TensorFlow识别纯数字验证码过程记录

亦凉 2022-05-23 06:51 275阅读 0赞

最初的代码不是我写的，别人给我的，源代码的作者已经不知道了，抱歉，在简书上找到类似的[TensorFlow练习20: 使用深度学习破解字符验证码][TensorFlow_20_]。这里我作为一个初学者，打算从零开始记录我学习的全过程。  
**一、验证码识别**  
![1][]  
**1.下载图片存储并转为数组**

import os,requests,re
    import random
    import numpy as np
    from PIL import Image
    from io import BytesIO
    headers = {
    'Accept':'image/webp,image/apng,image/*,*/*;q=0.8',
    'Accept-Language':'zh-CN,zh;q=0.8',
    'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; WOW64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/59.0.3071.86 Safari/537.36',
    }
    
    def get_captcha(url,i=1):
        try:
            resp = requests.get(url,headers=headers).content
            # [Python中BytesIO的使用场景，什么时间应该使用BytesIO？](https://www.zhihu.com/question/49102468)
            # [Python图像处理库PIL的Image模块介绍（一）](https://blog.csdn.net/icamera0/article/details/50654910)
            pic = Image.open(BytesIO(resp))
            # print(i)
            name ='F:\验证码\\test\\' + str(i) + '.jpg'
            pic.save(name)
            #
            # Image.ANTIALIAS 高质量
            captcha_image = pic.resize((160, 60),Image.ANTIALIAS)
            # [图像数组为什么要转换成字符数组](https://zhidao.baidu.com/question/1513565352058747300.html)
            #[图片怎么转为矩阵数组](https://zhidao.baidu.com/question/1580093335915387140.html)
            captcha_image = np.array(captcha_image)
            return name,captcha_image
        except Exception as e:
            print('错误类型：',e)

**2.将彩色图片转为灰色图像**  
色彩对识别验证码没什么用，转成灰色，去除多余的杂音，同时也对图片进行降维度。  
通过ImageCaptcha生成的图片，可以看到`img.shape`的值是`<class 'tuple'>: (60, 160, 3)`  
[python PIL 图像处理 （二）][python PIL _]中图像数组表示，第一个元组表示图像数组的大小（行、列、颜色通道），那么这里的3就是颜色通道，而60\*160应该就是像素。

def convert2gray(img):
        if len(img.shape) > 2:
            # 
            gray = np.mean(img, -1)
            return gray
        else:
            plt.imshow(img)
            plt.show()
            return img

**3 将图片一唯化**

image = image.flatten() / 255

**4 定义CNN模型**  
看了下面的代码，懵逼了，不知道在说什么。  
先从这篇[深度学习之卷积神经网络CNN及tensorflow代码实现示例][CNN_tensorflow]文章了解到什么是CNN，  
《TensorFlow 官方文档》中提到卷积神经网络是为图像识别量身定做的一个模型  
[卷积神经网络CNN原理以及TensorFlow实现][CNN_TensorFlow]  
[卷积神经网络全面解析][Link 1]  
[CNN笔记：通俗理解卷积神经网络][CNN]  
理解CNN模型本身就是件困难的事情，那么另外的思路就是搞清楚，为什么要这么传参，

# 申请三个占位符
    # [【Tensorflow】tf.placeholder函数](https://blog.csdn.net/zj360202/article/details/70243127)
    X = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT*IMAGE_WIDTH])
    Y = tf.placeholder(tf.float32, [None, MAX_CAPTCHA*CHAR_SET_LEN])
    keep_prob = tf.placeholder(tf.float32) # dropout
    
    def crack_captcha_cnn(w_alpha=0.01, b_alpha=0.1):
        # [TensorFlow的reshape操作 tf.reshape](https://blog.csdn.net/lxg0807/article/details/53021859)
        x = tf.reshape(X, shape=[-1, IMAGE_HEIGHT, IMAGE_WIDTH, 1])
    
        # 3 conv layer # 3 个 转换层
        # [tf.random_normal()函数](https://blog.csdn.net/dcrmg/article/details/79028043)，从服从指定正太分布的数值中取出指定个数的值
        # [3, 3, 1, 32]是[batch, height, width, channels]这样的shape
        w_c1 = tf.Variable(w_alpha*tf.random_normal([3, 3, 1, 32]))
        b_c1 = tf.Variable(b_alpha*tf.random_normal([32]))
        # [tf.nn.bias_add和tf.add、tf.add_n](https://blog.csdn.net/weixin_38698649/article/details/80100737)，bias必须是一唯的
        # [TF-卷积函数 tf.nn.conv2d 介绍](https://www.cnblogs.com/qggg/p/6832342.html)
        # [tensorflow激活函数relu()的详解](https://blog.csdn.net/jiao_mrswang/article/details/74537066)
        # [ReLu(Rectified Linear Units)激活函数](http://www.mamicode.com/info-detail-873243.html)
        conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
        # [【TensorFlow】tf.nn.max_pool实现池化操作](https://blog.csdn.net/mao_xiao_feng/article/details/53453926)
        conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        # [TensorFlow学习---tf.nn.dropout防止过拟合](https://blog.csdn.net/huahuazhu/article/details/73649389)
        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 # 最后连接层
        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)
    
        # 输出层
        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]))
        # [mdumplingtf.multiply（）和tf.matmul（）区别](https://www.cnblogs.com/mdumpling/p/8094310.html)
        out = tf.add(tf.matmul(dense, w_out), b_out)
        #out = tf.nn.softmax(out)
        return out

**5 执行CNN模型**

def crack_captcha(captcha_image):
        output = crack_captcha_cnn()
        # [Tensorflow系列——Saver的用法](https://blog.csdn.net/u011500062/article/details/51728830)
        saver = tf.train.Saver()
        with tf.Session() as sess:
            # 重载模型的参数，继续训练或用于测试数据
            # tf.train.latest_checkpoint自动获取最后一次的模型
            saver.restore(sess, tf.train.latest_checkpoint('.\\'))
            # [tf.argmax()以及axis解析](https://blog.csdn.net/qq575379110/article/details/70538051/),返回最大的那个数值所在的下标
            predict = tf.argmax(tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2)
            text_list = sess.run(predict, feed_dict={X: [captcha_image], keep_prob: 1})
            text = text_list[0].tolist()
            vector = np.zeros(MAX_CAPTCHA*CHAR_SET_LEN)
            i = 0
            for n in text:
                vector[i*CHAR_SET_LEN + n] = 1
                i += 1
            return vec2text(vector)

**6 向量转文本**  
调试后，我很纳闷，为啥下图中text中已经得到了验证码的内容，还要大费周折，做向量转文本呢？  
我猜唯一的解释就是如果既有数字，又有字母通用，下次遇到既有数字，又有字母验证码的我再调试看看。  
![2][]  
[Python ord() 函数][Python ord_]用于获取 ASCII 数值

text, image = get_image_text()
    # print("验证码图像channel:", image.shape) # (60, 160, 3)
    # 图像大小
    IMAGE_HEIGHT = 60
    IMAGE_WIDTH = 160
    MAX_CAPTCHA = len(text) # 如果验证码的长度是固定的，这个可以设置成固定值
    number =  ['0','1','2','3','4','5','6','7','8','9']
    char_set = number +['_']  # 如果验证码长度小于4, '_'用来补齐
    CHAR_SET_LEN = len(char_set)
    def vec2text(vec):
        char_pos = vec.nonzero()[0]
        text=[]
        for i, c in enumerate(char_pos):
            char_at_pos = i #c/63
            char_idx = c % CHAR_SET_LEN
            if char_idx < 10:
                char_code = char_idx + ord('0')
            elif char_idx <36:
                char_code = char_idx - 10 + ord('A')
            elif char_idx < 62:
                char_code = char_idx-  36 + ord('a')
            elif char_idx == 62:
                char_code = ord('_')
            else:
                raise ValueError('error')
            text.append(chr(char_code))
        return "".join(text)

为什么是160\*160的图片呢？查看`from captcha.image import ImageCaptcha`代码，产生的验证码像素默认就是160\*60  
![1][1 1]

**二、训练模型**  
![1][1 2]  
**1、构建CNN模型**  
上述已经描述了  
**2、构建训练模型**

def train_crack_captcha_cnn():
        output = crack_captcha_cnn()
        # [tensorflow学习之常用函数总结：tensorflow官方例子中的诸如tf.reduce_mean()这类函数](https://blog.csdn.net/qq_32166627/article/details/52734387)
        #loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(output, Y))
        # [Tensorflow四种交叉熵函数计算公式：tf.nn.cross_entropy](https://blog.csdn.net/QW_sunny/article/details/72885403)
        loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=output, labels=Y))
        # 最后一层用来分类的softmax和sigmoid有什么不同？
        # optimizer 为了加快训练 learning_rate应该开始大，然后慢慢衰
        # [TensorFlow学习（四）：优化器Optimizer](https://blog.csdn.net/xierhacker/article/details/53174558)
        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)
        # correct_pred = tf.equal(tf.argmax(y_conv,1), y_)
        # [【Tensorflow】tf.cast 类型转换 函数](https://blog.csdn.net/zj360202/article/details/70260265)
        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)
                _, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75})
                print(step, loss_)
    
                # 每100 step计算一次准确率
                if step % 100 == 0:
                    batch_x_test, batch_y_test = get_next_batch(50)
                    acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.})
                    print('正确率:',step, acc)
                    # 如果准确率大于0.99995,保存模型,完成训练
                    if acc > 0.99995:
                        saver.save(sess, ".\\crack_capcha.model", global_step=step)
                        break
                step += 1

**3、执行训练模型**

# 生成一个训练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])
    
        # 有时生成图像大小不是(60, 160, 3)
        def wrap_gen_captcha_text_and_image():
            ''' 获取一张图，判断其是否符合（60，160，3）的规格'''
            while True:
                try:
                    text, image = get_image_text()
                except:
                    pass
                if image.shape == (60, 160, 3):#此部分应该与开头部分图片宽高吻合
                    return text, image
    
        for i in range(batch_size):
            text, image = wrap_gen_captcha_text_and_image()
            #灰度化
            image = convert2gray(image)
    
            # 将图片数组一维化 同时将文本也对应在两个二维组的同一行
            batch_x[i,:] = image.flatten() / 255 # (image.flatten()-128)/128 mean为0
            batch_y[i,:] = text2vec(text)
        # 返回该训练批次
        return batch_x, batch_y

**4、评估训练模型**  
《TensorFlow官方文档》中关于评估模型，提到`tf.argmax`,它能给出某个tensor对象在某一维上的其数据最大值所在的索引值。由于标签向量是由0、1组成，因此最大值1所在的索引位置就是类别标签。  
`acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.})`这段代码用于计算所学习到的模型在测试数据集上面的正确率。

[TensorFlow_20_]: https://www.jianshu.com/p/0b165b187b06
[1]: /images/20220523/5ac6755d54474b1e9c731149e54ea190.png
[python PIL _]: https://blog.csdn.net/u013378306/article/details/70156842
[CNN_tensorflow]: https://blog.csdn.net/cxmscb/article/details/71023576
[CNN_TensorFlow]: https://blog.csdn.net/xukaiwen_2016/article/details/70880694
[Link 1]: http://www.moonshile.com/post/juan-ji-shen-jing-wang-luo-quan-mian-jie-xi
[CNN]: https://blog.csdn.net/v_july_v/article/details/51812459
[2]: /images/20220523/ba0e863bf63f4a4a93c04eb1a9e7d9a3.png
[Python ord_]: http://www.runoob.com/python/python-func-ord.html
[1 1]: /images/20220523/d4868cac41c54980a463d3070b32ea53.png
[1 2]: /images/20220523/3f5e4aa431994088bd54a43518e3f8f1.png