tensorflow訓練驗證碼識別模型的樣本可以使用captcha生成，captcha在linux中的安裝也很簡單：
pip install captcha# -*- coding: utf-8 -*-
from captcha.image import ImageCaptcha # pip install captcha
import numpy as np
from PIL import Image
import random
import cv2
import os
# 驗證碼中的字符
number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
# 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']
# 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']
# 驗證碼長度為4個字符
def random_captcha_text(char_set=number, captcha_size=4):
captcha_text = []
for i in range(captcha_size):
c = random.choice(char_set)
captcha_text.append(c)
return captcha_text
# 生成字符對應的驗證碼
def gen_captcha_text_and_image():
image = ImageCaptcha()
captcha_text = random_captcha_text()
captcha_text = ''.join(captcha_text)
captcha = image.generate(captcha_text)
captcha_image = Image.open(captcha)
captcha_image = np.array(captcha_image)
return captcha_text, captcha_image
if __name__ == '__main__':
#保存路徑
path = './trainImage'
# path = './validImage'
for i in range(10000):
text, image = gen_captcha_text_and_image()
fullPath = os.path.join(path, text + ".jpg")
cv2.imwrite(fullPath, image)
print "{0}/10000".format(i)
print "/nDone!"分別生成訓練樣本和測試樣本，生成的樣本圖片如下：
使用tensorflow執行訓練：
# -*- coding: utf-8 -*-
import numpy as np
import tensorflow as tf
import cv2
import os
import random
import time
number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
# 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']
# 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']
image_filename_list = []
total = 0
def get_image_file_name(imgFilePath):
fileName = []
total = 0
for filePath in os.listdir(imgFilePath):
captcha_name = filePath.split('/')[-1]
fileName.append(captcha_name)
total += 1
return fileName, total
image_filename_list, total = get_image_file_name('./trainImage')
random.seed(time.time())
# 打亂順序
random.shuffle(image_filename_list)
def gen_captcha_text_and_image(imageFilePath, imageAmount):
num = random.randint(0, imageAmount - 1)
img = cv2.imread(os.path.join(imageFilePath, image_filename_list[num]), 0)
img = np.float32(img)
text = image_filename_list[num].split('.')[0]
return text, img
# 圖像大小
IMAGE_HEIGHT = 60
IMAGE_WIDTH = 160
MAX_CAPTCHA = 4
# 文本轉向量
char_set = number
CHAR_SET_LEN = len(char_set)
# 例如，如果驗證碼是 ‘0296’ ，則對應的標簽是
# [1 0 0 0 0 0 0 0 0 0
# 0 0 1 0 0 0 0 0 0 0
# 0 0 0 0 0 0 0 0 0 1
# 0 0 0 0 0 0 1 0 0 0]
def name2label(name):
label = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
for i, c in enumerate(name):
idx = i * CHAR_SET_LEN + ord(c) - ord('0')
label[idx] = 1
return label
# label to name
def label2name(digitalStr):
digitalList = []
for c in digitalStr:
digitalList.append(ord(c) - ord('0'))
return np.array(digitalList)
# 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):
if c == '_':
k = 62
return k
k = ord(c) - 48
if k > 9:
k = ord(c) - 55
if k > 35:
k = ord(c) - 61
if k > 61:
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(imageFilePath, batch_size=128):
batch_x = np.zeros([batch_size, IMAGE_HEIGHT * IMAGE_WIDTH])
batch_y = np.zeros([batch_size, MAX_CAPTCHA * CHAR_SET_LEN])
def wrap_gen_captcha_text_and_image(imageFilePath, imageAmount):
while True:
text, image = gen_captcha_text_and_image(imageFilePath, imageAmount)
if image.shape == (60, 160):
return text, image
for listNum in os.walk(imageFilePath):
pass
imageAmount = len(listNum[2])
for i in range(batch_size):
text, image = wrap_gen_captcha_text_and_image(imageFilePath, imageAmount)
batch_x[i, :] = image.flatten() / 255 # (image.flatten()-128)/128 mean為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])
# 3 conv layer
w_c1 = tf.Variable(w_alpha * tf.random_normal([3, 3, 1, 32]))
b_c1 = tf.Variable(b_alpha * tf.random_normal([32]))
conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv1 = tf.nn.dropout(conv1, keep_prob)
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)
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
w_d = tf.Variable(w_alpha * tf.random_normal([8 * 20 * 64, 1024]))
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():
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(logits=output, labels=Y))
# optimizer 為了加快訓練 learning_rate應該開始大，然后慢慢減小
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('./trainImage', 128)
_, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75})
print(step, loss_)
# 每100 step計算一次准確率
if step % 100 == 0:
batch_x_test, batch_y_test = get_next_batch('./validImage', 128)
acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.})
print(step, acc)
# 訓練結束條件
if acc > 0.94 or step > 3000:
saver.save(sess, "./crack_capcha.model", global_step=step)
break
step += 1
def predict_captcha(captcha_image):
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)
text_list = sess.run(predict, feed_dict={X: [captcha_image], keep_prob: 1})
text = text_list[0].tolist()
vector = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
i = 0
for n in text:
vector[i * CHAR_SET_LEN + n] = 1
i += 1
return vec2text(vector)
# 執行訓練
train_crack_captcha_cnn()
print "訓練完成，開始測試…"
time.sleep(3000)
# -------------------------------------------------------------------
大約執行1600輪迭代（batchsize=128）之后訓練完成：
訓練結果在當前目錄文件夾下生成4個文件：
測試單張驗證碼圖片：
# -*- coding: utf-8 -*-
import numpy as np
import tensorflow as tf
import cv2
import os
import random
import time
number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
# 圖像大小
IMAGE_HEIGHT = 60
IMAGE_WIDTH = 160
MAX_CAPTCHA = 4
char_set = number
CHAR_SET_LEN = len(char_set)
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])
# 3 conv layer
w_c1 = tf.Variable(w_alpha * tf.random_normal([3, 3, 1, 32]))
b_c1 = tf.Variable(b_alpha * tf.random_normal([32]))
conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
conv1 = tf.nn.dropout(conv1, keep_prob)
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)
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
w_d = tf.Variable(w_alpha * tf.random_normal([8 * 20 * 64, 1024]))
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
# 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)
def predict_captcha(captcha_image):
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)
text_list = sess.run(predict, feed_dict={X: [captcha_image], keep_prob: 1})
text = text_list[0].tolist()
vector = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
i = 0
for n in text:
vector[i * CHAR_SET_LEN + n] = 1
i += 1
return vec2text(vector)
#單張圖片預測
image = np.float32(cv2.imread('./validImage/2792.jpg', 0))
text = '2792'
image = image.flatten() / 255
predict_text = predict_captcha(image)
print("正確: {0} 預測: {1}".format(text, predict_text))由於captcha生成的驗證碼條件相對單一，使用訓練出來的模型即便只有0.94的精度也比人工識別的精度要高了。預測結果正確：
識別過程中加載測試圖片注意進行精度轉換（np.float32()）。
這里可以下載訓練好的模型文件： http://download.csdn.net/download/dcrmg/10195217
20180114補充： 訓練代碼詳細解讀
# -*- coding: utf-8 -*-
import numpy as np
import tensorflow as tf
import cv2
import os
import random
import time
#number
number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
# 圖像大小
IMAGE_HEIGHT = 60 #80
IMAGE_WIDTH = 160 #250
MAX_CAPTCHA = 8
char_set = number
CHAR_SET_LEN = len(char_set) #
image_filename_list = []
total = 0
def get_image_file_name(imgFilePath):
fileName = []
total = 0
for filePath in os.listdir(imgFilePath):
captcha_name = filePath.split('/')[-1]
fileName.append(captcha_name)
total += 1
random.seed(time.time())
# 打亂順序
random.shuffle(fileName)
return fileName, total
# 獲取訓練數據的名稱列表
image_filename_list, total = get_image_file_name('./trainImage')
# 獲取測試數據的名稱列表
image_filename_list_valid, total = get_image_file_name('./validImage')
# 讀取圖片和標簽
def gen_captcha_text_and_image(imageFilePath, image_filename_list,imageAmount):
num = random.randint(0, imageAmount - 1)
img = cv2.imread(os.path.join(imageFilePath, image_filename_list[num]), 0)
img = cv2.resize(img,(160,60))
img = np.float32(img)
text = image_filename_list[num].split('.')[0]
return text, img
# 文本轉向量
# 例如，如果驗證碼是 ‘0296’ ，則對應的標簽是
# [1 0 0 0 0 0 0 0 0 0
# 0 0 1 0 0 0 0 0 0 0
# 0 0 0 0 0 0 0 0 0 1
# 0 0 0 0 0 0 1 0 0 0]
def name2label(name):
label = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
for i, c in enumerate(name):
idx = i * CHAR_SET_LEN + ord(c) - ord('0')
label[idx] = 1
return label
# label to name
def label2name(digitalStr):
digitalList = []
for c in digitalStr:
digitalList.append(ord(c) - ord('0'))
return np.array(digitalList)
# 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):
if c == '_':
k = 62
return k
k = ord(c) - 48
if k > 9:
k = ord(c) - 55
if k > 35:
k = ord(c) - 61
if k > 61:
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(imageFilePath, image_filename_list= None,batch_size=128):
batch_x = np.zeros([batch_size, IMAGE_HEIGHT * IMAGE_WIDTH])
batch_y = np.zeros([batch_size, MAX_CAPTCHA * CHAR_SET_LEN])
def wrap_gen_captcha_text_and_image(imageFilePath, imageAmount):
while True:
text, image = gen_captcha_text_and_image(imageFilePath,image_filename_list, imageAmount)
if image.shape == (60, 160):
return text, image
for listNum in os.walk(imageFilePath):
pass
imageAmount = len(listNum[2])
for i in range(batch_size):
text, image = wrap_gen_captcha_text_and_image(imageFilePath, imageAmount)
batch_x[i, :] = image.flatten() / 255 # (image.flatten()-128)/128 mean為0
batch_y[i, :] = text2vec(text)
return batch_x, batch_y
####################################################################
# 占位符，X和Y分別是輸入訓練數據和其標簽，標簽轉換成8*10的向量
X = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT * IMAGE_WIDTH])
Y = tf.placeholder(tf.float32, [None, MAX_CAPTCHA * CHAR_SET_LEN])
# 聲明dropout占位符變量
keep_prob = tf.placeholder(tf.float32) # dropout
# 定義CNN
def crack_captcha_cnn(w_alpha=0.01, b_alpha=0.1):
# 把 X reshape 成 IMAGE_HEIGHT*IMAGE_WIDTH*1的格式,輸入的是灰度圖片，所有通道數是1;
# shape 里的-1表示數量不定，根據實際情況獲取，這里為每輪迭代輸入的圖像數量（batchsize）的大小;
x = tf.reshape(X, shape=[-1, IMAGE_HEIGHT, IMAGE_WIDTH, 1])
# 搭建第一層卷積層
# shape[3, 3, 1, 32]里前兩個參數表示卷積核尺寸大小，即patch;
# 第三個參數是圖像通道數，第四個參數是該層卷積核的數量，有多少個卷積核就會輸出多少個卷積特征圖像
w_c1 = tf.Variable(w_alpha * tf.random_normal([3, 3, 1, 32]))
# 每個卷積核都配置一個偏置量，該層有多少個輸出，就應該配置多少個偏置量
b_c1 = tf.Variable(b_alpha * tf.random_normal([32]))
# 圖片和卷積核卷積，並加上偏執量，卷積結果28x28x32
# tf.nn.conv2d() 函數實現卷積操作
# tf.nn.conv2d()中的padding用於設置卷積操作對邊緣像素的處理方式，在tf中有VALID和SAME兩種模式
# padding='SAME'會對圖像邊緣補0,完成圖像上所有像素（特別是邊緣象素）的卷積操作
# padding='VALID'會直接丟棄掉圖像邊緣上不夠卷積的像素
# strides：卷積時在圖像每一維的步長，是一個一維的向量，長度4，並且strides[0]=strides[3]=1
# tf.nn.bias_add() 函數的作用是將偏置項b_c1加到卷積結果value上去;
# 注意這里的偏置項b_c1必須是一維的，並且數量一定要與卷積結果value最后一維數量相同
# tf.nn.relu() 函數是relu激活函數，實現輸出結果的非線性轉換，即features=max(features, 0)，輸出tensor的形狀和輸入一致
conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
# tf.nn.max_pool()函數實現最大池化操作，進一步提取圖像的抽象特征，並且降低特征維度
# ksize=[1, 2, 2, 1]定義最大池化操作的核尺寸為2*2, 池化結果14x14x32 卷積結果乘以池化卷積核
conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# tf.nn.dropout是tf里為了防止或減輕過擬合而使用的函數，一般用在全連接層;
# Dropout機制就是在不同的訓練過程中根據一定概率（大小可以設置，一般情況下訓練推薦0.5）隨機扔掉（屏蔽）一部分神經元，
# 不參與本次神經網絡迭代的計算（優化）過程，權重保留但不做更新;
# tf.nn.dropout()中 keep_prob用於設置概率，需要是一個占位變量，在執行的時候具體給定數值
conv1 = tf.nn.dropout(conv1, keep_prob)
# 原圖像HEIGHT = 60 WIDTH = 160，經過神經網絡第一層卷積（圖像尺寸不變、特征×32）、池化（圖像尺寸縮小一半，特征不變）之后;
# 輸出大小為 30*80*32
# 搭建第二層卷積層
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)
# 原圖像HEIGHT = 60 WIDTH = 160，經過神經網絡第一層后輸出大小為 30*80*32
# 經過神經網絡第二層運算后輸出為 16*40*64 (30*80的圖像經過2*2的卷積核池化，padding為SAME，輸出維度是16*40)
# 搭建第三層卷積層
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)
# 原圖像HEIGHT = 60 WIDTH = 160，經過神經網絡第一層后輸出大小為 30*80*32 經過第二層后輸出為 16*40*64
# 經過神經網絡第二層運算后輸出為 16*40*64 ; 經過第三層輸出為 8*20*64，這個參數很重要，決定量后邊全連接層的維度
# 搭建全連接層
# 二維張量，第一個參數8*20*64的patch，這個參數由最后一層卷積層的輸出決定，第二個參數代表卷積個數共1024個，即輸出為1024個特征
w_d = tf.Variable(w_alpha * tf.random_normal([ 8 * 20 * 64, 1024]))
# 偏置項為1維，個數跟卷積核個數保持一致
b_d = tf.Variable(b_alpha * tf.random_normal([1024]))
# w_d.get_shape()作用是把張量w_d的形狀轉換為元組tuple的形式，w_d.get_shape().as_list()是把w_d轉為元組再轉為list形式
# w_d 的 形狀是[ 8 * 20 * 64, 1024]，w_d.get_shape().as_list()結果為 8*20*64=10240 ;
# 所以tf.reshape(conv3, [-1, w_d.get_shape().as_list()[0]])的作用是把最后一層隱藏層的輸出轉換成一維的形式
dense = tf.reshape(conv3, [-1, w_d.get_shape().as_list()[0]])
# tf.matmul(dense, w_d)函數是矩陣相乘，輸出維度是 -1*1024
dense = tf.nn.relu(tf.add(tf.matmul(dense, w_d), b_d))
dense = tf.nn.dropout(dense, keep_prob)
# 經過全連接層之后，輸出為 一維，1024個向量
# w_out定義成一個形狀為 [1024, 8 * 10] = [1024, 80]
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 的輸出為 8*10 的向量， 8代表識別結果的位數，10是每一位上可能的結果（0到9）
out = tf.add(tf.matmul(dense, w_out), b_out)
# out = tf.nn.softmax(out)
# 輸出神經網絡在當前參數下的預測值
return out
# 訓練
def train_crack_captcha_cnn():
output = crack_captcha_cnn()
# loss
# loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(output, Y))
# tf.nn.sigmoid_cross_entropy_with_logits()函數計算交叉熵,輸出的是一個向量而不是數;
# 交叉熵刻畫的是實際輸出（概率）與期望輸出（概率）的距離，也就是交叉熵的值越小，兩個概率分布就越接近
# tf.reduce_mean()函數求矩陣的均值
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=output, labels=Y))
# optimizer 為了加快訓練 learning_rate應該開始大，然后慢慢減小
# tf.train.AdamOptimizer（）函數實現了Adam算法的優化器
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('./trainImage',image_filename_list, 64)
_, loss_ = sess.run([optimizer, loss], feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.75})
print(step, loss_)
# 每100 step計算一次准確率
if step % 100 == 0:
batch_x_test, batch_y_test = get_next_batch('./vaildImage',image_filename_list_valid, 128)
acc = sess.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test, keep_prob: 1.})
print(step, acc)
# 訓練結束條件
if acc > 0.97 or step > 5500:
saver.save(sess, "./crack_capcha.model", global_step=step)
break
step += 1
def predict_captcha(captcha_image):
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)
text_list = sess.run(predict, feed_dict={X: [captcha_image], keep_prob: 1})
text = text_list[0].tolist()
vector = np.zeros(MAX_CAPTCHA * CHAR_SET_LEN)
i = 0
for n in text:
vector[i * CHAR_SET_LEN + n] = 1
i += 1
return vec2text(vector)
# 執行訓練
train_crack_captcha_cnn()
print "訓練完成，開始測試…"
# time.sleep(3000)