来输入数据，这里我们每次都直接调用python-captcha这个库来根据随机生成的label来生成相应的验证码图片。这样我们的训练集相当于是无穷大的。

class OCRIter(mx.io.DataIter):

def __init__(self, count, batch_size, num_label, height, width):

super(OCRIter, self).__init__()

self.captcha = ImageCaptcha(fonts=['./data/OpenSans-Regular.ttf'])

self.batch_size = batch_size

self.count = count

self.height = height

self.width = width

self.provide_data = [('data', (batch_size, 3, height, width))]

self.provide_label = [('softmax_label', (self.batch_size, num_label))]

def __iter__(self):

for k in range(self.count / self.batch_size):

data = []

label = []

for i in range(self.batch_size):

# 生成一个四位数字的随机字符串

num = gen_rand()

# 生成随机字符串对应的验证码图片

img = self.captcha.generate(num)

img = np.fromstring(img.getvalue(), dtype='uint8')

img = cv2.imdecode(img, cv2.IMREAD_COLOR)

img = cv2.resize(img, (self.width, self.height))

cv2.imwrite("./tmp" + str(i % 10) + ".png", img)

img = np.multiply(img, 1/255.0)

img = img.transpose(2, 0, 1)

data.append(img)

label.append(get_label(num))

data_all = [mx.nd.array(data)]

label_all = [mx.nd.array(label)]

data_names = ['data']

label_names = ['softmax_label']

data_batch = OCRBatch(data_names, data_all, label_names, label_all)

yield data_batch

def reset(self):

pass

然后我们用如下的网络来训练这个数据集：

def get_ocrnet():

data = mx.symbol.Variable('data')

label = mx.symbol.Variable('softmax_label')

conv1 = mx.symbol.Convolution(data=data, kernel=(5,5), num_filter=32)

pool1 = mx.symbol.Pooling(data=conv1, pool_type="max", kernel=(2,2), stride=(1, 1))

relu1 = mx.symbol.Activation(data=pool1, act_type="relu")

conv2 = mx.symbol.Convolution(data=relu1, kernel=(5,5), num_filter=32)

pool2 = mx.symbol.Pooling(data=conv2, pool_type="avg", kernel=(2,2), stride=(1, 1))

relu2 = mx.symbol.Activation(data=pool2, act_type="relu")

conv3 = mx.symbol.Convolution(data=relu2, kernel=(3,3), num_filter=32)

pool3 = mx.symbol.Pooling(data=conv3, pool_type="avg", kernel=(2,2), stride=(1, 1))

relu3 = mx.symbol.Activation(data=pool3, act_type="relu")

flatten = mx.symbol.Flatten(data = relu3)

fc1 = mx.symbol.FullyConnected(data = flatten, num_hidden = 512)

fc21 = mx.symbol.FullyConnected(data = fc1, num_hidden = 10)

fc22 = mx.symbol.FullyConnected(data = fc1, num_hidden = 10)

fc23 = mx.symbol.FullyConnected(data = fc1, num_hidden = 10)

fc24 = mx.symbol.FullyConnected(data = fc1, num_hidden = 10)

fc2 = mx.symbol.Concat(*[fc21, fc22, fc23, fc24], dim = 0)

label = mx.symbol.transpose(data = label)

label = mx.symbol.Reshape(data = label, target_shape = (0, ))

return mx.symbol.SoftmaxOutput(data = fc2, label = label, name = "softmax")

上面这个网络要稍微解释一下。因为这个问题是一个有顺序的多label的图片分类问题。我们在fc1的层上面接了4个Full Connect层(fc21,fc22,fc23,fc24)，用来对应不同位置的4个数字label。然后将它们Concat在一起。然后同时学习这4个label。目前用上面的网络训练，4位数字全部预测正确的精度可以达到90%左右。

更新，经过比较长时间的训练，精度可以达到98%左右，最后几轮迭代的结果如下：

2016-05-22 21:58:34,859 Epoch[14] Batch [1250] Speed: 117.29 samples/sec Train-Accuracy=0.980800

2016-05-22 21:58:48,527 Epoch[14] Batch [1300] Speed: 117.06 samples/sec Train-Accuracy=0.982000

2016-05-22 21:59:02,174 Epoch[14] Batch [1350] Speed: 117.24 samples/sec Train-Accuracy=0.981200

2016-05-22 21:59:16,509 Epoch[14] Batch [1400] Speed: 111.62 samples/sec Train-Accuracy=0.976800

2016-05-22 21:59:31,031 Epoch[14] Batch [1450] Speed: 110.18 samples/sec Train-Accuracy=0.975600

2016-05-22 21:59:45,323 Epoch[14] Batch [1500] Speed: 111.95 samples/sec Train-Accuracy=0.975600

2016-05-22 21:59:59,634 Epoch[14] Batch [1550] Speed: 111.81 samples/sec Train-Accuracy=0.985600

2016-05-22 22:00:13,997 Epoch[14] Batch [1600] Speed: 111.39 samples/sec Train-Accuracy=0.978800

2016-05-22 22:00:28,270 Epoch[14] Batch [1650] Speed: 112.11 samples/sec Train-Accuracy=0.983200

2016-05-22 22:00:42,713 Epoch[14] Batch [1700] Speed: 110.78 samples/sec Train-Accuracy=0.985200

2016-05-22 22:00:56,668 Epoch[14] Batch [1750] Speed: 114.65 samples/sec Train-Accuracy=0.975600

2016-05-22 22:01:11,000 Epoch[14] Batch [1800] Speed: 111.64 samples/sec Train-Accuracy=0.981200

2016-05-22 22:01:25,450 Epoch[14] Batch [1850] Speed: 110.73 samples/sec Train-Accuracy=0.979600

2016-05-22 22:01:39,860 Epoch[14] Batch [1900] Speed: 111.03 samples/sec Train-Accuracy=0.978400

2016-05-22 22:01:54,272 Epoch[14] Batch [1950] Speed: 111.02 samples/sec Train-Accuracy=0.978800

2016-05-22 22:02:08,939 Epoch[14] Batch [2000] Speed: 109.09 samples/sec Train-Accuracy=0.981600

2016-05-22 22:02:08,939 Epoch[14] Resetting Data Iterator

2016-05-22 22:02:08,939 Epoch[14] Time cost=568.681

2016-05-22 22:02:14,124 Epoch[14] Validation-Accuracy=0.986000

另外这个Slide提供了关于深度学习进行验证码识别的详细描述。