專案介紹

  在文章CNN大戰驗證碼中，我們利用TensorFlow搭建了簡單的CNN模型來破解某個網站的驗證碼。驗證碼如下：

在本文中，我們將會用Keras來搭建一個稍微複雜的CNN模型來破解以上的驗證碼。

資料集

  對於驗證碼圖片的處理過程在本文中將不再具體敘述，有興趣的讀者可以參考文章CNN大戰驗證碼。
  在這個專案中，我們現在的樣本一共是1668個樣本，每個樣本都是一個字元圖片，字元圖片的大小為16*20。樣本的特徵為字元圖片的畫素，0代表白色，1代表黑色，每個樣本為320個特徵，取值為0或1，特徵變數名稱為v1到v320，樣本的類別標籤即為該字元。整個資料集的部分如下：

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：

將資料集分為訓練集和測試集，佔比為8：2，該模型訓練的程式碼如下：

# -*- coding: utf-8 -*-
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt

from 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

# 讀取資料
df = pd.read_csv('F://verifycode_data/data.csv')

# 標籤值
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])

# 將資料分為訓練集和測試集
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))

# 對標籤值進行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()

# 卷積層和池化層
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層
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())

# 全連線層
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)

# 模型訓練
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)

# 繪製驗證集上的準確率曲線
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()

在上述程式碼中，我們訓練模型的時候採用了early stopping技巧。early stopping是用於提前停止訓練的callbacks。具體地，可以達到當訓練集上的loss不在減小（即減小的程度小於某個閾值）的時候停止繼續訓練。

模型訓練

  執行上述模型訓練程式碼，輸出的結果如下：

......(忽略之前的輸出)
Epoch 22/100

  64/1167 [>.............................] - ETA: 3s - loss: 0.0399 - acc: 1.0000
 128/1167 [==>...........................] - ETA: 3s - loss: 0.1195 - acc: 0.9844
 192/1167 [===>..........................] - ETA: 2s - loss: 0.1085 - acc: 0.9792
 256/1167 [=====>........................] - ETA: 2s - loss: 0.1132 - acc: 0.9727
 320/1167 [=======>......................] - ETA: 2s - loss: 0.1045 - acc: 0.9750
 384/1167 [========>.....................] - ETA: 2s - loss: 0.1006 - acc: 0.9740
 448/1167 [==========>...................] - ETA: 2s - loss: 0.1522 - acc: 0.9643
 512/1167 [============>.................] - ETA: 1s - loss: 0.1450 - acc: 0.9648
 576/1167 [=============>................] - ETA: 1s - loss: 0.1368 - acc: 0.9653
 640/1167 [===============>..............] - ETA: 1s - loss: 0.1353 - acc: 0.9641
 704/1167 [=================>............] - ETA: 1s - loss: 0.1280 - acc: 0.9659
 768/1167 [==================>...........] - ETA: 1s - loss: 0.1243 - acc: 0.9674
 832/1167 [====================>.........] - ETA: 0s - loss: 0.1577 - acc: 0.9639
 896/1167 [======================>.......] - ETA: 0s - loss: 0.1488 - acc: 0.9665
 960/1167 [=======================>......] - 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，一共訓練了21次，最近一次的訓練後，在測試集上的準確率為96.83%。在測試集的準確率曲線如下圖：

模型預測

  模型訓練完後，我們對新的驗證碼進行預測。新的100張驗證碼如下圖：

  使用訓練好的CNN模型，對這些新的驗證碼進行預測，預測的Python程式碼如下：

# -*- coding: utf-8 -*-

import os
import cv2
import numpy as np

def split_picture(imagepath):

    # 以灰度模式讀取圖片
    gray = cv2.imread(imagepath, 0)

    # 將圖片的邊緣變為白色
    height, width = gray.shape
    for i in range(width):
        gray[0, i] = 255
        gray[height-1, i] = 255
    for j in range(height):
        gray[j, 0] = 255
        gray[j, width-1] = 255

    # 中值濾波
    blur = cv2.medianBlur(gray, 3) #模板大小3*3

    # 二值化
    ret,thresh1 = cv2.threshold(blur, 200, 255, cv2.THRESH_BINARY)

    # 提取單個字元
    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 = item
        cv2.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.shape
    corner_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.shape

    file_name = imagepath.split('/')[-1].split(r'.')[0]
    # 將圖片重新分裂成parts部分
    step = width//parts     # 步長
    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 += step

def resplit(imagepath):

    image = cv2.imread(imagepath, 0)
    height, width = image.shape

    if 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
    # 讀取圖片
    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)
    # 儲存圖片
    cv2.imwrite('%s/%s' % (dir, file), img)

# 讀取圖片的資料，並轉化為0-1值
def Read_Data(dir, file):

    imagepath = dir+'/'+file
    # 讀取圖片
    image = cv2.imread(imagepath, 0)
    # 二值化
    ret, thresh = cv2.threshold(image, 127, 255, cv2.THRESH_BINARY)
    # 顯示圖片
    bin_values = [1 if pixel==255 else 0 for pixel in thresh.ravel()]

    return bin_values

def predict(VerifyCodePath):

    dir = 'F://test_verifycode/chars'
    files = os.listdir(dir)

    # 清空原有的檔案
    if files:
        for file in files:
            os.remove(dir + '/' + file)

    split_picture(VerifyCodePath)

    files = os.listdir(dir)
    if not files:
        print('檢視的資料夾為空！')
    else:

        # 去除噪聲圖片
        for file in files:
            remove_edge_picture(dir + '/' + file)

        # 對黏連圖片進行重分割
        for file in os.listdir(dir):
            resplit(dir + '/' + file)

        # 將圖片統一調整至16*20大小
        for file in os.listdir(dir):
            convert(dir, file)

        # 圖片中的字元代表的向量
        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)

        # 模型儲存地址
        mp = 'F://verifycode_data/verifycode_Keras.h5'
        # 載入模型
        from keras.models import load_model
        cnn = load_model(mp)
        # 模型預測
        y_pred = cnn.predict(table)
        predictions = np.argmax(y_pred, axis=1)

        # 標籤字典
        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 = 0
    for i, file in enumerate(os.listdir(dir)):
        true_label = file.split('.')[0]
        VerifyCodePath = dir+file
        pred = predict(VerifyCodePath)

        if true_label == pred:
            correct += 1
        print(i+1, (true_label, pred), true_label == pred, correct)

    total = len(os.listdir(dir))
    print('\n總共圖片：%d張\n識別正確：%d張\n識別準確率:%.2f%%.'\
          %(total, correct, correct*100/total))

main()

以下是該CNN模型的預測結果：

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

總共圖片：100張
識別正確：92張
識別準確率:92.00%.

可以看到，該訓練後的CNN模型，其預測新驗證的準確率在90%以上。

總結