Tensorflow是目前最流行的深度學習框架，我們可以用它來搭建自己的卷積神經網路並訓練自己的分類器，本文介紹怎樣使用Tensorflow構建自己的CNN，怎樣訓練用於簡單的驗證碼識別的分類器。本文假設你已經安裝好了Tensorflow，瞭解過CNN的一些知識。

下面將分步介紹怎樣獲得訓練資料，怎樣使用tensorflow構建卷積神經網路，怎樣訓練，以及怎樣測試訓練出來的分類器

1. 準備訓練樣本

        使用Python的庫captcha來生成我們需要的訓練樣本，程式碼如下：

import sys

import os
import shutil
import random
import time
#captcha是用於生成驗證碼圖片的庫，可以 pip install captcha 來安裝它
from captcha.image import ImageCaptcha

#用於生成驗證碼的字符集
CHAR_SET = ['0','1','2','3','4','5','6','7','8','9']
#字符集的長度
CHAR_SET_LEN = 10
#驗證碼的長度，每個驗證碼由4個數字組成
CAPTCHA_LEN = 4

#驗證碼圖片的存放路徑
CAPTCHA_IMAGE_PATH = 'E:/Tensorflow/captcha/images/'
#用於模型測試的驗證碼圖片的存放路徑，它裡面的驗證碼圖片作為測試集
TEST_IMAGE_PATH = 'E:/Tensorflow/captcha/test/'
#用於模型測試的驗證碼圖片的個數，從生成的驗證碼圖片中取出來放入測試集中
TEST_IMAGE_NUMBER = 50

#生成驗證碼圖片，4位的十進位制數字可以有10000種驗證碼
def generate_captcha_image(charSet = CHAR_SET, charSetLen=CHAR_SET_LEN, captchaImgPath=CAPTCHA_IMAGE_PATH):   
    k  = 0
    total = 1
    for i in range(CAPTCHA_LEN):
        total *= charSetLen
        
    for i in range(charSetLen):
        for j in range(charSetLen):
            for m in range(charSetLen):
                for n in range(charSetLen):
                    captcha_text = charSet[i] + charSet[j] + charSet[m] + charSet[n]
                    image = ImageCaptcha()
                    image.write(captcha_text, captchaImgPath + captcha_text + '.jpg')
                    k += 1
                    sys.stdout.write("\rCreating %d/%d" % (k, total))
                    sys.stdout.flush()
                    
#從驗證碼的圖片集中取出一部分作為測試集，這些圖片不參加訓練，只用於模型的測試                    
def prepare_test_set():
    fileNameList = []    
    for filePath in os.listdir(CAPTCHA_IMAGE_PATH):
        captcha_name = filePath.split('/')[-1]
        fileNameList.append(captcha_name)
    random.seed(time.time())
    random.shuffle(fileNameList) 
    for i in range(TEST_IMAGE_NUMBER):
        name = fileNameList[i]
        shutil.move(CAPTCHA_IMAGE_PATH + name, TEST_IMAGE_PATH + name)
                        
if __name__ == '__main__':
    generate_captcha_image(CHAR_SET, CHAR_SET_LEN, CAPTCHA_IMAGE_PATH)
    prepare_test_set()
    sys.stdout.write("\nFinished")
    sys.stdout.flush()  
 

執行上面的程式碼，可以生成驗證碼圖片，

生成的驗證碼圖片如下圖所示：

2. 構建CNN，訓練分類器

     程式碼如下：

import tensorflow as tf
import numpy as np
from PIL import Image
import os
import random
import time

#驗證碼圖片的存放路徑
CAPTCHA_IMAGE_PATH = 'E:/Tensorflow/captcha/images/'
#驗證碼圖片的寬度
CAPTCHA_IMAGE_WIDHT = 160
#驗證碼圖片的高度
CAPTCHA_IMAGE_HEIGHT = 60

CHAR_SET_LEN = 10
CAPTCHA_LEN = 4

#60%的驗證碼圖片放入訓練集中
TRAIN_IMAGE_PERCENT = 0.6
#訓練集，用於訓練的驗證碼圖片的檔名
TRAINING_IMAGE_NAME = []
#驗證集，用於模型驗證的驗證碼圖片的檔名
 

VALIDATION_IMAGE_NAME = []

#存放訓練好的模型的路徑
MODEL_SAVE_PATH = 'E:/Tensorflow/captcha/models/'

def get_image_file_name(imgPath=CAPTCHA_IMAGE_PATH):
    fileName = []
    total = 0
    for filePath in os.listdir(imgPath):
        captcha_name = filePath.split('/')[-1]
        fileName.append(captcha_name)
        total += 1
    return fileName, total
    
#將驗證碼轉換為訓練時用的標籤向量，維數是 40   
#例如，如果驗證碼是 ‘0296’ ，則對應的標籤是
# [1 0 0 0 0 0 0 0 0 0
#  0 0 1 0 0 0 0 0 0 0
#  0 0 0 0 0 0 0 0 0 1
#  0 0 0 0 0 0 1 0 0 0]
def name2label(name):
    label = np.zeros(CAPTCHA_LEN * CHAR_SET_LEN)
    for i, c in enumerate(name):
        idx = i*CHAR_SET_LEN + ord(c) - ord('0')
        label[idx] = 1
    return label
    
#取得驗證碼圖片的資料以及它的標籤        
def get_data_and_label(fileName, filePath=CAPTCHA_IMAGE_PATH):
    pathName = os.path.join(filePath, fileName)
    img = Image.open(pathName)
    #轉為灰度圖
    img = img.convert("L")       
    image_array = np.array(img)    
    image_data = image_array.flatten()/255
    image_label = name2label(fileName[0:CAPTCHA_LEN])
    return image_data, image_label
    
#生成一個訓練batch    
def get_next_batch(batchSize=32, trainOrTest='train', step=0):
    batch_data = np.zeros([batchSize, CAPTCHA_IMAGE_WIDHT*CAPTCHA_IMAGE_HEIGHT])
    batch_label = np.zeros([batchSize, CAPTCHA_LEN * CHAR_SET_LEN])
    fileNameList = TRAINING_IMAGE_NAME
    if trainOrTest == 'validate':        
        fileNameList = VALIDATION_IMAGE_NAME
        
    totalNumber = len(fileNameList) 
    indexStart = step*batchSize    
    for i in range(batchSize):
        index = (i + indexStart) % totalNumber
        name = fileNameList[index]        
        img_data, img_label = get_data_and_label(name)
        batch_data[i, : ] = img_data
        batch_label[i, : ] = img_label  

    return batch_data, batch_label
    
#構建卷積神經網路並訓練
def train_data_with_CNN():
    #初始化權值
    def weight_variable(shape, name='weight'):
        init = tf.truncated_normal(shape, stddev=0.1)
        var = tf.Variable(initial_value=init, name=name)
        return var
    #初始化偏置    
    def bias_variable(shape, name='bias'):
        init = tf.constant(0.1, shape=shape)
        var = tf.Variable(init, name=name)
        return var
    #卷積    
    def conv2d(x, W, name='conv2d'):
        return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding='SAME', name=name)
    #池化 
    def max_pool_2X2(x, name='maxpool'):
        return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME', name=name)     
   
    #輸入層
    #請注意 X 的 name，在測試model時會用到它
    X = tf.placeholder(tf.float32, [None, CAPTCHA_IMAGE_WIDHT * CAPTCHA_IMAGE_HEIGHT], name='data-input')
    Y = tf.placeholder(tf.float32, [None, CAPTCHA_LEN * CHAR_SET_LEN], name='label-input')    
    x_input = tf.reshape(X, [-1, CAPTCHA_IMAGE_HEIGHT, CAPTCHA_IMAGE_WIDHT, 1], name='x-input')
    #dropout,防止過擬合
    #請注意 keep_prob 的 name，在測試model時會用到它
    keep_prob = tf.placeholder(tf.float32, name='keep-prob')
    #第一層卷積
    W_conv1 = weight_variable([5,5,1,32], 'W_conv1')
    B_conv1 = bias_variable([32], 'B_conv1')
    conv1 = tf.nn.relu(conv2d(x_input, W_conv1, 'conv1') + B_conv1)
    conv1 = max_pool_2X2(conv1, 'conv1-pool')
    conv1 = tf.nn.dropout(conv1, keep_prob)
    #第二層卷積
    W_conv2 = weight_variable([5,5,32,64], 'W_conv2')
    B_conv2 = bias_variable([64], 'B_conv2')
    conv2 = tf.nn.relu(conv2d(conv1, W_conv2,'conv2') + B_conv2)
    conv2 = max_pool_2X2(conv2, 'conv2-pool')
    conv2 = tf.nn.dropout(conv2, keep_prob)
    #第三層卷積
    W_conv3 = weight_variable([5,5,64,64], 'W_conv3')
    B_conv3 = bias_variable([64], 'B_conv3')
    conv3 = tf.nn.relu(conv2d(conv2, W_conv3, 'conv3') + B_conv3)
    conv3 = max_pool_2X2(conv3, 'conv3-pool')
    conv3 = tf.nn.dropout(conv3, keep_prob)
    #全連結層
    #每次池化後，圖片的寬度和高度均縮小為原來的一半，進過上面的三次池化，寬度和高度均縮小8倍
    W_fc1 = weight_variable([20*8*64, 1024], 'W_fc1')
    B_fc1 = bias_variable([1024], 'B_fc1')
    fc1 = tf.reshape(conv3, [-1, 20*8*64])
    fc1 = tf.nn.relu(tf.add(tf.matmul(fc1, W_fc1), B_fc1))
    fc1 = tf.nn.dropout(fc1, keep_prob)
    #輸出層
    W_fc2 = weight_variable([1024, CAPTCHA_LEN * CHAR_SET_LEN], 'W_fc2')
    B_fc2 = bias_variable([CAPTCHA_LEN * CHAR_SET_LEN], 'B_fc2')
    output = tf.add(tf.matmul(fc1, W_fc2), B_fc2, 'output')
    
    loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=Y, logits=output))
    optimizer = tf.train.AdamOptimizer(0.001).minimize(loss)
    
    predict = tf.reshape(output, [-1, CAPTCHA_LEN, CHAR_SET_LEN], name='predict')
    labels = tf.reshape(Y, [-1, CAPTCHA_LEN, CHAR_SET_LEN], name='labels')
    #預測結果
    #請注意 predict_max_idx 的 name，在測試model時會用到它
    predict_max_idx = tf.argmax(predict, axis=2, name='predict_max_idx')
    labels_max_idx = tf.argmax(labels, axis=2, name='labels_max_idx')
    predict_correct_vec = tf.equal(predict_max_idx, labels_max_idx)
    accuracy = tf.reduce_mean(tf.cast(predict_correct_vec, tf.float32))
    
    saver = tf.train.Saver()
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        steps = 0
        for epoch in range(6000):
            train_data, train_label = get_next_batch(64, 'train', steps)
            sess.run(optimizer, feed_dict={X : train_data, Y : train_label, keep_prob:0.75})
            if steps % 100 == 0:
                test_data, test_label = get_next_batch(100, 'validate', steps)
                acc = sess.run(accuracy, feed_dict={X : test_data, Y : test_label, keep_prob:1.0})
                print("steps=%d, accuracy=%f" % (steps, acc))
                if acc > 0.99:
                    saver.save(sess, MODEL_SAVE_PATH+"crack_captcha.model", global_step=steps)
                    break
            steps += 1

if __name__ == '__main__':    
    image_filename_list, total = get_image_file_name(CAPTCHA_IMAGE_PATH)
    random.seed(time.time())
    #打亂順序
    random.shuffle(image_filename_list)
    trainImageNumber = int(total * TRAIN_IMAGE_PERCENT)
    #分成測試集
    TRAINING_IMAGE_NAME = image_filename_list[ : trainImageNumber]
    #和驗證集
    VALIDATION_IMAGE_NAME = image_filename_list[trainImageNumber : ]
    train_data_with_CNN()    
    print('Training finished')
 

執行上面的程式碼，開始訓練，訓練要花些時間，如果沒有GPU的話，會慢些，

訓練完後，輸出如下結果，經過4100次的迭代，訓練出來的分類器模型在驗證集上識別的準確率為99.5%

生成的模型檔案如下，在模型測試時將用到這些檔案

3. 測試模型

編寫程式碼，對訓練出來的模型進行測試

import tensorflow as tf

import numpy as np
from PIL import Image
import os
import matplotlib.pyplot as plt 

CAPTCHA_LEN = 4

MODEL_SAVE_PATH = 'E:/Tensorflow/captcha/models/'
TEST_IMAGE_PATH = 'E:/Tensorflow/captcha/test/'

def get_image_data_and_name(fileName, filePath=TEST_IMAGE_PATH):
    pathName = os.path.join(filePath, fileName)
    img = Image.open(pathName)
    #轉為灰度圖
    img = img.convert("L")       
    image_array = np.array(img)    
    image_data = image_array.flatten()/255
    image_name = fileName[0:CAPTCHA_LEN]
    return image_data, image_name

def digitalStr2Array(digitalStr):
    digitalList = []
    for c in digitalStr:
        digitalList.append(ord(c) - ord('0'))
    return np.array(digitalList)

def model_test():
    nameList = []
    for pathName in os.listdir(TEST_IMAGE_PATH):
        nameList.append(pathName.split('/')[-1])
    totalNumber = len(nameList)
    #載入graph
    saver = tf.train.import_meta_graph(MODEL_SAVE_PATH+"crack_captcha.model-4100.meta")
    graph = tf.get_default_graph()
    #從graph取得 tensor，他們的name是在構建graph時定義的(檢視上面第2步裡的程式碼)
    input_holder = graph.get_tensor_by_name("data-input:0")
    keep_prob_holder = graph.get_tensor_by_name("keep-prob:0")
    predict_max_idx = graph.get_tensor_by_name("predict_max_idx:0")
    with tf.Session() as sess:
        saver.restore(sess, tf.train.latest_checkpoint(MODEL_SAVE_PATH))
        count = 0
        for fileName in nameList:
            img_data, img_name = get_image_data_and_name(fileName, TEST_IMAGE_PATH)
            predict = sess.run(predict_max_idx, feed_dict={input_holder:[img_data], keep_prob_holder : 1.0})            
            filePathName = TEST_IMAGE_PATH + fileName
            print(filePathName)
            img = Image.open(filePathName)
            plt.imshow(img)
            plt.axis('off')
            plt.show()
            predictValue = np.squeeze(predict)
            rightValue = digitalStr2Array(img_name)
            if np.array_equal(predictValue, rightValue):
                result = '正確'
                count += 1
            else: 
                result = '錯誤'            
            print('實際值：{}， 預測值：{}，測試結果：{}'.format(rightValue, predictValue, result))
            print('\n')
            
        print('正確率：%.2f%%(%d/%d)' % (count*100/totalNumber, count, totalNumber))

if __name__ == '__main__':
    model_test()

對模型的測試結果如下，在測試集上識別的準確率為 94%

下面是兩個識別錯誤的驗證碼

訓練出的模型放在了下面的雲盤裡，有興趣的同學可以用它做下驗證碼的識別