卷积神经网络（CNN）对验证码图片识别案例

数据集

数据集下载

链接：https://pan.baidu.com/s/1ypNNQkR1_ZK-_KO92x6Phw?pwd=6753 
提取码：6753

图片1  -->NZPP   一个样本对应四个目标值

NZPP  ---【13,25,15,15】

使用one-hot编码转换

第一个位置：[0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0]

第二个位置：[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1]

第三个位置：[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,]

第四个位置：[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,]

如何衡量损失？

       softmax交叉熵  ---适合一个样本对应一个目标值

       sigmoid交叉熵  ---每个类别独立且不互相排斥 适合多个目标值情况

准确率如何计算？

        核心：对比真实值和预测值最大值所在位置是否一致 

                 本案例要比对4个26列 4个所有一致才能说明识别正确

                  y_predict[None,4,26]

                 tf.argmax(y_predict,axis = 2)  将返回[[True],[True],[True],[True]]等结果

                  判断全是True才能返回True  ---->tf.reduce_all()

流程分析

6000张图片

1）读取图片数据

              filename --->标签值

2）解析CSV文件，将标签值处理成数字形式  例 NZPP  ---[13,25,15,15]

3）将filename和标签值联系起来

4）构建卷积神经网络 --->y_predict

5）构造损失函数  sigmoid交叉熵损失

6）优化损失

7）计算准确率

8）开启会话、开启线程

代码实现

import tensorflow as tf
import glob
import pandas as pd
import numpy as np
tf.compat.v1.disable_eager_execution()

# 1) 读取图片数据filename  --标签值
def read_pic():
    """
    读取图片数据
    :return:
    """
    #1、构建文件名队列
    file_names = glob.glob("./tmp/GenPics/*.jpg")
    # print("file_names:\n",file_names)
    file_queue = tf.compat.v1.train.string_input_producer(file_names)
    # 2、读取与解码
    reader = tf.compat.v1.WholeFileReader()
    filename,image = reader.read(file_queue)
    # 解码
    decoded = tf.image.decode_jpeg(image)
    # 更新图像，将图片形状确定下来
    decoded.set_shape([20,80,3])
    # 修改图像类型
    image_cast = tf.cast(decoded,tf.float32)
    # 批处理
    filename_batch,image_batch = tf.compat.v1.train.batch([filename,image_cast],batch_size=100,num_threads=1,capacity=200)
    return filename_batch,image_batch
# 2） 解析CSV使得成为 将标签值NZPP->[13, 25, 15, 15]
def parse_csv():
    """
    解析CSV文件，建立文件名和标签值对应表格
    :return:
    """
    csv_data = pd.read_csv("./tmp/GenPics/labels.csv", names=["file_num", "chars"], index_col="file_num")

    labels = []
    for label in csv_data["chars"]:
        tmp = []
        for letter in label:
            tmp.append(ord(letter) - ord("A"))
        labels.append(tmp)
    csv_data["labels"] = labels
    print(csv_data)
    return csv_data
# 3）将filename和标签值联系起来
def filename2label(filenames, csv_data):
    """
    将filename和标签值联系起来
    :param filenames:
    :param csv_data:
    :return:
    """
    labels = []
    # 将b'文件名中的数字提取出来并索引相应的标签值
    for filename in filenames:
        digit_str = "".join(list(filter(str.isdigit, str(filename))))
        label = csv_data.loc[int(digit_str), "labels"]
        labels.append(label)

     #print("labels:\n", labels)

    return np.array(labels)
# 4）构建卷积神经网络
def create_weights(shape):

    return tf.Variable(initial_value=tf.compat.v1.random_normal(shape=shape, stddev=0.01))

def create_model(x):
    """
    构建卷积神经网络
    :param x:[None, 20, 80, 3]
    :return:
    """
    # 1）第一个卷积大层
    with tf.compat.v1.variable_scope("conv1"):

        # 卷积层
        # 定义filter和偏置
        conv1_weights = create_weights(shape=[5, 5, 3, 32])
        conv1_bias = create_weights(shape=[32])
        conv1_x = tf.nn.conv2d(input=x, filters=conv1_weights, strides=[1, 1, 1, 1], padding="SAME") + conv1_bias

        # 激活层
        relu1_x = tf.nn.relu(conv1_x)

        # 池化层
        pool1_x = tf.nn.max_pool(input=relu1_x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")

    # 2）第二个卷积大层
    with tf.compat.v1.variable_scope("conv2"):
        # [None, 20, 80, 3] --> [None, 10, 40, 32]
        # 卷积层
        # 定义filter和偏置
        conv2_weights = create_weights(shape=[5, 5, 32, 64])
        conv2_bias = create_weights(shape=[64])
        conv2_x = tf.nn.conv2d(input=pool1_x, filters=conv2_weights, strides=[1, 1, 1, 1], padding="SAME") + conv2_bias

        # 激活层
        relu2_x = tf.nn.relu(conv2_x)

        # 池化层
        pool2_x = tf.nn.max_pool(input=relu2_x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")

    # 3）全连接层
    with tf.compat.v1.variable_scope("full_connection"):
        # [None, 10, 40, 32] -> [None, 5, 20, 64]
        # [None, 5, 20, 64] -> [None, 5 * 20 * 64]
        # [None, 5 * 20 * 64] * [5 * 20 * 64, 4 * 26] = [None, 4 * 26]
        x_fc = tf.reshape(pool2_x, shape=[-1, 5 * 20 * 64])
        weights_fc = create_weights(shape=[5 * 20 * 64, 4 * 26])
        bias_fc = create_weights(shape=[104])
        y_predict = tf.matmul(x_fc, weights_fc) + bias_fc

    return y_predict



if __name__ == "__main__":
    filename,image = read_pic()
    csv_data = parse_csv()
    # 1、准备数据
    x = tf.compat.v1.placeholder(tf.float32, shape=[None, 20, 80, 3])
    y_true = tf.compat.v1.placeholder(tf.float32, shape=[None, 4 * 26])

    # 2、构建模型
    y_predict = create_model(x)

    # 3、构造损失函数
    loss_list = tf.nn.sigmoid_cross_entropy_with_logits(labels=y_true, logits=y_predict)
    loss = tf.reduce_mean(loss_list)

    # 4、优化损失
    optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=0.001).minimize(loss)

    # 5、计算准确率
    equal_list = tf.reduce_all(
        tf.equal(tf.argmax(tf.reshape(y_predict, shape=[-1, 4, 26]), axis=2),
                 tf.argmax(tf.reshape(y_true, shape=[-1, 4, 26]), axis=2)), axis=1)
    accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32))

    # 初始化变量
    init = tf.compat.v1.global_variables_initializer()

    # 开启会话
    with tf.compat.v1.Session() as sess:
        # 初始化变量
        sess.run(init)
        coord = tf.compat.v1.train.Coordinator()
        threads = tf.compat.v1.train.start_queue_runners(sess=sess, coord=coord)
        for i in range(500):
            filename_value, image_value = sess.run([filename, image])
            # print("filename_value:\n",filename_value)
            # print("image_value:\n",image_value)
            labels = filename2label(filename_value, csv_data)
            # 将标签值转换成one-hot
            labels_value = tf.reshape(tf.one_hot(labels, depth=26), [-1, 4 * 26]).eval()

            _, error, accuracy_value = sess.run([optimizer, loss, accuracy],
                                                feed_dict={x: image_value, y_true: labels_value})

            print("第%d次训练后损失为%f，准确率为%f" % (i + 1, error, accuracy_value))

        # 回收线程
        coord.request_stop()
        coord.join(threads)

数据读取解码

解析CSV文件，文件名与标签值对应起来，

训练结果：