本項目介紹利用深度學習技術（tensorflow），來識別驗證碼（4位驗證碼，具體的驗證碼的長度可以自己生成，可以在自己進行訓練）

程序分為四個部分

1、生成驗證碼的程序，可生成數字+字母大小寫的任意長度驗證碼

# coding:utf-8
# name:captcha_gen.pyimport random
import numpy as np
from PIL import Image
from captcha.image import ImageCaptchaNUMBER = ['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
CAPTCHA_LEN = 4         # 驗證碼長度
CAPTCHA_HEIGHT = 60     # 驗證碼高度
CAPTCHA_WIDTH = 160     # 驗證碼寬度def random_captcha_text(char_set=CAPTCHA_LIST, captcha_size=CAPTCHA_LEN):"""隨機生成定長字符串:param char_set: 備選字符串列表:param captcha_size: 字符串長度:return: 字符串"""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: 驗證碼圖片寬度:param height: 驗證碼圖片高度:param save: 是否保存（None）:return: 驗證碼字符串，驗證碼圖像np數組"""image = ImageCaptcha(width=width, height=height)# 驗證碼文本captcha_text = random_captcha_text()captcha = image.generate(captcha_text)# 保存if save:image.write(captcha_text, './img/' + captcha_text + '.jpg')captcha_image = Image.open(captcha)# 轉化為np數組captcha_image = np.array(captcha_image)return captcha_text, captcha_imageif __name__ == '__main__':t, im = gen_captcha_text_and_image(save=True)print(t, im.shape)      # (60, 160, 3)

?2、工具庫，用于調用驗證碼生成程序來生成訓練集

# -*- coding:utf-8 -*-
# name: util.pyimport numpy as np
from captcha_gen import gen_captcha_text_and_image
from captcha_gen import CAPTCHA_LIST, CAPTCHA_LEN, CAPTCHA_HEIGHT, CAPTCHA_WIDTHdef convert2gray(img):"""圖片轉為黑白，3維轉1維:param img: np:return:  灰度圖的np"""if len(img.shape) > 2:img = np.mean(img, -1)return imgdef text2vec(text, captcha_len=CAPTCHA_LEN, captcha_list=CAPTCHA_LIST):"""驗證碼文本轉為向量 啞編碼 方式:param text::param captcha_len::param captcha_list::return: vector 文本對應的向量形式"""text_len = len(text)    # 欲生成驗證碼的字符長度if text_len > captcha_len:raise ValueError('驗證碼最長4個字符')vector = np.zeros(captcha_len * len(captcha_list))      # 生成一個一維向量 驗證碼長度*字符列表長度for i in range(text_len):vector[captcha_list.index(text[i])+i*len(captcha_list)] = 1     # 找到字符對應在字符列表中的下標值+字符列表長度*i 的 一維向量 賦值為 1return vectordef vec2text(vec, captcha_list=CAPTCHA_LIST, captcha_len=CAPTCHA_LEN):"""驗證碼向量轉為文本:param vec::param captcha_list::param captcha_len::return: 向量的字符串形式"""vec_idx = vectext_list = [captcha_list[int(v)] for v in vec_idx]return ''.join(text_list)def wrap_gen_captcha_text_and_image(shape=(CAPTCHA_HEIGHT, CAPTCHA_WIDTH, 3)):"""返回特定shape圖片:param shape::return:"""while True:t, im = gen_captcha_text_and_image()if im.shape == shape:return t, imdef get_next_batch(batch_count=60, width=CAPTCHA_WIDTH, height=CAPTCHA_HEIGHT):"""獲取訓練圖片組:param batch_count: default 60:param width: 驗證碼寬度:param height: 驗證碼高度:return: batch_x, batch_yc"""batch_x = np.zeros([batch_count, width * height])batch_y = np.zeros([batch_count, CAPTCHA_LEN * len(CAPTCHA_LIST)])for i in range(batch_count):    # 生成對應的訓練集text, image = wrap_gen_captcha_text_and_image()image = convert2gray(image)     # 轉灰度numpy# 將圖片數組一維化 同時將文本也對應在兩個二維組的同一行batch_x[i, :] = image.flatten() / 255batch_y[i, :] = text2vec(text)  # 驗證碼文本的向量形式# 返回該訓練批次return batch_x, batch_yif __name__ == '__main__':x, y = get_next_batch(batch_count=1)    # 默認為1用于測試集print(x, y)

3、訓練程序，并將準確率超過0.95的模型保存到 ./model/ 文件夾下

# -*- coding:utf-8 -*-
# name: model_train.pyimport tensorflow as tf
from datetime import datetime
from util import get_next_batch
from captcha_gen import CAPTCHA_HEIGHT, CAPTCHA_WIDTH, CAPTCHA_LEN, CAPTCHA_LISTdef weight_variable(shape, w_alpha=0.01):"""初始化權值: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):"""初始化偏置項:param shape::param b_alpha::return:"""initial = b_alpha * tf.random_normal(shape)return tf.Variable(initial)def conv2d(x, w):"""卷基層 ：局部變量線性組合，步長為1，模式‘SAME’代表卷積后圖片尺寸不變，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):"""池化層：max pooling,取出區域內最大值為代表特征， 2x2 的pool，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):"""三層卷積神經網絡:param x:   訓練集 image x:param keep_prob:   神經元利用率:param size:        大小 (高,寬):param captcha_list::param captcha_len::return: y_conv"""# 需要將圖片reshape為4維向量image_height, image_width = sizex_image = tf.reshape(x, shape=[-1, image_height, image_width, 1])# 第一層# filter 定義為3x3x1， 輸出32個特征, 即32個filterw_conv1 = weight_variable([3, 3, 1, 32])    # 3*3的采樣窗口，32個（通道）卷積核從1個平面抽取特征得到32個特征平面b_conv1 = bias_variable([32])h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)    # rulu激活函數h_pool1 = max_pool_2x2(h_conv1)     # 池化h_drop1 = tf.nn.dropout(h_pool1, keep_prob)      # dropout 防止過擬合# 第二層w_conv2 = weight_variable([3, 3, 32, 64])b_conv2 = bias_variable([64])h_conv2 = tf.nn.relu(conv2d(h_drop1, w_conv2) + b_conv2)h_pool2 = max_pool_2x2(h_conv2)h_drop2 = tf.nn.dropout(h_pool2, keep_prob)# 第三層w_conv3 = weight_variable([3, 3, 64, 64])b_conv3 = bias_variable([64])h_conv3 = tf.nn.relu(conv2d(h_drop2, w_conv3) + b_conv3)h_pool3 = max_pool_2x2(h_conv3)h_drop3 = tf.nn.dropout(h_pool3, keep_prob)"""原始：60*160圖片 第一次卷積后 60*160 第一池化后 30*80*32第二次卷積后 30*80*32 ，第二次池化后 15*40*64第三次卷積后 15*40*64 ，第三次池化后 7.5*20*64 = > 向下取整 7*20*64經過上面操作后得到 64 個 7*20的平面"""# 全連接層image_height = int(h_drop3.shape[1])image_width = int(h_drop3.shape[2])w_fc = weight_variable([image_height*image_width*64, 1024])     # 上一層有64個神經元 全連接層有1024個神經元b_fc = bias_variable([1024])h_drop3_re = tf.reshape(h_drop3, [-1, image_height*image_width*64])h_fc = tf.nn.relu(tf.matmul(h_drop3_re, w_fc) + b_fc)h_drop_fc = tf.nn.dropout(h_fc, keep_prob)# 輸出層w_out = weight_variable([1024, len(captcha_list)*captcha_len])b_out = bias_variable([len(captcha_list)*captcha_len])y_conv = tf.matmul(h_drop_fc, w_out) + b_outreturn y_convdef optimize_graph(y, y_conv):"""優化計算圖:param y: 正確值:param y_conv:  預測值:return: optimizer"""# 交叉熵代價函數計算loss 注意 logits 輸入是在函數內部進行sigmod操作# sigmod_cross適用于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適用于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(labels=y, logits=y_conv))# 最小化loss優化 AdaminOptimizer優化optimizer = tf.train.AdamOptimizer(1e-3).minimize(loss)return optimizerdef accuracy_graph(y, y_conv, width=len(CAPTCHA_LIST), height=CAPTCHA_LEN):"""偏差計算圖，正確值和預測值，計算準確度:param y: 正確值 標簽:param y_conv:  預測值:param width:   驗證碼預備字符列表長度:param height:  驗證碼的大小，默認為4:return:    正確率"""# 這里區分了大小寫 實際上驗證碼一般不區分大小寫,有四個值，不同于手寫體識別# 預測值predict = tf.reshape(y_conv, [-1, height, width])   #max_predict_idx = tf.argmax(predict, 2)# 標簽label = tf.reshape(y, [-1, height, width])max_label_idx = tf.argmax(label, 2)correct_p = tf.equal(max_predict_idx, max_label_idx)    # 判斷是否相等accuracy = tf.reduce_mean(tf.cast(correct_p, tf.float32))return accuracydef train(height=CAPTCHA_HEIGHT, width=CAPTCHA_WIDTH, y_size=len(CAPTCHA_LIST)*CAPTCHA_LEN):"""cnn訓練:param height:  驗證碼高度:param width:   驗證碼寬度:param y_size:  驗證碼預備字符列表長度*驗證碼長度（默認為4）:return:"""# cnn在圖像大小是2的倍數時性能最高, 如果圖像大小不是2的倍數，可以在圖像邊緣補無用像素# 在圖像上補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     # 預設模型準確率標準# 按照圖片大小申請占位符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 神經元使用率keep_prob = tf.placeholder(tf.float32)# cnn 模型y_conv = cnn_graph(x, keep_prob, (height, width))# 優化optimizer = optimize_graph(y, y_conv)# 計算準確率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())     # 初始化step = 0    # 步數while 1:print(step)batch_x, batch_y = get_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 % 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: 1.0})print(datetime.now().strftime('%c'), ' step:', step, ' accuracy:', acc)# 準確率滿足要求，保存模型if acc > acc_rate:model_path = "./model/captcha.model"saver.save(sess, model_path, global_step=step)acc_rate += 0.01if acc_rate > 0.99:     # 準確率達到99%則退出breakstep += 1sess.close()if __name__ == '__main__':train()

4、測試模型效果：

# -*- coding:utf-8 -*-
# name: model_test.pyimport tensorflow as tf
from model_train import cnn_graph
from captcha_gen import gen_captcha_text_and_image
from util import vec2text, convert2gray
from util import CAPTCHA_LIST, CAPTCHA_WIDTH, CAPTCHA_HEIGHT, CAPTCHA_LEN
from PIL import Imagedef captcha2text(image_list, height=CAPTCHA_HEIGHT, width=CAPTCHA_WIDTH):"""驗證碼圖片轉化為文本:param image_list::param height::param width::return:"""x = tf.placeholder(tf.float32, [None, height * width])keep_prob = tf.placeholder(tf.float32)y_conv = cnn_graph(x, keep_prob, (height, width))saver = tf.train.Saver()with tf.Session() as sess:saver.restore(sess, tf.train.latest_checkpoint('model/'))predict = tf.argmax(tf.reshape(y_conv, [-1, CAPTCHA_LEN, len(CAPTCHA_LIST)]), 2)vector_list = sess.run(predict, feed_dict={x: image_list, keep_prob: 1})vector_list = vector_list.tolist()text_list = [vec2text(vector) for vector in vector_list]return text_listif __name__ == '__main__':text, image = gen_captcha_text_and_image()img = Image.fromarray(image)image = convert2gray(image)image = image.flatten() / 255pre_text = captcha2text([image])print("驗證碼正確值:", text, ' 模型預測值:', pre_text)img.show()

?