本文系做者原創,轉載請註明出處:http://www.javashuo.com/article/p-wfufrieu-g.html html
深度學習一個必要的前提就是須要大量的訓練樣本數據,絕不誇張的說,訓練樣本數據的多少直接決定模型的預測準確度。而本節的訓練樣本數據(驗證碼:字母和數字組成)經過調用Image模塊(圖像處理庫)中相關函數生成。python
安裝:pip install pillowgit
驗證碼生成步驟:隨機在字母和數字中選擇4個字符 -> 建立背景圖片 -> 添加噪聲 -> 字符扭曲算法
具體樣本以下所示:網絡
對於上圖的驗證碼,若是用傳統方式破解,其步驟通常是:session
圖片分割:採用分割算法分割出每個字符;app
字符識別:由分割出的每一個字符圖片,根據OCR光學字符識別出每一個字符圖片對應的字符;dom
難點在於:對於圖片字符有黏連(2個,3個,或者4個所有黏連),圖片是沒法徹底分割出來的,也就是說,即便分割出來了,字符識別基本上都是錯誤的,特別對於人眼都沒法分辨的驗證碼,用傳統的這種破解方法,成功率基本上是極其低的。ide
黏連驗證碼函數
人眼幾乎沒法分辨驗證碼
第一張是 0ymo or 0ynb ?第二張是 7e9l or 1e9l ?
對於以上傳統方法破解驗證碼的短板,咱們採用深度學習之卷積神經網絡來進行破解。
前向傳播組成:3個卷積層(3*3*1*32,3*3*32*64,3*3*64*64),3個池化層,4個dropout防過擬合層,2個全鏈接層((8*20*64,1024),(1024, MAX_CAPTCHA*CHAR_SET_LEN])),4個Relu激活函數。
反向傳播組成:計算損失(sigmoid交叉熵),計算梯度,目標預測,計算準確率,參數更新。
tensorboard生成結構圖(圖片可能不是很清楚,在圖片位置點擊鼠標右鍵->在新標籤頁面打開圖片,就能夠放縮圖片了。)
這裏特別要注意數據流的變化:
(?,60,160,1) + conv1->(?,60,160,32)+ relu ->(?,60,160,32) + pool1 ->(?,30,80,32) + dropout -> (?,30,80,32)
+ conv2->(?,30,80,64) + relu ->(?,30,80,64) + pool2 ->(?,15,40,64) + dropout -> (?,15,40,64)
+ conv3->(?,15,40,64) + relu ->(?,15,40,64) + pool3 ->(?,8,20,64) + dropout -> (?,8,20,64)
+ fc1 ->(?,1024) + relu ->(?,1024) + dropout ->(?,1024)
+ fc2 ->(?,MAX_CAPTCHA*CHAR_SET_LEN)
只要把握住一點,卷積過程跟全鏈接運算是不同的。
卷積過程:矩陣對應位置相乘再相加,要求相乘的兩個矩陣寬、高必須相同(好比大小都是m * n),獲得結果就是一個數值。
全鏈接(矩陣乘法):它要求第一個矩陣的列和第二個矩陣的行必須相同,好比矩陣A大小m * n,矩陣B大小n * k,紅色部分必須相同,獲得結果大小就是m * k。
參數保存:
tensorflow對於參數保存功能已幫咱們作好了,咱們只要直接使用就能夠了。使用也很簡單,就兩步,獲取保存對象,調用保存方法。
獲取保存對象:
saver = tf.train.Saver()
調用保存方法:
saver.save(sess, "./model/crack_capcha.model99", global_step=step)
global_step=step :在保存文件時,會統計運行了多少次。
參數使用:
獲取保存對象->獲取最後一次生成文件的路徑->導入參數到session會話中
獲取保存對象與參數保存是同樣的。
獲取最後一次生成文件的路徑:在參數保存時會生成一個checkpoint文件(個人是在model文件下),裏面會記錄最後一次生成文件的文件名。model文件
checkpoint內容
導入參數到session會話中:首先要開啓session會話,而後調用保存對象的restore方法便可。
saver.restore(sess, checkpoint.model_checkpoint_path)
1. 在session調用run方法時,必定不能遺漏某個操做結果對應的參數賦值,這表述比較繞口,咱們來看下面的例子。
_, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75})
X:輸入數據,Y:標籤數據,keep_prob:防過擬合機率因子(超參),這些參數在獲取損失函數loss,計算梯度optimizer時被用到,
在tensorflow的CNN中只是做爲佔位符處理的,因此在session調用run方法時,必定要對這些參數賦值,並用feed_dict做爲字典參數傳入,注意大小寫也要相同。
2. 在訓練前須要將文本轉爲向量,在預測判斷是否準確時須要將向量轉爲文本字符串。
這裏的樣例總長度63:數字10個(0-9),小寫字母26(a-z),大寫字母26(A-Z),'_':若是不夠4個字符,用來補齊。
向量長度範圍:字符4*(10 + 26 + 26 + 1) = 252
文本轉向量:經過某種規則(char2pos),計算字符數值,而後根據該字符在4個字符中的位置,計算向量索引
idx = i * CHAR_SET_LEN + char2pos(c)
向量轉文本:跟文本轉向量操做相反(vec2text)
在letterAndNumber.py文件中,train = 0 表示訓練,1表示預測。
在訓練時,採用的batch_size = 64,每訓練100次計算一次準確率,若是準確率大於0.8,就將參數保存到model文件中,準確率大於0.9,在保存參數的同時結束訓練。
在預測時,隨機採用100幅圖片,觀察其準確率;另外,對於以前展現的黏連驗證碼,人眼不能較好分辨的驗證碼,單獨進行識別。
letterAndNumber.py
1 import numpy as np 2 import tensorflow as tf 3 from captcha.image import ImageCaptcha 4 import numpy as np 5 import matplotlib.pyplot as plt 6 from PIL import Image 7 import random 8 9 number = ['0','1','2','3','4','5','6','7','8','9'] 10 alphabet = ['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'] 11 ALPHABET = ['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'] 12 13 def random_captcha_text(char_set=number+alphabet+ALPHABET, captcha_size=4): 14 #def random_captcha_text(char_set=number, captcha_size=4): 15 captcha_text = [] 16 for i in range(captcha_size): 17 c = random.choice(char_set) 18 captcha_text.append(c) 19 return captcha_text 20 21 22 def gen_captcha_text_and_image(i = 0): 23 # 建立圖像實例對象 24 image = ImageCaptcha() 25 # 隨機選擇4個字符 26 captcha_text = random_captcha_text() 27 # array 轉化爲 string 28 captcha_text = ''.join(captcha_text) 29 # 生成驗證碼 30 captcha = image.generate(captcha_text) 31 if i%100 == 0 : 32 image.write(captcha_text, "./generateImage/" + captcha_text + '.jpg') 33 34 captcha_image = Image.open(captcha) 35 captcha_image = np.array(captcha_image) 36 return captcha_text, captcha_image 37 38 def convert2gray(img): 39 if len(img.shape) > 2: 40 gray = np.mean(img, -1) 41 # 上面的轉法較快,正規轉法以下 42 # r, g, b = img[:,:,0], img[:,:,1], img[:,:,2] 43 # gray = 0.2989 * r + 0.5870 * g + 0.1140 * b 44 return gray 45 else: 46 return img 47 48 49 # 文本轉向量 50 def text2vec(text): 51 text_len = len(text) 52 if text_len > MAX_CAPTCHA: 53 raise ValueError('驗證碼最長4個字符') 54 55 vector = np.zeros(MAX_CAPTCHA*CHAR_SET_LEN) 56 57 def char2pos(c): 58 if c =='_': 59 k = 62 60 return k 61 k = ord(c)-48 62 if k > 9: 63 k = ord(c) - 55 64 if k > 35: 65 k = ord(c) - 61 66 if k > 61: 67 raise ValueError('No Map') 68 return k 69 70 for i, c in enumerate(text): 71 #idx = i * CHAR_SET_LEN + int(c) 72 idx = i * CHAR_SET_LEN + char2pos(c) 73 vector[idx] = 1 74 return vector 75 # 向量轉回文本 76 def vec2text(vec): 77 char_pos = vec[0] 78 text=[] 79 for i, c in enumerate(char_pos): 80 char_at_pos = i #c/63 81 char_idx = c % CHAR_SET_LEN 82 if char_idx < 10: 83 char_code = char_idx + ord('0') 84 elif char_idx <36: 85 char_code = char_idx - 10 + ord('A') 86 elif char_idx < 62: 87 char_code = char_idx- 36 + ord('a') 88 elif char_idx == 62: 89 char_code = ord('_') 90 else: 91 raise ValueError('error') 92 text.append(chr(char_code)) 93 """ 94 text=[] 95 char_pos = vec.nonzero()[0] 96 for i, c in enumerate(char_pos): 97 number = i % 10 98 text.append(str(number)) 99 """ 100 return "".join(text) 101 102 """ 103 #向量(大小MAX_CAPTCHA*CHAR_SET_LEN)用0,1編碼 每63個編碼一個字符,這樣順利有,字符也有 104 vec = text2vec("F5Sd") 105 text = vec2text(vec) 106 print(text) # F5Sd 107 vec = text2vec("SFd5") 108 text = vec2text(vec) 109 print(text) # SFd5 110 """ 111 112 # 生成一個訓練batch 113 def get_next_batch(batch_size=128): 114 batch_x = np.zeros([batch_size, IMAGE_HEIGHT*IMAGE_WIDTH]) 115 batch_y = np.zeros([batch_size, MAX_CAPTCHA*CHAR_SET_LEN]) 116 117 # 有時生成圖像大小不是(60, 160, 3) 118 def wrap_gen_captcha_text_and_image(i): 119 while True: 120 text, image = gen_captcha_text_and_image(i) 121 if image.shape == (60, 160, 3): 122 return text, image 123 124 for i in range(batch_size): 125 text, image = wrap_gen_captcha_text_and_image(i) 126 image = convert2gray(image) 127 128 batch_x[i,:] = image.flatten() / 255 # (image.flatten()-128)/128 mean爲0 129 batch_y[i,:] = text2vec(text) 130 131 return batch_x, batch_y 132 133 134 135 # 定義CNN 136 def crack_captcha_cnn(w_alpha=0.01, b_alpha=0.1): 137 x = tf.reshape(X, shape=[-1, IMAGE_HEIGHT, IMAGE_WIDTH, 1]) 138 139 #w_c1_alpha = np.sqrt(2.0/(IMAGE_HEIGHT*IMAGE_WIDTH)) # 140 #w_c2_alpha = np.sqrt(2.0/(3*3*32)) 141 #w_c3_alpha = np.sqrt(2.0/(3*3*64)) 142 #w_d1_alpha = np.sqrt(2.0/(8*32*64)) 143 #out_alpha = np.sqrt(2.0/1024) 144 145 # 3 conv layer 146 w_c1 = tf.Variable(w_alpha*tf.random_normal([3, 3, 1, 32])) 147 b_c1 = tf.Variable(b_alpha*tf.random_normal([32])) 148 # 卷積 + Relu激活函數 149 conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1)) 150 # 池化 151 conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 152 # dropout 防止過擬合 153 conv1 = tf.nn.dropout(conv1, rate = 1 - keep_prob) 154 155 w_c2 = tf.Variable(w_alpha*tf.random_normal([3, 3, 32, 64])) 156 b_c2 = tf.Variable(b_alpha*tf.random_normal([64])) 157 # 卷積 + Relu激活函數 158 conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv1, w_c2, strides=[1, 1, 1, 1], padding='SAME'), b_c2)) 159 # 池化 160 conv2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 161 # dropout 防止過擬合 162 conv2 = tf.nn.dropout(conv2, rate = 1 - keep_prob) 163 164 w_c3 = tf.Variable(w_alpha*tf.random_normal([3, 3, 64, 64])) 165 b_c3 = tf.Variable(b_alpha*tf.random_normal([64])) 166 # 卷積 + Relu激活函數 167 conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv2, w_c3, strides=[1, 1, 1, 1], padding='SAME'), b_c3)) 168 # 池化 169 conv3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 170 # dropout 防止過擬合 171 conv3 = tf.nn.dropout(conv3, rate = 1 - keep_prob) 172 173 # Fully connected layer 174 w_d = tf.Variable(w_alpha*tf.random_normal([8*20*64, 1024])) 175 b_d = tf.Variable(b_alpha*tf.random_normal([1024])) 176 dense = tf.reshape(conv3, [-1, w_d.get_shape().as_list()[0]]) 177 # 全鏈接 + Relu 178 dense = tf.nn.relu(tf.add(tf.matmul(dense, w_d), b_d)) 179 dense = tf.nn.dropout(dense, rate = 1 - keep_prob) 180 181 w_out = tf.Variable(w_alpha*tf.random_normal([1024, MAX_CAPTCHA*CHAR_SET_LEN])) 182 b_out = tf.Variable(b_alpha*tf.random_normal([MAX_CAPTCHA*CHAR_SET_LEN])) 183 # 全鏈接 184 out = tf.add(tf.matmul(dense, w_out), b_out) 185 return out 186 187 # 訓練 188 def train_crack_captcha_cnn(): 189 output = crack_captcha_cnn() 190 # 計算損失 191 loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits= output, labels= Y)) 192 # 計算梯度 193 optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss) 194 # 目標預測 195 predict = tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN]) 196 # 目標預測最大值 197 max_idx_p = tf.argmax(predict, 2) 198 # 真實標籤最大值 199 max_idx_l = tf.argmax(tf.reshape(Y, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2) 200 correct_pred = tf.equal(max_idx_p, max_idx_l) 201 # 準確率 202 accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32)) 203 204 saver = tf.train.Saver() 205 with tf.Session() as sess: 206 # 打印tensorboard流程圖 207 tf.summary.FileWriter("./tensorboard/", sess.graph) 208 sess.run(tf.global_variables_initializer()) 209 210 step = 0 211 while True: 212 batch_x, batch_y = get_next_batch(64) 213 _, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75}) 214 print(step, loss_) 215 216 # 每100 step計算一次準確率 217 if step % 100 == 0: 218 batch_x_test, batch_y_test = get_next_batch(100) 219 acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.}) 220 print(step, acc) 221 # 若是準確率大於80%,保存模型,完成訓練 222 if acc > 0.90: 223 saver.save(sess, "./model/crack_capcha.model99", global_step=step) 224 break 225 if acc > 0.80: 226 saver.save(sess, "./model/crack_capcha.model88", global_step=step) 227 228 step += 1 229 def crack_captcha(captcha_image, output): 230 231 saver = tf.train.Saver() 232 233 with tf.Session() as sess: 234 sess.run(tf.initialize_all_variables()) 235 # 獲取訓練後的參數 236 checkpoint = tf.train.get_checkpoint_state("model") 237 if checkpoint and checkpoint.model_checkpoint_path: 238 saver.restore(sess, checkpoint.model_checkpoint_path) 239 print("Successfully loaded:", checkpoint.model_checkpoint_path) 240 else: 241 print("Could not find old network weights") 242 243 predict = tf.argmax(tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2) 244 text_list = sess.run(predict, feed_dict={X: [captcha_image], keep_prob: 1}) 245 #text = text_list[0].tolist() 246 text = vec2text(text_list) 247 return text 248 if __name__ == '__main__': 249 train = 0 # 0: 訓練 1: 預測 250 if train == 0: 251 number = ['0','1','2','3','4','5','6','7','8','9'] 252 alphabet = ['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'] 253 ALPHABET = ['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'] 254 255 text, image = gen_captcha_text_and_image() 256 print("驗證碼圖像channel:", image.shape) # (60, 160, 3) 257 # 圖像大小 258 IMAGE_HEIGHT = 60 259 IMAGE_WIDTH = 160 260 MAX_CAPTCHA = len(text) 261 print("驗證碼文本最長字符數", MAX_CAPTCHA) 262 # 文本轉向量 263 char_set = number + alphabet + ALPHABET + ['_'] # 若是驗證碼長度小於4, '_'用來補齊 264 #char_set = number 265 CHAR_SET_LEN = len(char_set) 266 # placeholder佔位符,做用域:整個頁面,不須要聲明時初始化 267 X = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT*IMAGE_WIDTH]) 268 Y = tf.placeholder(tf.float32, [None, MAX_CAPTCHA*CHAR_SET_LEN]) 269 keep_prob = tf.placeholder(tf.float32) # dropout 270 271 train_crack_captcha_cnn() 272 # 預測時須要將訓練的變量初始化,且只能初始化一次。 273 if train == 1: 274 # 天然計數 275 step = 0 276 # 正確預測計數 277 rightCnt = 0 278 # 設置測試次數 279 count = 100 280 number = ['0','1','2','3','4','5','6','7','8','9'] 281 alphabet = ['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'] 282 ALPHABET = ['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'] 283 284 IMAGE_HEIGHT = 60 285 IMAGE_WIDTH = 160 286 287 char_set = number + alphabet + ALPHABET + ['_'] 288 CHAR_SET_LEN = len(char_set) 289 MAX_CAPTCHA = 4 # len(text) 290 # placeholder佔位符,做用域:整個頁面,不須要聲明時初始化 291 X = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT*IMAGE_WIDTH]) 292 Y = tf.placeholder(tf.float32, [None, MAX_CAPTCHA*CHAR_SET_LEN]) 293 keep_prob = tf.placeholder(tf.float32) # dropout 294 output = crack_captcha_cnn() 295 296 saver = tf.train.Saver() 297 with tf.Session() as sess: 298 sess.run(tf.global_variables_initializer()) 299 # 獲取訓練後參數路徑 300 checkpoint = tf.train.get_checkpoint_state("model") 301 if checkpoint and checkpoint.model_checkpoint_path: 302 saver.restore(sess, checkpoint.model_checkpoint_path) 303 print("Successfully loaded:", checkpoint.model_checkpoint_path) 304 else: 305 print("Could not find old network weights.") 306 307 while True: 308 # image = Image.open("D:/Project/python/myProject/CNN/tensorflow/captchaIdentify/11/0sHB.jpg") 309 # image = np.array(image) 310 # text = '0sHB' 311 text, image = gen_captcha_text_and_image() 312 # f = plt.figure() 313 # ax = f.add_subplot(111) 314 # ax.text(0.1, 0.9,text, ha='center', va='center', transform=ax.transAxes) 315 # plt.imshow(image) 316 # 317 # plt.show() 318 319 image = convert2gray(image) 320 image = image.flatten() / 255 321 predict = tf.math.argmax(tf.reshape(output, [-1, MAX_CAPTCHA, CHAR_SET_LEN]), 2) 322 text_list = sess.run(predict, feed_dict= { X: [image], keep_prob : 1}) 323 predict_text = vec2text(text_list) 324 predict_text = crack_captcha(image, output) 325 # predict_text_list = [str(x) for x in predict_text] 326 # predict_text_new = ''.join(predict_text_list) 327 print("step:{} 真實值: {} 預測: {} 預測結果: {}".format(str(step), text, predict_text, "正確" if text.lower()==predict_text.lower() else "錯誤")) 328 if text.lower()==predict_text.lower(): 329 rightCnt += 1 330 if step == count - 1: 331 print("測試總數: {} 測試準確率: {}".format(str(count), str(rightCnt/count))) 332 break 333 step += 1 334 335 336 337
隨機採用100幅圖片,運行結果以下:
黏連驗證碼
運行結果
人眼較難識別驗證碼
運行結果
結果分析:隨機選取100張驗證碼測試,準確率有73%,這個準確率在同類型的驗證碼中已經比較可觀了。固然,能夠在訓練時將測試準確率繼續提升,好比0.95或更高,這樣,在預測時的準確率應該還會有提高的,你們有興趣的話能夠試試。
不要讓懶惰佔據你的大腦,不要讓妥協拖垮了你的人生。青春就是一張票,能不能遇上時代的快車,你的步伐就掌握在你的腳下。