神经网络实现Discuz验证码识别

最近自己尝试了网上的验证码识别代码项目，该小项目见以下链接：
>https://cuijiahua.com/blog/2018/01/dl_5.html

数据也就用了作者上传的60000张Discuz验证码。作者是创建了一个 类 封装了所有的变量和函数，我看了他的代码之后自己尝试着不用类去实现该网络。

作者说自己可以训练到90%以上的精度。然而我看了他的代码后发现，作者是用训练过的数据来进行测试，即训练集和测试集是一样的。

我想着，测试集应该是不能参与训练过程中的，比如说我们在做mnist手写数字识别的时候，训练集与测试集就一定是不一样的。

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=False)

Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz

于是我在自己实现的过程中，将数据集打乱后取10000个作为测试集，不参与训练，剩余的50000张验证码作为训练集。

训练过程中发现只有将学习率设置为0.001时，loss才会降下去，太高，loss会卡在0.07；其次，我的训练精度最多只能到50%左右，但是我用训练数据来测试保存的模型，精度确实达到了90%，即作者看到的精度。不过这个模型不具有泛化能力，它在没见过的测试集上只有50%的精确度。

同时这个代码还有问题：测算精确度时，一张图中4个验证码有两个错误的话，正确率是50%而不是0.当一张图中4个验证码识别有一个错误时，该验证码识别就应该是失败的。因此这个精确度实在是有相当大的水分。

于是要考虑解决办法。首先我尝试着下调学习率，发现永远还是到50%就上不去了。

接下来我在原来的3层卷积层上，又加了一层卷积层。然而这并没有提升多少精度。

随后我又加入了一层全连接层，期望可以拟合得更好一些，但是这样让我陷入了麻烦。

我的loss值卡在了0.07，无论我的学习率是0.1还是0.00001.哪怕迭代一百万次也是如此。这时候的测试精度只有······3%。

我不知道是什么问题更不知道如何改进。

这更让我觉得没有人带我，多么地难受；同时也更深刻地体验到理论知识是多么地重要（当然我一直知道）。

我自己的代码附上，大家可以相互交流。数据可以在文章顶部的链接里下载，作者压缩好的。

以下是训练脚本：（理论上python3和python2应该都能跑。我是用2写的）

训练中我使用了学习率衰减，本来还想用dropout结果发现这个训练基本不给我过拟合的机会所以训练加了没有意义。

from __future__ import print_function, division, absolute_import
import tensorflow as tf
import os
import cv2
import matplotlib.pyplot as plt
import random
import numpy as np
from optparse import OptionParser

path = 'Discuz/' #存放数据的路径
imgs = os.listdir(path) #以列表形式读取所有图片名称
random.shuffle(imgs) #打乱
max_steps = 1000000 #最大迭代步数
save_path = 'model4cnn-1fcn' #保存模型的路径，会自动生成
dropout = 1 #没用到

trainnum = 50000 #定义训练集和测试集的大小
testnum = 10000

traindatas = imgs[:trainnum] #取出训练集和测试集及其标签
trainlabels = list(map(lambda x: x.split('.')[0],traindatas))

testdatas = imgs[trainnum:]
testlabels = list(map(lambda x: x.split('.')[0],testdatas))

#定义取数据集的指针
train_ptr = 0
test_ptr = 0

def next_batch(batch=100, train_flag=True):
	global train_ptr
	global test_ptr
	batch_x = np.zeros([batch,30*100])
	batch_y = np.zeros([batch, 4*63])

	if train_flag == True:
		if batch + train_ptr < trainnum:
			trains = traindatas[train_ptr:(train_ptr+batch)]
			labels = trainlabels[train_ptr:(train_ptr+batch)]
			train_ptr += batch
		else:
			new_ptr = (train_ptr + batch) % trainnum
			trains = traindatas[train_ptr:] + traindatas[:new_ptr]
			labels = trainlabels[train_ptr:] + traindatas[:new_ptr]
			train_ptr = new_ptr

		for index, train in enumerate(trains):
			img = np.mean(cv2.imread(path + train), -1)
			batch_x[index,:] = img.flatten() /255
		for index, label in enumerate(labels):
			batch_y[index,:] = text2vec(label)

	else:
		if batch + test_ptr < testnum:
			tests = testdatas[test_ptr:(test_ptr+batch)]
			labels = testlabels[test_ptr:(test_ptr+batch)]
			test_ptr += batch
		else:
			new_ptr = (test_ptr + batch) % testnum
			tests = testdatas[test_ptr:] + testdatas[:new_ptr]
			labels = testlabels[test_ptr:] + testlabels[:new_ptr]
			test_ptr = new_ptr

		for index, test in enumerate(tests):
			img = np.mean(cv2.imread(path + test), -1)
			batch_x[index, :] = img.flatten() /255
		for index, label in enumerate(labels):
			batch_y[index,:] = text2vec(label)

	return batch_x, batch_y

def text2vec(text):
	if len(text) > 4:
		raise ValueError('too long captcha')

	vector = np.zeros(4*63)
	def char2pos(c):
		if c == '_':
			k = 62
			return k
		k = ord(c)-48
		if k > 9:
			k = ord(c)-55
			if k > 35:
				k = ord(c) - 61
				if k > 61:
					raise ValueError('No Map')

		return k

	for i, c in enumerate(text):
		idx = i*63 + char2pos(c)
		vector[idx] = 1

	return vector

X = tf.placeholder(tf.float32, [None, 30*100])
Y = tf.placeholder(tf.float32, [None,4*63])
_lr = tf.placeholder(tf.float32)
keep_prob = tf.placeholder(tf.float32)

def conv2d(x, W, b, strides=1):
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)

def max_pool2d(x, k=2):
    x = tf.nn.max_pool(
        x, ksize=[
            1, k, k, 1], strides=[
            1, k, k, 1], padding='SAME')
    return x

weights = {
        'wc1': tf.Variable(0.01*tf.random_normal([3, 3, 1, 32])),
        'wc2': tf.Variable(0.01*tf.random_normal([3, 3, 32, 64])),
        'wc3': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
        'wc4': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
        'wf1': tf.Variable(0.01*tf.random_normal([2 * 7 * 64, 1024])),
        'wf2': tf.Variable(0.01*tf.random_normal([1024, 1024])),
        'wout': tf.Variable(0.01*tf.random_normal([1024, 4*63]))
        }

biases = {
        'bc1': tf.Variable(0.1*tf.random_normal([32])),
        'bc2': tf.Variable(0.1*tf.random_normal([64])),
        'bc3': tf.Variable(0.1*tf.random_normal([64])),
        'bc4': tf.Variable(0.1*tf.random_normal([64])),
        'bf1': tf.Variable(0.1*tf.random_normal([1024])),
        'bf2': tf.Variable(0.1*tf.random_normal([1024])),
        'bout': tf.Variable(0.1*tf.random_normal([4*63]))
    }

def conv_net(x, weights, biases, dropout):
	x = tf.reshape(x, [-1,100,30,1])

	conv1 = conv2d(x, weights['wc1'], biases['bc1'], 1)
	conv1 = max_pool2d(conv1, 2)

	conv2 = conv2d(conv1, weights['wc2'], biases['bc2'], 1)
	conv2 = max_pool2d(conv2, 2)

	conv3 = conv2d(conv2, weights['wc3'], biases['bc3'], 1)
	conv3 = max_pool2d(conv3, 2)

	conv4 = conv2d(conv3, weights['wc4'], biases['bc4'], 1)
	conv4 = max_pool2d(conv4, 2)

	fc1 = tf.reshape(
        conv4, shape=[-1, weights['wf1'].get_shape().as_list()[0]])
	fc1 = tf.matmul(fc1, weights['wf1'])
	fc1 = tf.add(fc1, biases['bf1'])
	fc1 = tf.nn.relu(fc1)

	out = tf.add(tf.matmul(fc1, weights['wout']), biases['bout'])

	return out

output = conv_net(X, weights, biases, keep_prob)

loss_op = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
            logits=output, labels=Y))
optimizer = tf.train.AdamOptimizer(learning_rate=_lr).minimize(loss_op)

y = tf.reshape(output, [-1,4,63])
y_ = tf.reshape(Y, [-1,4,63])

correct_pred = tf.equal(tf.argmax(y, 2), tf.argmax(y_,2))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
init = tf.global_variables_initializer()
lr = 0.001
saver = tf.train.Saver()
with tf.Session() as sess:
	sess.run(init)
	for step in range(1,1+max_steps):
		batch_x, batch_y = next_batch(100,True)
		loss_value,_ = sess.run([loss_op, optimizer],
			feed_dict = {X:batch_x, Y:batch_y, keep_prob:dropout,_lr:lr})
		if step % 10 == 0:
			batch_x_test, batch_y_test = next_batch(100, False)
			acc = sess.run(accuracy,
				feed_dict={X:batch_x_test, Y:batch_y_test,keep_prob:1})
			print('step{}, loss={}, accuracy={}'.format(step,loss_value, acc))

		if step % 500 == 0:
			random.shuffle(traindatas)
			trainlabels = list(map(lambda x: x.split('.')[0],traindatas))

		if step % 3000 == 0:
			lr *= 0.9

		if step % 10000 == 0:
			saver.save(sess, save_path + "/model.ckpt-%d" % step)
			print('model saved!')

接下来是我写的一个直观观察训练效果的，新建一个脚本，添加如下代码，然后运行该脚本，将会随机展示4张验证码和你的预测结果，终端还会显示本次预测的精确度。

from __future__ import print_function, division, absolute_import
import tensorflow as tf
import os
import cv2
import matplotlib.pyplot as plt
import random
import numpy as np
from datasplit import use
#from optparse import OptionParser

testnumber = 4 #要更改的话需要改画图部分的代码否则会出错
path = 'Discuz/'
imgs = os.listdir(path)
model_path = 'model4cnn-1fcn/model.ckpt-500000' #读取你训练好的模型
testdatas = random.sample(imgs,testnumber)
testlabels = list(map(lambda x: x.split('.')[0],testdatas))
#testnum = len(testdatas)
#test_ptr = 0

X = tf.placeholder(tf.float32, [None, 30*100])
Y = tf.placeholder(tf.float32, [None,4*63])
keep_prob = tf.placeholder(tf.float32)

def text2vec(text):
	if len(text) > 4:
		raise ValueError('too long captcha')

	vector = np.zeros(4*63)
	def char2pos(c):
		if c == '_':
			k = 62
			return k
		k = ord(c)-48
		if k > 9:
			k = ord(c)-55
			if k > 35:
				k = ord(c) - 61
				if k > 61:
					raise ValueError('No Map')

		return k

	for i, c in enumerate(text):
		idx = i*63 + char2pos(c)
		vector[idx] = 1

	return vector

def vec2text(vec):

    char_pos = vec.nonzero()[0]
    text = []
    for i, c in enumerate(char_pos):
        char_at_pos = i #c/63
        char_idx = c % 63
        if char_idx < 10:
            char_code = char_idx + ord('0')
        elif char_idx < 36:
            char_code = char_idx - 10 + ord('A')
        elif char_idx < 62:
            char_code = char_idx - 36 + ord('a')
        elif char_idx == 62:
            char_code = ord('_')
        else:
            raise ValueError('error')
        text.append(chr(char_code))
    return "".join(text)

batch_x = np.zeros([testnumber,30*100])
batch_y = np.zeros([testnumber, 4*63])

for index, test in enumerate(testdatas):
	img = np.mean(cv2.imread(path + test), -1)
	batch_x[index, :] = img.flatten() /255
for index, label in enumerate(testlabels):
	batch_y[index, :] = text2vec(label)

def conv2d(x, W, b, strides=1):
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)

def max_pool2d(x, k=2):
    x = tf.nn.max_pool(
        x, ksize=[
            1, k, k, 1], strides=[
            1, k, k, 1], padding='SAME')
    return x

weights = {
        'wc1': tf.Variable(0.01*tf.random_normal([3, 3, 1, 32])),
        'wc2': tf.Variable(0.01*tf.random_normal([3, 3, 32, 64])),
        'wc3': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
        'wc4': tf.Variable(0.01*tf.random_normal([3, 3, 64, 64])),
        'wf1': tf.Variable(0.01*tf.random_normal([2 * 7 * 64, 1024])),
        'wf2': tf.Variable(0.01*tf.random_normal([1024, 1024])),
        'wout': tf.Variable(0.01*tf.random_normal([1024, 4*63]))
        }

biases = {
        'bc1': tf.Variable(0.1*tf.random_normal([32])),
        'bc2': tf.Variable(0.1*tf.random_normal([64])),
        'bc3': tf.Variable(0.1*tf.random_normal([64])),
        'bc4': tf.Variable(0.1*tf.random_normal([64])),
        'bf1': tf.Variable(0.1*tf.random_normal([1024])),
        'bf2': tf.Variable(0.1*tf.random_normal([1024])),
        'bout': tf.Variable(0.1*tf.random_normal([4*63]))
    }

def conv_net(x, weights, biases, dropout):
	x = tf.reshape(x, [-1,100,30,1])

	conv1 = conv2d(x, weights['wc1'], biases['bc1'], 1)
	conv1 = max_pool2d(conv1, 2)

	conv2 = conv2d(conv1, weights['wc2'], biases['bc2'], 1)
	conv2 = max_pool2d(conv2, 2)

	conv3 = conv2d(conv2, weights['wc3'], biases['bc3'], 1)
	conv3 = max_pool2d(conv3, 2)

	conv4 = conv2d(conv3, weights['wc4'], biases['bc4'], 1)
	conv4 = max_pool2d(conv4, 2)

	fc1 = tf.reshape(
        conv4, shape=[-1, weights['wf1'].get_shape().as_list()[0]])
	fc1 = tf.matmul(fc1, weights['wf1'])
	fc1 = tf.add(fc1, biases['bf1'])
	fc1 = tf.nn.relu(fc1)

	out = tf.add(tf.matmul(fc1, weights['wout']), biases['bout'])

	return out

output = conv_net(X, weights, biases, keep_prob)

y = tf.reshape(output, [-1,4,63])
y_ = tf.reshape(Y, [-1,4,63])

predict = tf.argmax(y,2)
correct_pred = tf.equal(predict, tf.argmax(y_,2))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
saver = tf.train.Saver()

with tf.Session() as sess:
	saver.restore(sess, model_path)

	pred, acc = sess.run([predict,accuracy], feed_dict ={ X:batch_x, Y:batch_y,keep_prob:1})
	print('accuracy={}'.format(acc))
	for i in range(1,testnumber+1):

		plt.subplot(2,2,i)
		img = cv2.imread(path+testdatas[i-1])
		plt.imshow(img)
		plt.title('number%d' %i)
		plt.xticks([])
		plt.yticks([])
		vect = np.zeros([4*63])

		#print(pred[i-1])
		for ind,j in enumerate(pred[i-1]):
			vect[ind*63+j] = 1

		xlabel = 'True label:{};Pred label:{}'.format(testlabels[i-1], vec2text(vect))
		plt.xlabel(xlabel)

	plt.show()

有任何问题欢迎讨论。