Tensorflow驗證碼識別應用

簡單的Tensorflow驗證碼識別應用，供大家參考，具體內容如下

1.Tensorflow的安裝方式簡單,在此就不贅述了.

2.訓練集訓練集以及測試及如下(純手工打造,所以數量不多):

3.實現代碼部分(參考了網上的一些實現來完成的)

main.py(主要的神經網絡代碼)

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238

from gen_check_code import gen_captcha_text_and_image_new,gen_captcha_text_and_image from gen_check_code import number from test_check_code import get_test_captcha_text_and_image import numpy as np import tensorflow as tf   text, image = gen_captcha_text_and_image_new() print ( "驗證碼圖像channel:" , image.shape) # (60, 160, 3) # 圖像大小 IMAGE_HEIGHT = image.shape[ 0 ] IMAGE_WIDTH = image.shape[ 1 ] image_shape = image.shape MAX_CAPTCHA = len (text) print ( "驗證碼文本最長字符數" , MAX_CAPTCHA) # 驗證碼最長4字符; 我全部固定爲4,可以不固定. 如果驗證碼長度小於4，用'_'補齊     # 把彩色圖像轉爲灰度圖像（色彩對識別驗證碼沒有什麼用） # 度化是將三分量轉化成一樣數值的過程 def convert2gray(img):   if len (img.shape) > 2 :    gray = np.mean(img, - 1 )    # 上面的轉法較快，正規轉法如下    # r, g, b = img[:,:,0], img[:,:,1], img[:,:,2]    # gray = 0.2989 * r + 0.5870 * g + 0.1140 * b    # int gray = (int) (0.3 * r + 0.59 * g + 0.11 * b);    return gray   else :    return img     """ cnn在圖像大小是2的倍數時性能最高, 如果你用的圖像大小不是2的倍數，可以在圖像邊緣補無用像素。 np.pad(image,((2,3),(2,2)), 'constant', constant_values=(255,)) # 在圖像上補2行，下補3行，左補2行，右補2行 """     char_set = number # 如果驗證碼長度小於4, '_'用來補齊 CHAR_SET_LEN = len (char_set)   # 文本轉向量 def text2vec(text):   text_len = len (text)   if text_len > MAX_CAPTCHA:    raise ValueError( '驗證碼最長4個字符' )     vector = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)     def char2pos(c):    try :     k = ord (c) - ord ( '0' )    except :     raise ValueError( 'No Map' )    return k     for i, c in enumerate (text):    idx = i * CHAR_SET_LEN + char2pos(c)    vector[idx] = 1   return vector     # 向量轉回文本 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)     # 生成一個訓練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_new()       if image.shape = = image_shape:      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爲0    batch_y[i, :] = text2vec(text)     return batch_x, batch_y     ####################################################################   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     # 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 ])     # 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)     # 定義三層的卷積神經網絡     # 定義第一層的卷積神經網絡   # 定義第一層權重   w_c1 = tf.Variable(w_alpha * tf.random_normal([ 3 , 3 , 1 , 32 ]))   # 定義第一層的偏置   b_c1 = tf.Variable(b_alpha * tf.random_normal([ 32 ]))   # 定義第一層的激勵函數   conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides = [ 1 , 1 , 1 , 1 ], padding = 'SAME' ), b_c1))   # conv1 爲輸入 ksize 表示使用2*2池化，即將2*2的色塊轉化成1*1的色塊   conv1 = tf.nn.max_pool(conv1, ksize = [ 1 , 2 , 2 , 1 ], strides = [ 1 , 2 , 2 , 1 ], padding = 'SAME' )   # dropout防止過擬合。   conv1 = tf.nn.dropout(conv1, keep_prob)     # 定義第二層的卷積神經網絡   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)     # 定義第三層的卷積神經網絡   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   # 隨機生成權重   w_d = tf.Variable(w_alpha * tf.random_normal([ 1536 , 1024 ]))   # 隨機生成偏置   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)   # out = tf.nn.softmax(out)   return out     # 訓練 def train_crack_captcha_cnn():   # 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   output = crack_captcha_cnn()   # loss   # loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(output, Y))   loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(output, Y))   # 最後一層用來分類的softmax和sigmoid有什麼不同？   # optimizer 爲了加快訓練 learning_rate應該開始大，然後慢慢衰   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 )      _, loss_ = sess.run([optimizer, loss], feed_dict = {X: batch_x, Y: batch_y, keep_prob: 0.75 })      print (step, loss_)        # 每100 step計算一次準確率      if step % 100 = = 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, acc)       # 如果準確率大於50%,保存模型,完成訓練       if acc > 0.99 :        saver.save(sess, "./crack_capcha.model" , global_step = step)        break      step + = 1   ## 訓練(如果要訓練則去掉下面一行的註釋) train_crack_captcha_cnn()     def crack_captcha():   output = crack_captcha_cnn()     saver = tf.train.Saver()   with tf.Session() as sess:    saver.restore(sess, tf.train.latest_checkpoint( '.' ))      predict = tf.argmax(tf.reshape(output, [ - 1 , MAX_CAPTCHA, CHAR_SET_LEN]), 2 )    count = 0    # 因爲測試集共40個...寫的很草率    for i in range ( 40 ):     text, image = get_test_captcha_text_and_image(i)     image = convert2gray(image)     captcha_image = image.flatten() / 255     text_list = sess.run(predict, feed_dict = {X: [captcha_image], keep_prob: 1 })     predict_text = text_list[ 0 ].tolist()     predict_text = str (predict_text)     predict_text = predict_text.replace( "[" , " ").replace(" ] ", " ").replace(" , ", " ").replace(" "," ")     if text = = predict_text:      count + = 1      check_result = "，預測結果正確"     else :      check_result = "，預測結果不正確"      print ( "正確: {} 預測: {}" . format (text, predict_text) + check_result)      print ( "正確率:" + str (count) + "/40" ) # 測試(如果要測試則去掉下面一行的註釋) # crack_captcha()

gen_check_code.py(得到訓練集輸入,需要注意修改root_dir爲訓練集的輸入文件夾,下同)

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

from