keras入门（三）搭建CNN模型破解网站验证码

项目介绍

在文章CNN大战验证码中，我们利用TensorFlow搭建了简单的CNN模型来破解某个网站的验证码。验证码如下：

网站验证码

在本文中，我们将会用Keras来搭建一个稍微复杂的CNN模型来破解以上的验证码。

数据集

对于验证码图片的处理过程在本文中将不再具体叙述，有兴趣的读者可以参考文章CNN大战验证码。
在这个项目中，我们现在的样本一共是1668个样本，每个样本都是一个字符图片，字符图片的大小为16*20。样本的特征为字符图片的像素，0代表白色，1代表黑色，每个样本为320个特征，取值为0或1，特征变量名称为v1到v320，样本的类别标签即为该字符。整个数据集的部分如下：

data.csv（部分）

CNN模型

利用Keras可以快速方便地搭建CNN模型，本文搭建的CNN模型如下：

将数据集分为训练集和测试集，占比为8：2，该模型训练的代码如下：

# -*- coding: utf-8 -*-
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as pltfrom keras.utils import np_utils, plot_model
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.callbacks import EarlyStopping
from keras.layers import Conv2D, MaxPooling2D# 读取数据
df = pd.read_csv('F://verifycode_data/data.csv')# 标签值
vals = range(31)
keys = ['1','2','3','4','5','6','7','8','9','A','B','C','D','E','F','G','H','J','K','L','N','P','Q','R','S','T','U','V','X','Y','Z']
label_dict = dict(zip(keys, vals))x_data = df[['v'+str(i+1) for i in range(320)]]
y_data = pd.DataFrame({'label':df['label']})
y_data['class'] = y_data['label'].apply(lambda x: label_dict[x])# 将数据分为训练集和测试集
X_train, X_test, Y_train, Y_test = train_test_split(x_data, y_data['class'], test_size=0.3, random_state=42)
x_train = np.array(X_train).reshape((1167, 20, 16, 1))
x_test = np.array(X_test).reshape((501, 20, 16, 1))# 对标签值进行one-hot encoding
n_classes = 31
y_train = np_utils.to_categorical(Y_train, n_classes)
y_val = np_utils.to_categorical(Y_test, n_classes)input_shape = x_train[0].shape# CNN模型
model = Sequential()# 卷积层和池化层
model.add(Conv2D(32, kernel_size=(3, 3), input_shape=input_shape, padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))# Dropout层
model.add(Dropout(0.25))model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))model.add(Dropout(0.25))model.add(Conv2D(128, kernel_size=(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(128, kernel_size=(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))model.add(Dropout(0.25))model.add(Flatten())# 全连接层
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(128, activation='relu'))
model.add(Dense(n_classes, activation='softmax'))model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])# plot model
plot_model(model, to_file=r'./model.png', show_shapes=True)# 模型训练
callbacks = [EarlyStopping(monitor='val_acc', patience=5, verbose=1)]
batch_size = 64
n_epochs = 100
history = model.fit(x_train, y_train, batch_size=batch_size, epochs=n_epochs, \verbose=1, validation_data=(x_test, y_val), callbacks=callbacks)mp = 'F://verifycode_data/verifycode_Keras.h5'
model.save(mp)# 绘制验证集上的准确率曲线
val_acc = history.history['val_acc']
plt.plot(range(len(val_acc)), val_acc, label='CNN model')
plt.title('Validation accuracy on verifycode dataset')
plt.xlabel('epochs')
plt.ylabel('accuracy')
plt.legend()
plt.show()

在上述代码中，我们训练模型的时候采用了early stopping技巧。early stopping是用于提前停止训练的callbacks。具体地，可以达到当训练集上的loss不在减小（即减小的程度小于某个阈值）的时候停止继续训练。

模型训练

运行上述模型训练代码，输出的结果如下：

......(忽略之前的输出)
Epoch 22/10064/1167 [>.............................] - ETA: 3s - loss: 0.0399 - acc: 1.0000128/1167 [==>...........................] - ETA: 3s - loss: 0.1195 - acc: 0.9844192/1167 [===>..........................] - ETA: 2s - loss: 0.1085 - acc: 0.9792256/1167 [=====>........................] - ETA: 2s - loss: 0.1132 - acc: 0.9727320/1167 [=======>......................] - ETA: 2s - loss: 0.1045 - acc: 0.9750384/1167 [========>.....................] - ETA: 2s - loss: 0.1006 - acc: 0.9740448/1167 [==========>...................] - ETA: 2s - loss: 0.1522 - acc: 0.9643512/1167 [============>.................] - ETA: 1s - loss: 0.1450 - acc: 0.9648576/1167 [=============>................] - ETA: 1s - loss: 0.1368 - acc: 0.9653640/1167 [===============>..............] - ETA: 1s - loss: 0.1353 - acc: 0.9641704/1167 [=================>............] - ETA: 1s - loss: 0.1280 - acc: 0.9659768/1167 [==================>...........] - ETA: 1s - loss: 0.1243 - acc: 0.9674832/1167 [====================>.........] - ETA: 0s - loss: 0.1577 - acc: 0.9639896/1167 [======================>.......] - ETA: 0s - loss: 0.1488 - acc: 0.9665960/1167 [=======================>......] - ETA: 0s - loss: 0.1488 - acc: 0.9656
1024/1167 [=========================>....] - ETA: 0s - loss: 0.1427 - acc: 0.9668
1088/1167 [==========================>...] - ETA: 0s - loss: 0.1435 - acc: 0.9669
1152/1167 [============================>.] - ETA: 0s - loss: 0.1383 - acc: 0.9688
1167/1167 [==============================] - 4s 3ms/step - loss: 0.1380 - acc: 0.9683 - val_loss: 0.0835 - val_acc: 0.9760
Epoch 00022: early stopping

可以看到，一共训练了21次，最近一次的训练后，在测试集上的准确率为96.83%。在测试集的准确率曲线如下图：

测试集上的准确率曲线

模型预测

模型训练完后，我们对新的验证码进行预测。新的100张验证码如下图：

新的验证码（部分）

使用训练好的CNN模型，对这些新的验证码进行预测，预测的Python代码如下：

# -*- coding: utf-8 -*-import os
import cv2
import numpy as npdef split_picture(imagepath):# 以灰度模式读取图片gray = cv2.imread(imagepath, 0)# 将图片的边缘变为白色height, width = gray.shapefor i in range(width):gray[0, i] = 255gray[height-1, i] = 255for j in range(height):gray[j, 0] = 255gray[j, width-1] = 255# 中值滤波blur = cv2.medianBlur(gray, 3) #模板大小3*3# 二值化ret,thresh1 = cv2.threshold(blur, 200, 255, cv2.THRESH_BINARY)# 提取单个字符chars_list = []image, contours, hierarchy = cv2.findContours(thresh1, 2, 2)for cnt in contours:# 最小的外接矩形x, y, w, h = cv2.boundingRect(cnt)if x != 0 and y != 0 and w*h >= 100:chars_list.append((x,y,w,h))sorted_chars_list = sorted(chars_list, key=lambda x:x[0])for i,item in enumerate(sorted_chars_list):x, y, w, h = itemcv2.imwrite('F://test_verifycode/chars/%d.jpg'%(i+1), thresh1[y:y+h, x:x+w])def remove_edge_picture(imagepath):image = cv2.imread(imagepath, 0)height, width = image.shapecorner_list = [image[0,0] < 127,image[height-1, 0] < 127,image[0, width-1]<127,image[ height-1, width-1] < 127]if sum(corner_list) >= 3:os.remove(imagepath)def resplit_with_parts(imagepath, parts):image = cv2.imread(imagepath, 0)os.remove(imagepath)height, width = image.shapefile_name = imagepath.split('/')[-1].split(r'.')[0]# 将图片重新分裂成parts部分step = width//parts     # 步长start = 0             # 起始位置for i in range(parts):cv2.imwrite('F://test_verifycode/chars/%s.jpg'%(file_name+'-'+str(i)), \image[:, start:start+step])start += stepdef resplit(imagepath):image = cv2.imread(imagepath, 0)height, width = image.shapeif width >= 64:resplit_with_parts(imagepath, 4)elif width >= 48:resplit_with_parts(imagepath, 3)elif width >= 26:resplit_with_parts(imagepath, 2)# rename and convert to 16*20 size
def convert(dir, file):imagepath = dir+'/'+file# 读取图片image = cv2.imread(imagepath, 0)# 二值化ret, thresh = cv2.threshold(image, 127, 255, cv2.THRESH_BINARY)img = cv2.resize(thresh, (16, 20), interpolation=cv2.INTER_AREA)# 保存图片cv2.imwrite('%s/%s' % (dir, file), img)# 读取图片的数据，并转化为0-1值
def Read_Data(dir, file):imagepath = dir+'/'+file# 读取图片image = cv2.imread(imagepath, 0)# 二值化ret, thresh = cv2.threshold(image, 127, 255, cv2.THRESH_BINARY)# 显示图片bin_values = [1 if pixel==255 else 0 for pixel in thresh.ravel()]return bin_valuesdef predict(VerifyCodePath):dir = 'F://test_verifycode/chars'files = os.listdir(dir)# 清空原有的文件if files:for file in files:os.remove(dir + '/' + file)split_picture(VerifyCodePath)files = os.listdir(dir)if not files:print('查看的文件夹为空！')else:# 去除噪声图片for file in files:remove_edge_picture(dir + '/' + file)# 对黏连图片进行重分割for file in os.listdir(dir):resplit(dir + '/' + file)# 将图片统一调整至16*20大小for file in os.listdir(dir):convert(dir, file)# 图片中的字符代表的向量files = sorted(os.listdir(dir), key=lambda x: x[0])table = np.array([Read_Data(dir, file) for file in files]).reshape(-1,20,16,1)# 模型保存地址mp = 'F://verifycode_data/verifycode_Keras.h5'# 载入模型from keras.models import load_modelcnn = load_model(mp)# 模型预测y_pred = cnn.predict(table)predictions = np.argmax(y_pred, axis=1)# 标签字典keys = range(31)vals = ['1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'N','P', 'Q', 'R', 'S', 'T', 'U', 'V', 'X', 'Y', 'Z']label_dict = dict(zip(keys, vals))return ''.join([label_dict[pred] for pred in predictions])def main():dir = 'F://VerifyCode/'correct = 0for i, file in enumerate(os.listdir(dir)):true_label = file.split('.')[0]VerifyCodePath = dir+filepred = predict(VerifyCodePath)if true_label == pred:correct += 1print(i+1, (true_label, pred), true_label == pred, correct)total = len(os.listdir(dir))print('\n总共图片：%d张\n识别正确：%d张\n识别准确率:%.2f%%.'\%(total, correct, correct*100/total))main()

以下是该CNN模型的预测结果：

Using TensorFlow backend.
2018-10-25 15:13:50.390130: I C:\tf_jenkins\workspace\rel-win\M\windows\PY\35\tensorflow\core\platform\cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
1 ('ZK6N', 'ZK6N') True 1
2 ('4JPX', '4JPX') True 2
3 ('5GP5', '5GP5') True 3
4 ('5RQ8', '5RQ8') True 4
5 ('5TQP', '5TQP') True 5
6 ('7S62', '7S62') True 6
7 ('8R2Z', '8R2Z') True 7
8 ('8RFV', '8RFV') True 8
9 ('9BBT', '9BBT') True 9
10 ('9LNE', '9LNE') True 10
11 ('67UH', '67UH') True 11
12 ('74UK', '74UK') True 12
13 ('A5T2', 'A5T2') True 13
14 ('AHYV', 'AHYV') True 14
15 ('ASEY', 'ASEY') True 15
16 ('B371', 'B371') True 16
17 ('CCQL', 'CCQL') True 17
18 ('CFD5', 'GFD5') False 17
19 ('CJLJ', 'CJLJ') True 18
20 ('D4QV', 'D4QV') True 19
21 ('DFQ8', 'DFQ8') True 20
22 ('DP18', 'DP18') True 21
23 ('E3HC', 'E3HC') True 22
24 ('E8VB', 'E8VB') True 23
25 ('DE1U', 'DE1U') True 24
26 ('FK1R', 'FK1R') True 25
27 ('FK91', 'FK91') True 26
28 ('FSKP', 'FSKP') True 27
29 ('FVZP', 'FVZP') True 28
30 ('GC6H', 'GC6H') True 29
31 ('GH62', 'GH62') True 30
32 ('H9FQ', 'H9FQ') True 31
33 ('H67Q', 'H67Q') True 32
34 ('HEKC', 'HEKC') True 33
35 ('HV2B', 'HV2B') True 34
36 ('J65Z', 'J65Z') True 35
37 ('JZCX', 'JZCX') True 36
38 ('KH5D', 'KH5D') True 37
39 ('KXD2', 'KXD2') True 38
40 ('1GDH', '1GDH') True 39
41 ('LCL3', 'LCL3') True 40
42 ('LNZR', 'LNZR') True 41
43 ('LZU5', 'LZU5') True 42
44 ('N5AK', 'N5AK') True 43
45 ('N5Q3', 'N5Q3') True 44
46 ('N96Z', 'N96Z') True 45
47 ('NCDG', 'NCDG') True 46
48 ('NELS', 'NELS') True 47
49 ('P96U', 'P96U') True 48
50 ('PD42', 'PD42') True 49
51 ('PECG', 'PEQG') False 49
52 ('PPZF', 'PPZF') True 50
53 ('PUUL', 'PUUL') True 51
54 ('Q2DN', 'D2DN') False 51
55 ('QCQ9', 'QCQ9') True 52
56 ('QDB1', 'QDBJ') False 52
57 ('QZUD', 'QZUD') True 53
58 ('R3T5', 'R3T5') True 54
59 ('S1YT', 'S1YT') True 55
60 ('SP7L', 'SP7L') True 56
61 ('SR2K', 'SR2K') True 57
62 ('SUP5', 'SVP5') False 57
63 ('T2SP', 'T2SP') True 58
64 ('U6V9', 'U6V9') True 59
65 ('UC9P', 'UC9P') True 60
66 ('UFYD', 'UFYD') True 61
67 ('V9NJ', 'V9NH') False 61
68 ('V35X', 'V35X') True 62
69 ('V98F', 'V98F') True 63
70 ('VD28', 'VD28') True 64
71 ('YGHE', 'YGHE') True 65
72 ('YNKD', 'YNKD') True 66
73 ('YVXV', 'YVXV') True 67
74 ('ZFBS', 'ZFBS') True 68
75 ('ET6X', 'ET6X') True 69
76 ('TKVC', 'TKVC') True 70
77 ('2UCU', '2UCU') True 71
78 ('HNBK', 'HNBK') True 72
79 ('X8FD', 'X8FD') True 73
80 ('ZGNX', 'ZGNX') True 74
81 ('LQCU', 'LQCU') True 75
82 ('JNZY', 'JNZVY') False 75
83 ('RX34', 'RX34') True 76
84 ('811E', '811E') True 77
85 ('ETDX', 'ETDX') True 78
86 ('4CPR', '4CPR') True 79
87 ('FE91', 'FE91') True 80
88 ('B7XH', 'B7XH') True 81
89 ('1RUA', '1RUA') True 82
90 ('UBCX', 'UBCX') True 83
91 ('KVT5', 'KVT5') True 84
92 ('HZ3A', 'HZ3A') True 85
93 ('3XLR', '3XLR') True 86
94 ('VC7T', 'VC7T') True 87
95 ('7PG1', '7PQ1') False 87
96 ('4F21', '4F21') True 88
97 ('3HLJ', '3HLJ') True 89
98 ('1KT7', '1KT7') True 90
99 ('1RHE', '1RHE') True 91
100 ('1TTA', '1TTA') True 92总共图片：100张
识别正确：92张
识别准确率:92.00%.

可以看到，该训练后的CNN模型，其预测新验证的准确率在90%以上。

总结

在文章CNN大战验证码中，笔者使用TensorFlow搭建了CNN模型，代码较长，训练时间在两个小时以上，而使用Keras搭建该模型，代码简洁，且使用early stopping技巧后能缩短训练时间，同时保证模型的准确率，由此可见Keras的优势所在。
该项目已开源，Github地址为：https://github.com/percent4/C...。

注意：本人现已开通微信公众号： Python爬虫与算法（微信号为：easy_web_scrape），欢迎大家关注哦~~