（附完整python源码）基于tensorflow、opencv的入门案例_发票识别三：发票数据集制作和cnn网络训练

 

1 制作数据集合

1.1 在word上输入一行数字，我用的是Calibri字体，已经比较接近发票数字了。网友们可以自行定义字体。

 

1.2 读入图片为灰度图，threshold化；并将图片颜色反转：字体为白背景为黑；对图像进行各种类型的膨胀，多样化数据。

 

1. # encoding: utf-8

2. import cv2

3. import numpy as np

4. import os

5.  

6. img0 = cv2.imread("./number.jpg",0)

7. _, img0 = cv2.threshold(img0, 100, 255, cv2.THRESH_BINARY)

8. img0 = 255 - img0 #反转：文字置为白色，背景置为黑色

9. #进行三种尺度的膨胀

10. element1 = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 1))

11. element2 = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 3))

12. element3 = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))

13.  

14. img1 = cv2.dilate(img0, element1, iterations = 1)

15. img2 = cv2.dilate(img0, element2, iterations = 1)

16. img3 = cv2.dilate(img0, element3, iterations = 1)

17. #保存原始图和膨胀图，共四张

18. img = [img0,img1,img2,img3]

 

1.3 计算垂直直方图，用投影法分割出各个数字。

1. #对img0、img1、img2、img3进行分割，并保存分割后的图片

2. img_seg = []

3. for img_ in img:

4. img_seg.append(seg_num(img_))

编写的函数如下：

1. def cal_hist(img_,flag = 0):

2. return np.sum(img_,axis = flag)

3.  

4. class Find_num_region:

5. def __init__(self,img,hist):

6. self.cursor = -1

7. self.hist = hist

8. self.img = img

9. def next(self):

10. #将游标的位置前移一步，并返回所在检索位的矩形框

11. self.cursor = self.cursor+1

12. return self.hist[self.cursor]

13. def hasNext(self):

14. #判断是否已经检查完了所有矩形框

15. return len(self.hist)> self.cursor + 1

16. def find_start(self):

17. while(self.hasNext()):

18. hist_num = self.next()

19. if hist_num > 0:

20. return self.cursor

21. def find_end(self):

22. while(self.hasNext()):

23. hist_num = self.next()

24. if hist_num == 0:

25. return self.cursor

26. def get_num_region(self,flag = 0):

27. start = self.find_start()

28. end = self.find_end()

29. if flag == 0:

30. return self.img.copy()[:,start:end]

31. else:

32. return self.img.copy()[start:end,:]

33.  

34. def seg_num(img_):

35. hist_1 = cal_hist(img_,1)

36. find_num_ = Find_num_region(img_,hist_1)

37. img = find_num_.get_num_region(1)

38. hist_2 = cal_hist(img_)

39. find_num = Find_num_region(img_,hist_2)

40. img_number = []

41. for i in range(10):

42. print i

43. img_number.append( find_num.get_num_region() )

44. return img_number

 

1.4 图片的size设置为（28,28）对图片进行，进行缩放、旋转（仿射变换），然后增加随机噪声。

1.5 将数据保存npy格式，共10000组，每个数字满足“粗细、旋转角度、缩放比例、噪声分布”的多样化。数据集制作完毕。（数据量大的话，建议使用tfrecord格式）

1. img_arr = np.zeros((10000,28,28))

2. label_arr = np.zeros((10000,10))

3. for num in range(250): # 随机角度、缩放、噪声，200次

4.     for i in range(10): #“0-9”共10个数

5.         for j in range(4): #不同膨胀比

6.             angle_ = np.random.uniform(-20,20)

7.             scale_ = np.random.uniform(0.9,1.2)

8.             img_c = change_(img_seg[j][i],angle_,scale_)

9.             noise_num = np.random.randint(8,38)#噪声点个数

10.             img_c = add_noise(img_c,noise_num)

11.             img_arr[40*num+4*i+j] = img_c

12.             label_arr[40*num+4*i+j][i] = 1

13. np.save("img.npy",np.array(img_arr))

14. np.save("label.npy",np.array(label_arr))

15.  

编写的函数如下：

1. #将图像设置为28×28,不拉伸，全0填充

2. def change_(img,angle_,scale_):

3. length = 28

4. h,w = img.shape

5. H = np.float32([[1,0,(length-w)/2],[0,1,(length-h)/2]])

6. img = cv2.warpAffine(img,H,(length,length))

7. M = cv2.getRotationMatrix2D((length/2,length/2),angle_,scale_)

8. return cv2.warpAffine(img,M,(length,length))

9. def add_noise(img,amout):

10.     length = 28

11.     for i in range(amout):

12.         rand_ = int(np.random.rand()*length*length)

13.         row = int(rand_/length)

14.         col = int(rand_%length)

15.         if img[row,col] == 0:

16.             img[row,col] = 255

17.         else:

18.             img[row,col] = 0

19.     return img

 

2 cnn训练

2.1  使用网络上的一个常见数字识别网络结构：2层卷积，2层全连接。

1. #encoding:utf-8

2. import tensorflow as tf

3. import numpy as np

4. def weight_variable(shape):

5. initial = tf.truncated_normal(shape, stddev=0.1)

6. return tf.Variable(initial)

7. def bias_variable(shape):

8. initial = tf.constant(0.1, shape=shape)

9. return tf.Variable(initial)

10. def conv2d(x, W):

11. return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

12. def pool(x):

13. return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

14. def inference(x_,keep_prob):

15. with tf.variable_scope("layer_conv1"): #卷积

16. W_conv1 = weight_variable([5, 5, 1, 32])

17. b_conv1 = bias_variable([32])

18. layer_conv1 = tf.nn.relu(conv2d(x_, W_conv1) + b_conv1)

19. layer_pool1 = pool(layer_conv1)

20. with tf.variable_scope("layer_conv2"):#卷积

21. W_conv2 = weight_variable([5, 5, 32, 64])

22. b_conv2 = bias_variable([64])

23. layer_conv2 = tf.nn.relu(conv2d(layer_pool1, W_conv2) + b_conv2)

24. layer_pool2 = pool(layer_conv2)

25. with tf.variable_scope("layer_fc3"): #全连接

26. W_fc3 = weight_variable([7 * 7 * 64, 1024])

27. b_fc3 = bias_variable([1024])

28. reshape_pool3 = tf.reshape(layer_pool2, [-1, 7*7*64])

29. layer_fc3 = tf.nn.relu(tf.matmul(reshape_pool3, W_fc3) + b_fc3)

30. with tf.variable_scope("dropout4"):

31. fc_drop4 = tf.nn.dropout(layer_fc3, keep_prob)

32. with tf.variable_scope("layer_fc5"): #全连接

33. W_fc5 = weight_variable([1024, 10])

34. b_fc5 = bias_variable([10])

35. predict_ = tf.nn.softmax(tf.matmul(fc_drop4, W_fc5) + b_fc5)

36. return predict_

37. def train(x_train,y_train,x_test,y_test):

38. batch_size = 230 #每个批次的大小

39. all_size = y_train.shape[0] #train集的大小

40. x_ = tf.placeholder("float", shape=[None,28,28,1],name='x_input')

41. y_ = tf.placeholder("float", shape=[None,10],name='y_input')

42. keep_prob = tf.placeholder("float")

43. predict_ = inference(x_,keep_prob)

44. cross_entropy = -tf.reduce_sum(y_*tf.log(predict_))

45. train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

46. #测试准确率

47. correct_prediction = tf.equal(tf.argmax(predict_,1),tf.argmax(y_,1))

48. accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

49. saver = tf.train.Saver()

50. with tf.Session() as sess:

51. sess.run(tf.initialize_all_variables())

52. for i in range(500):

53. start = (i*batch_size)%(all_size-1)

54. if start <(all_size-batch_size):

55. end1 = start+batch_size

56. end2 = 0

57. else:

58. end1 = (all_size-1)

59. end2 = start+batch_size-(all_size-1)

60. in_x = np.concatenate((x_train[start:end1],x_train[0:end2]),axis=0)

61. in_y = np.concatenate((y_train[start:end1],y_train[0:end2]),axis=0)

62. sess.run(train_step, feed_dict={x_:in_x,y_:in_y,keep_prob:0.5})

63. if i%50 == 0:

64. print "第",i,"次迭代："

65. print "test集精度：", sess.run(accuracy, feed_dict={x_:x_test,y_:y_test,keep_prob:1.0})

66. print "train集精度：",sess.run(accuracy, feed_dict={x_:in_x,y_:in_y,keep_prob:1.0})

67. saver.save(sess,"./Model/model.ckpt") #

68. def main():

69. x_input = np.load("img.npy") #读入图片

70. y_input = np.load("label.npy") #读入标签

71. x_input = x_input/255.0 #原二值化图像的像素值分别为“0”,“255”。将“255”置为“1”

72. x_train,y_train,x_test,y_test = x_input[0:8000,:],y_input[0:8000,:],x_input[8000:10000,:],y_input[8000:10000,:]

73. train(x_train,y_train,x_test,y_test)

74. if __name__ == "__main__":

75. main()

2.2 模型很快收敛，精度100%。