深度学习卷积神经网络识别验证码

一、引言

验证码识别，本身使用来判断访问网站的用户是不是一个真人，但是随着人工智能的发展，尤其是深度卷积神经网络的发展，使得验证码识别机器还是人的底线再被层层攻破，本文所使用的captcha 可以生成语音和图片验证码，验证码是由数字、大写字母、小写字母组成，可以自动生成验证码。

最好掌握的预备知识：

• 有python脚本编程基础；
• 了解图像处理、计算机视觉理论基础；
• 对Opencv有一定的认识；
• 对图像处理有操作经验；

目的：通过Python语言与其各种资源库如numpy,pillow,tensorflow,keras等来实现。将效果实现出来。了解如何使用catptcha自动生成验证码数据，同时使用tensorflow撰写网络，来完成验证码训练，将使用高级框架的keras作为手写字母的网络主体。

二、导入必要的库

我们需要导入一个叫 captcha 的库来生成验证码。

我们生成验证码的字符由数字和大写字母组成。

导入成功后，我们使用jupyter notebook编写代码，使用py3.6.9，TensorFlow2.0+

from captcha.image import ImageCaptcha
import matplotlib.pyplot as plt
import numpy as np
import random
import tensorflow as tf%matplotlib inline
%config InlineBackend.figure_format = 'retina'#输出所有的ascii，用来生成验证码
import string
characters = string.digits + string.ascii_uppercase
print(characters)#验证码的长、宽、字符数、类别数
width, height, n_len, n_class = 128, 64, 4, len(characters)

输出结果：

0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ

三、防止 tensorflow 占用所有显存

#下面就是实现按需分配的代码
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:try:#设置仅在需要时申请显存空间for gpu in gpus:tf.config.experimental.set_memory_growth(gpu, True)logical_gpus = tf.config.experimental.list_logical_devices('GPU')print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")except RuntimeError as e:print(e)

四、定义数据生成器并测试

from tensorflow.keras.utils import Sequenceclass CaptchaSequence(Sequence):#这里传入的参数与上面的导入模块常量相对应def __init__(self, characters, batch_size, steps, n_len=4, width=128, height=64):self.characters = charactersself.batch_size = batch_sizeself.steps = stepsself.n_len = n_lenself.width = widthself.height = heightself.n_class = len(characters)self.generator = ImageCaptcha(width=width, height=height)def __len__(self):return self.stepsdef __getitem__(self, idx):X = np.zeros((self.batch_size, self.height, self.width, 3), dtype=np.float32)y = [np.zeros((self.batch_size, self.n_class), dtype=np.uint8) for i in range(self.n_len)]for i in range(self.batch_size):random_str = ''.join([random.choice(self.characters) for j in range(self.n_len)])X[i] = np.array(self.generator.generate_image(random_str)) / 255.0for j, ch in enumerate(random_str):y[j][i, :] = 0y[j][i, self.characters.find(ch)] = 1return X, y

测试代码：

def decode(y):y = np.argmax(np.array(y), axis=2)[:,0]return ''.join([characters[x] for x in y])
#注意每次运行输出结果可能不同，因为我们每次随机选择一个字符
data = CaptchaSequence(characters, batch_size=1, steps=1)  
X, y = data[0]
plt.imshow(X[0])
plt.title(decode(y))

结果如下：

注意：生成的验证码中字可能会出现不清晰的情况

例如这样：

总结原因可能有以下几点：

1. 图片尺寸过小：在代码中，如果设置了验证码图片的宽度和高度为64和32，这可能导致生成的验证码图片较小，从而导致字体较小且不清晰。可以尝试增大图片尺寸，例如将宽度设置为128，高度设置为64。
2. 图片缩放导致失真：在生成图片时，使用了np.array(self.generator.generate_image(random_str)) / 255.0 将图像转换为numpy数组并进行了归一化处理。如果在这个过程中对图像进行了缩放操作，可能会导致图像失真，从而影响字体的清晰度。可以尝试去除归一化处理，或者调整归一化的方式，确保图像不失真。
3. 字体类型和大小：验证码生成器ImageCaptcha使用默认的字体类型和大小生成验证码。如果默认字体类型和大小不适合生成清晰的验证码，可以尝试更换字体类型并调整字体大小。可以参考ImageCaptcha的文档，了解如何更改字体类型和大小。
4. 图像参数调整：除了上述因素外，还可以尝试调整一些图像生成的参数，例如干扰线的宽度、噪点的数量等，以提高验证码的清晰度。

五、定义网络结构

train_data = CaptchaSequence(characters, batch_size=512, steps=1000)
valid_data = CaptchaSequence(characters, batch_size=128, steps=100)
x_train, y_train = train_data[0]
x_val, y_val = valid_data[0]

from tensorflow.keras.models import *
from tensorflow.keras.layers import *input_tensor = Input((height, width, 3))
x = input_tensor
for num_cnn,num_kernel in enumerate([32,16,8]):for j in range(num_cnn):x = Conv2D(num_kernel, kernel_size=(3,3), padding='same', kernel_initializer='he_uniform')(x)x = Conv2D(num_kernel, kernel_size=(3,3), padding='same')(x)x = BatchNormalization()(x)x = Activation('relu')(x)x = MaxPooling2D(pool_size=(2,2))(x)x = Flatten()(x)
x = [Dense(n_class, activation='softmax',name='c%d'%(i+1))(x) for i in range(n_len)]
model = Model(inputs=input_tensor, outputs=x)
model.summary()

网络结构总结如下：

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 64, 128, 3)] 0                                            
__________________________________________________________________________________________________
conv2d (Conv2D)                 (None, 64, 128, 16)  448         input_1[0][0]                    
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 64, 128, 16)  2320        conv2d[0][0]                     
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 64, 128, 16)  64          conv2d_1[0][0]                   
__________________________________________________________________________________________________
activation (Activation)         (None, 64, 128, 16)  0           batch_normalization[0][0]        
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D)    (None, 32, 64, 16)   0           activation[0][0]                 
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 32, 64, 8)    1160        max_pooling2d[0][0]              
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 32, 64, 8)    584         conv2d_2[0][0]                   
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 32, 64, 8)    32          conv2d_3[0][0]                   
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 32, 64, 8)    0           batch_normalization_1[0][0]      
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D)  (None, 16, 32, 8)    0           activation_1[0][0]               
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 16, 32, 8)    584         max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 16, 32, 8)    584         conv2d_4[0][0]                   
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 16, 32, 8)    32          conv2d_5[0][0]                   
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 16, 32, 8)    0           batch_normalization_2[0][0]      
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D)  (None, 8, 16, 8)     0           activation_2[0][0]               
__________________________________________________________________________________________________
flatten (Flatten)               (None, 1024)         0           max_pooling2d_2[0][0]            
__________________________________________________________________________________________________
c1 (Dense)                      (None, 36)           36900       flatten[0][0]                    
__________________________________________________________________________________________________
c2 (Dense)                      (None, 36)           36900       flatten[0][0]                    
__________________________________________________________________________________________________
c3 (Dense)                      (None, 36)           36900       flatten[0][0]                    
__________________________________________________________________________________________________
c4 (Dense)                      (None, 36)           36900       flatten[0][0]                    
==================================================================================================
Total params: 153,408
Trainable params: 153,344
Non-trainable params: 64

六、训练模型

开始训练

from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
from tensorflow.keras.optimizers import Adamcallbacks = [EarlyStopping(patience=3), ModelCheckpoint('cnn_best.h5', save_best_only=True)]model.compile(loss='categorical_crossentropy',optimizer=Adam(1e-3), metrics=['accuracy'])
history = model.fit(x_train, y_train, epochs=100, validation_data=(x_val,y_val),callbacks=callbacks, verbose=1)

载入最好的模型继续训练一会

model.load_weights('cnn_best.h5')callbacks = [EarlyStopping(patience=3),ModelCheckpoint('cnn_best.h5', save_best_only=True)]
#学习率降低，
model.compile(loss='categorical_crossentropy',optimizer=Adam(1e-4), metrics=['accuracy'])
model.fit(x_train, y_train, epochs=100, validation_data=(x_val,y_val),callbacks=callbacks, verbose=1)

保存最好的模型

model.save('../cnn_best.h5')

七、测试模型

X, y = data[0]
y_pred = model.predict(X)
plt.title('real: %s pred:%s'%(decode(y), decode(y_pred)))
plt.imshow(X[0], cmap='gray')
plt.axis('off')

可见测试模型效果并不理想，因为我们的训练迭代次数少，网络结构简单，不过我们可以增加网络层数，调节网络参数然后重新训练一个更好的模型。