一、前言

二、實戰

1、驗證碼生成

import random
import numpy as np
from PIL import Image
from captcha.image import ImageCaptcha


NUMBER = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
LOW_CASE = ['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']
UP_CASE = ['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']
CAPTCHA_LIST = NUMBER + LOW_CASE + UP_CASE
CAPTCHA_LEN = 4
CAPTCHA_HEIGHT = 60
CAPTCHA_WIDTH = 160


def random_captcha_text(char_set=CAPTCHA_LIST, captcha_size=CAPTCHA_LEN):
    '''
    隨機生成驗證碼文字
    :param char_set:
    :param captcha_size:
    :return:
    '''
    captcha_text = [random.choice(char_set) for _ in range(captcha_size)]
    return ''.join(captcha_text)


def gen_captcha_text_and_image(width=CAPTCHA_WIDTH, height=CAPTCHA_HEIGHT,save=None):
    '''
    生成隨機驗證碼
    :param width:
    :param height:
    :param save:
    :return: np陣列
    '''
    image = ImageCaptcha(width=width, height=height)
    # 驗證碼文字
    captcha_text = random_captcha_text()
    captcha = image.generate(captcha_text)
    # 儲存
    if save: image.write(captcha_text, captcha_text + '.jpg')
    captcha_image = Image.open(captcha)
    # 轉化為np陣列
    captcha_image = np.array(captcha_image)
    return captcha_text, captcha_image
 

2、權重、偏置及工具函式定義

def weight_variable(shape, w_alpha=0.01):
    '''
    增加噪音，隨機生成權重
    :param shape:
    :param w_alpha:
    :return:
    '''
    initial = w_alpha * tf.random_normal(shape)
    return tf.Variable(initial)


def bias_variable(shape, b_alpha=0.1):
    '''
    增加噪音，隨機生成偏置項
    :param shape:
    :param b_alpha:
    :return:
    '''
    initial = b_alpha * tf.random_normal(shape)
    return tf.Variable(initial)


def conv2d(x, w):
    '''
    區域性變數線性組合，步長為1，模式‘SAME’代表卷積後圖片尺寸不變，即零邊距
    :param x:
    :param w:
    :return:
    '''
    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
    '''
    max pooling,取出區域內最大值為代表特徵， 2x2pool，圖片尺寸變為1/2
    :param x:
    :return:
    '''
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
 

def cnn_graph(x, keep_prob, size, captcha_list=CAPTCHA_LIST, captcha_len=CAPTCHA_LEN):
    '''
    三層卷積神經網路計算圖
    :param x:
    :param keep_prob:
    :param size:
    :param captcha_list:
    :param captcha_len:
    :return:
    '''
    # 圖片reshape為4維向量
    image_height, image_width = size
    x_image = tf.reshape(x, shape=[-1, image_height, image_width, 1])

    # layer 1
    # filter定義為3x3x1， 輸出32個特徵, 即32個filter
    w_conv1 = weight_variable([3, 3, 1, 32])
    b_conv1 = bias_variable([32])
    # rulu啟用函式
    h_conv1 = tf.nn.relu(tf.nn.bias_add(conv2d(x_image, w_conv1), b_conv1))
    # 池化
    h_pool1 = max_pool_2x2(h_conv1)
    # dropout防止過擬合
    h_drop1 = tf.nn.dropout(h_pool1, keep_prob)

    # layer 2
    w_conv2 = weight_variable([3, 3, 32, 64])
    b_conv2 = bias_variable([64])
    h_conv2 = tf.nn.relu(tf.nn.bias_add(conv2d(h_drop1, w_conv2), b_conv2))
    h_pool2 = max_pool_2x2(h_conv2)
    h_drop2 = tf.nn.dropout(h_pool2, keep_prob)

    # layer 3
    w_conv3 = weight_variable([3, 3, 64, 64])
    b_conv3 = bias_variable([64])
    h_conv3 = tf.nn.relu(tf.nn.bias_add(conv2d(h_drop2, w_conv3), b_conv3))
    h_pool3 = max_pool_2x2(h_conv3)
    h_drop3 = tf.nn.dropout(h_pool3, keep_prob)

    # full connect layer
    image_height = int(h_drop3.shape[1])
    image_width = int(h_drop3.shape[2])
    w_fc = weight_variable([image_height*image_width*64, 1024])
    b_fc = bias_variable([1024])
    h_drop3_re = tf.reshape(h_drop3, [-1, image_height*image_width*64])
    h_fc = tf.nn.relu(tf.add(tf.matmul(h_drop3_re, w_fc), b_fc))
    h_drop_fc = tf.nn.dropout(h_fc, keep_prob)

    # out layer
    w_out = weight_variable([1024, len(captcha_list)*captcha_len])
    b_out = bias_variable([len(captcha_list)*captcha_len])
    y_conv = tf.add(tf.matmul(h_drop_fc, w_out), b_out)
    return y_conv
 

def optimize_graph(y, y_conv):
    '''
    優化計算圖
    :param y:
    :param y_conv:
    :return:
    '''
    # 交叉熵計算loss 注意logits輸入是在函式內部進行sigmod操作
    # sigmod_cross適用於每個類別相互獨立但不互斥，如圖中可以有字母和數字
    # softmax_cross適用於每個類別獨立且排斥的情況，如數字和字母不可以同時出現
    loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y_conv, labels=y))
    # 最小化loss優化
    optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
    return optimizer


def accuracy_graph(y, y_conv, width=len(CAPTCHA_LIST), height=CAPTCHA_LEN):
    '''
    偏差計算圖
    :param y:
    :param y_conv:
    :param width:
    :param height:
    :return:
    '''
    # 這裡區分了大小寫 實際上驗證碼一般不區分大小寫
    # 預測值
    predict = tf.reshape(y_conv, [-1, height, width])
    max_predict_idx = tf.argmax(predict, 2)
    # 標籤
    label = tf.reshape(y, [-1, height, width])
    max_label_idx = tf.argmax(label, 2)
    correct_p = tf.equal(max_predict_idx, max_label_idx)
    accuracy = tf.reduce_mean(tf.cast(correct_p, tf.float32))
    return accuracy

def train(height=CAPTCHA_HEIGHT, width=CAPTCHA_WIDTH, y_size=len(CAPTCHA_LIST)*CAPTCHA_LEN):
    '''
    cnn訓練
    :param height:
    :param width:
    :param y_size:
    :return:
    '''
    # cnn在影象大小是2的倍數時效能最高, 如果影象大小不是2的倍數，可以在影象邊緣補無用畫素
    # 在影象上補2行，下補3行，左補2行，右補2行
    # np.pad(image,((2,3),(2,2)), 'constant', constant_values=(255,))

    acc_rate = 0.95
    # 按照圖片大小申請佔位符
    x = tf.placeholder(tf.float32, [None, height * width])
    y = tf.placeholder(tf.float32, [None, y_size])
    # 防止過擬合 訓練時啟用 測試時不啟用
    keep_prob = tf.placeholder(tf.float32)
    # cnn模型
    y_conv = cnn_graph(x, keep_prob, (height, width))
    # 最優化
    optimizer = optimize_graph(y, y_conv)
    # 偏差
    accuracy = accuracy_graph(y, y_conv)
    # 啟動會話.開始訓練
    saver = tf.train.Saver()
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    step = 0
    while 1:
        batch_x, batch_y = next_batch(64)
        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y, keep_prob: 0.75})
        # 每訓練一百次測試一次
        if step % 100 == 0:
            batch_x_test, batch_y_test = next_batch(100)
            acc = sess.run(accuracy, feed_dict={x: batch_x_test, y: batch_y_test, keep_prob: 1.0})
            print(datetime.now().strftime('%c'), ' step:', step, ' accuracy:', acc)
            # 偏差滿足要求，儲存模型
            if acc > acc_rate:
                model_path = os.getcwd() + os.sep + str(acc_rate) + "captcha.model"
                saver.save(sess, model_path, global_step=step)
                acc_rate += 0.01
                if acc_rate > 0.99: break
        step += 1
    sess.close()

三、其他