主要使用Tensorflow深度學(xué)習(xí)框架和卷積神經(jīng)網(wǎng)絡(luò)（CNN）算法實(shí)現(xiàn)對(duì)驗(yàn)證碼識(shí)別的功能里烦。

步驟

1.captcha庫(kù)生成驗(yàn)證碼纺座。
2.構(gòu)造網(wǎng)絡(luò)的輸入數(shù)據(jù)和標(biāo)簽碌秸。
2.基于TensorFlow框架和卷積神經(jīng)網(wǎng)絡(luò)訓(xùn)練自己的驗(yàn)證碼識(shí)別模型

生成驗(yàn)證碼

使用 Python 的 captcha 庫(kù)來生成即可撤卢，這個(gè)庫(kù)默認(rèn)是沒有安裝的，所以需要先安裝這個(gè)庫(kù)集嵌，使用 pip3 安裝即可
代碼如下：

import numpy as np
import tensorflow as tf
from captcha.image import ImageCaptcha#驗(yàn)證碼生成框架
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image#圖片處理標(biāo)準(zhǔn)庫(kù)
import random

number = ['0','1','2','3','4','5','6','7','8','9']
alphabet = ['a','b','c','d','e','f','g','h','i','j','k','l','m','n','o','p','q','r','s','t','u','v','w','x','y','z']
ALPHABET = ['A','B','C','D','E','F','G','H','I','J','K','L','M','N','O','P','Q','R','S','T','U','V','W','X','Y','Z']

def random_captcha_text(char_set=number+alphabet+ALPHABET ,captcha_size=4):
    captcha_text=[]#list
    for i in range(captcha_size ):
        c=random.choice(char_set)
        captcha_text .append(c)
    return captcha_text
def gen_captcha_text_and_image():
    image=ImageCaptcha()
    captcha_text=random_captcha_text()
    captcha_text="".join(captcha_text )#把list中的所有元素放入一個(gè)字符串

    captcha=image.generate(captcha_text )

    captcha_image=Image.open(captcha)
    captcha_image=np.array(captcha_image)#將圖片保存為矩陣
    return captcha_text,captcha_image

if __name__ == '__main__':
    text, image = gen_captcha_text_and_image()
    f = plt.figure()
    ax = f.add_subplot(111)
    ax.text(0.1, 0.9, text, ha='center', va='center', transform=ax.transAxes)
    plt.imshow(image)
    plt.show()

結(jié)果如下：

構(gòu)造網(wǎng)絡(luò)的輸入數(shù)據(jù)和標(biāo)簽

這里將這個(gè)問題看做是分類問題缕粹，驗(yàn)證碼中一共有四個(gè)數(shù)字（或者字母），則共有（10+26+26）4=624 種可能的情況纸淮。這里為了方便進(jìn)行驗(yàn)證，僅僅構(gòu)造了數(shù)字的驗(yàn)證碼亚享，則每個(gè)數(shù)字有10種可能性咽块，共有10*4中可能的情況。
涉及到的主要代碼為：

def convert2gray(img):#將圖像轉(zhuǎn)化為灰度圖
    if len(img.shape) > 2:
        gray = np.mean(img, -1)
        # 上面的轉(zhuǎn)法較快欺税，正規(guī)轉(zhuǎn)法如下
        # r, g, b = img[:,:,0], img[:,:,1], img[:,:,2]
        # gray = 0.2989 * r + 0.5870 * g + 0.1140 * b
        return gray
    else:
        return img

#下面的 text2vec() 方法就是將真實(shí)文本轉(zhuǎn)化為 One-Hot 編碼侈沪，vec2text() 方法就是將 One-Hot 編碼轉(zhuǎn)回真實(shí)文本
def text2vec(text):
    text_len = len(text)
    if text_len > MAX_CAPTCHA:
        raise ValueError('驗(yàn)證碼最長(zhǎng)4個(gè)字符')

    vector = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
    for i, c in enumerate(text):
        idx = i * CHAR_SET_LEN + int(c)
        vector[idx] = 1
    return vector

# 向量轉(zhuǎn)回文本
def vec2text(vec):
    text = []
    char_pos = vec.nonzero()[0]#nonzero(a)返回?cái)?shù)組a中值不為零的元素的下標(biāo)
    for i, c in enumerate(char_pos):
        number = i % 10
        text.append(str(number))

    return "".join(text)

構(gòu)造CNN模型

這里僅僅構(gòu)造了三層卷積，每層卷積后跟池化層晚凿，一個(gè)全連接層+一個(gè)結(jié)果輸出層亭罪。

# 定義CNN
def crack_captcha_cnn(w_alpha=0.01, b_alpha=0.1):
    x = tf.reshape(X, shape=[-1, IMAGE_HEIGHT, IMAGE_WIDTH, 1])#為了滿足TensorFlow的要求
    # w_c1_alpha = np.sqrt(2.0/(IMAGE_HEIGHT*IMAGE_WIDTH)) #
    # w_c2_alpha = np.sqrt(2.0/(3*3*32))
    # w_c3_alpha = np.sqrt(2.0/(3*3*64))
    # w_d1_alpha = np.sqrt(2.0/(8*32*64))
    # out_alpha = np.sqrt(2.0/1024)
    # 3 conv layer
    #3個(gè)卷積+池化
    w_c1 = tf.Variable(w_alpha * tf.random_normal([3, 3, 1, 32]))
    b_c1 = tf.Variable(b_alpha * tf.random_normal([32]))
    conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
    conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    conv1 = tf.nn.dropout(conv1, keep_prob)

    w_c2 = tf.Variable(w_alpha * tf.random_normal([3, 3, 32, 64]))
    b_c2 = tf.Variable(b_alpha * tf.random_normal([64]))
    conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv1, w_c2, strides=[1, 1, 1, 1], padding='SAME'), b_c2))
    conv2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    conv2 = tf.nn.dropout(conv2, keep_prob)

    w_c3 = tf.Variable(w_alpha * tf.random_normal([3, 3, 64, 64]))
    b_c3 = tf.Variable(b_alpha * tf.random_normal([64]))
    conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv2, w_c3, strides=[1, 1, 1, 1], padding='SAME'), b_c3))
    conv3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    conv3 = tf.nn.dropout(conv3, keep_prob)

    # Fully connected layer
    w_d = tf.Variable(w_alpha * tf.random_normal([8 * 20 * 64, 1024]))
    b_d = tf.Variable(b_alpha * tf.random_normal([1024]))
    dense = tf.reshape(conv3, [-1, w_d.get_shape().as_list()[0]])
    dense = tf.nn.relu(tf.add(tf.matmul(dense, w_d), b_d))
    dense = tf.nn.dropout(dense, keep_prob)

    w_out = tf.Variable(w_alpha * tf.random_normal([1024, MAX_CAPTCHA * CHAR_SET_LEN]))
    b_out = tf.Variable(b_alpha * tf.random_normal([MAX_CAPTCHA * CHAR_SET_LEN]))
    out = tf.add(tf.matmul(dense, w_out), b_out)
    return out

補(bǔ)充padding='SAME'和padding='VALID'的區(qū)別。

tensorflow 中的padding方式有“SAME”和“VALID”
這兩個(gè)方式計(jì)算特征圖的輸出大小方式不一樣歼秽。

1. valid： 周圍不填0進(jìn)行卷積運(yùn)算应役，無法計(jì)算的部分（矩陣的右邊和下面）直接舍去。
   如果padding的方法為“VALID” ，那么輸出特征圖的長(zhǎng)和寬為:
   new_width=new_height=? W-F+1 / S?
2. same: 周圍填0進(jìn)行卷積運(yùn)算箩祥，0在矩陣的左右和上下均勻添加院崇，非均勻時(shí)多的加在右邊和下面。
   如果padding的方法為“SAME” 袍祖，那么輸出特征圖的長(zhǎng)和寬為:
   new_width=new_height=? W / S?
   W為輸入的size底瓣，F(xiàn)為filter為size，S為步長(zhǎng)蕉陋，? ?為向上取整符號(hào)

定義一次訓(xùn)練的batch捐凭，定義損失函數(shù)，訓(xùn)練優(yōu)化模型

# 生成一個(gè)訓(xùn)練batch
def get_next_batch(batch_size=128):
    batch_x = np.zeros([batch_size, IMAGE_HEIGHT * IMAGE_WIDTH])
    batch_y = np.zeros([batch_size, MAX_CAPTCHA * CHAR_SET_LEN])

    # 有時(shí)生成圖像大小不是(60, 160, 3)
    def wrap_gen_captcha_text_and_image():
        while True:
            text, image = gen_captcha_text_and_image()
            if image.shape == (60, 160, 3):
                return text, image
    for i in range(batch_size):
        text, image = wrap_gen_captcha_text_and_image()
        image = convert2gray(image)

        batch_x[i, :] = image.flatten() / 255  # (image.flatten()-128)/128  mean為0凳鬓。即值為從0到1之間
        batch_y[i, :] = text2vec(text)

    return batch_x, batch_y

#定義損失函數(shù)茁肠，進(jìn)行模型迭代優(yōu)化
def train_crack_captcha_cnn():
    output = crack_captcha_cnn()
    loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=output, labels=Y))
    optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
    predict = tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN])
    max_idx_p = tf.argmax(predict, 2)
    max_idx_l = tf.argmax(tf.reshape(Y, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2)
    correct_pred = tf.equal(max_idx_p, max_idx_l)
    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

    saver = tf.train.Saver()
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        step = 0
        while True:
            batch_x, batch_y = get_next_batch(64)#數(shù)據(jù)需要一個(gè)batch一個(gè)batch進(jìn)行傳入
            _, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75})
            #print(step, loss_)
            # 每10 step計(jì)算一次準(zhǔn)確率
            if step % 10 == 0:
                batch_x_test, batch_y_test = get_next_batch(100)
                acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.})
                print("Step: ",step,"準(zhǔn)確率：" ,acc)
                # 如果準(zhǔn)確率大于50%,保存模型,完成訓(xùn)練
                if acc > 0.60:
                    saver.save(sess, "./model/crack_capcha.model", global_step=step)
                    break
            step += 1