TensorFlow车牌识别完整版(含车牌数据集)

 

https://blog.youkuaiyun.com/ShadowN1ght/article/details/78571187

在之前发布的一篇博文《MNIST数据集实现车牌识别--初步演示版》中,我们演示了如何使用TensorFlow进行车牌识别,但是,当时采用的数据集是MNIST数字手写体,只能分类0-9共10个数字,无法分类省份简称和字母,局限性较大,无实际意义。

经过图像定位分割处理,博主收集了相关省份简称和26个字母的图片数据集,结合前述博文中贴出的python+TensorFlow代码,实现了完整的车牌识别功能。本着分享精神,在此送上全部代码和车牌数据集。

车牌数据集下载地址(约4000张图片):https://pan.baidu.com/s/1RyoMbHtLUlsMDsvLBCLZ2w

省份简称训练+识别代码(保存文件名为train-license-province.py)(拷贝代码请务必注意python文本缩进,只要有一处缩进错误,就无法得到正确结果,或者出现异常):

 

 
  1. #!/usr/bin/python3.5

  2. # -*- coding: utf-8 -*-

  3.  
  4. import sys

  5. import os

  6. import time

  7. import random

  8.  
  9. import numpy as np

  10. import tensorflow as tf

  11.  
  12. from PIL import Image

  13.  
  14.  
  15. SIZE = 1280

  16. WIDTH = 32

  17. HEIGHT = 40

  18. NUM_CLASSES = 6

  19. iterations = 300

  20.  
  21. SAVER_DIR = "train-saver/province/"

  22.  
  23. PROVINCES = ("京","闽","粤","苏","沪","浙")

  24. nProvinceIndex = 0

  25.  
  26. time_begin = time.time()

  27.  
  28.  
  29. # 定义输入节点,对应于图片像素值矩阵集合和图片标签(即所代表的数字)

  30. x = tf.placeholder(tf.float32, shape=[None, SIZE])

  31. y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASSES])

  32.  
  33. x_image = tf.reshape(x, [-1, WIDTH, HEIGHT, 1])

  34.  
  35.  
  36. # 定义卷积函数

  37. def conv_layer(inputs, W, b, conv_strides, kernel_size, pool_strides, padding):

  38. L1_conv = tf.nn.conv2d(inputs, W, strides=conv_strides, padding=padding)

  39. L1_relu = tf.nn.relu(L1_conv + b)

  40. return tf.nn.max_pool(L1_relu, ksize=kernel_size, strides=pool_strides, padding='SAME')

  41.  
  42. # 定义全连接层函数

  43. def full_connect(inputs, W, b):

  44. return tf.nn.relu(tf.matmul(inputs, W) + b)

  45.  
  46.  
  47. if __name__ =='__main__' and sys.argv[1]=='train':

  48. # 第一次遍历图片目录是为了获取图片总数

  49. input_count = 0

  50. for i in range(0,NUM_CLASSES):

  51. dir = './train_images/training-set/chinese-characters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  52. for rt, dirs, files in os.walk(dir):

  53. for filename in files:

  54. input_count += 1

  55.  
  56. # 定义对应维数和各维长度的数组

  57. input_images = np.array([[0]*SIZE for i in range(input_count)])

  58. input_labels = np.array([[0]*NUM_CLASSES for i in range(input_count)])

  59.  
  60. # 第二次遍历图片目录是为了生成图片数据和标签

  61. index = 0

  62. for i in range(0,NUM_CLASSES):

  63. dir = './train_images/training-set/chinese-characters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  64. for rt, dirs, files in os.walk(dir):

  65. for filename in files:

  66. filename = dir + filename

  67. img = Image.open(filename)

  68. width = img.size[0]

  69. height = img.size[1]

  70. for h in range(0, height):

  71. for w in range(0, width):

  72. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  73. if img.getpixel((w, h)) > 230:

  74. input_images[index][w+h*width] = 0

  75. else:

  76. input_images[index][w+h*width] = 1

  77. input_labels[index][i] = 1

  78. index += 1

  79.  
  80. # 第一次遍历图片目录是为了获取图片总数

  81. val_count = 0

  82. for i in range(0,NUM_CLASSES):

  83. dir = './train_images/validation-set/chinese-characters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  84. for rt, dirs, files in os.walk(dir):

  85. for filename in files:

  86. val_count += 1

  87.  
  88. # 定义对应维数和各维长度的数组

  89. val_images = np.array([[0]*SIZE for i in range(val_count)])

  90. val_labels = np.array([[0]*NUM_CLASSES for i in range(val_count)])

  91.  
  92. # 第二次遍历图片目录是为了生成图片数据和标签

  93. index = 0

  94. for i in range(0,NUM_CLASSES):

  95. dir = './train_images/validation-set/chinese-characters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  96. for rt, dirs, files in os.walk(dir):

  97. for filename in files:

  98. filename = dir + filename

  99. img = Image.open(filename)

  100. width = img.size[0]

  101. height = img.size[1]

  102. for h in range(0, height):

  103. for w in range(0, width):

  104. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  105. if img.getpixel((w, h)) > 230:

  106. val_images[index][w+h*width] = 0

  107. else:

  108. val_images[index][w+h*width] = 1

  109. val_labels[index][i] = 1

  110. index += 1

  111.  
  112. with tf.Session() as sess:

  113. # 第一个卷积层

  114. W_conv1 = tf.Variable(tf.truncated_normal([8, 8, 1, 16], stddev=0.1), name="W_conv1")

  115. b_conv1 = tf.Variable(tf.constant(0.1, shape=[16]), name="b_conv1")

  116. conv_strides = [1, 1, 1, 1]

  117. kernel_size = [1, 2, 2, 1]

  118. pool_strides = [1, 2, 2, 1]

  119. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  120.  
  121. # 第二个卷积层

  122. W_conv2 = tf.Variable(tf.truncated_normal([5, 5, 16, 32], stddev=0.1), name="W_conv2")

  123. b_conv2 = tf.Variable(tf.constant(0.1, shape=[32]), name="b_conv2")

  124. conv_strides = [1, 1, 1, 1]

  125. kernel_size = [1, 1, 1, 1]

  126. pool_strides = [1, 1, 1, 1]

  127. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  128.  
  129.  
  130. # 全连接层

  131. W_fc1 = tf.Variable(tf.truncated_normal([16 * 20 * 32, 512], stddev=0.1), name="W_fc1")

  132. b_fc1 = tf.Variable(tf.constant(0.1, shape=[512]), name="b_fc1")

  133. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  134. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  135.  
  136.  
  137. # dropout

  138. keep_prob = tf.placeholder(tf.float32)

  139.  
  140. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  141.  
  142.  
  143. # readout层

  144. W_fc2 = tf.Variable(tf.truncated_normal([512, NUM_CLASSES], stddev=0.1), name="W_fc2")

  145. b_fc2 = tf.Variable(tf.constant(0.1, shape=[NUM_CLASSES]), name="b_fc2")

  146.  
  147. # 定义优化器和训练op

  148. y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

  149. cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

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

  151.  
  152. correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))

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

  154.  
  155. # 初始化saver

  156. saver = tf.train.Saver()

  157.  
  158. sess.run(tf.global_variables_initializer())

  159.  
  160. time_elapsed = time.time() - time_begin

  161. print("读取图片文件耗费时间:%d秒" % time_elapsed)

  162. time_begin = time.time()

  163.  
  164. print ("一共读取了 %s 个训练图像, %s 个标签" % (input_count, input_count))

  165.  
  166. # 设置每次训练op的输入个数和迭代次数,这里为了支持任意图片总数,定义了一个余数remainder,譬如,如果每次训练op的输入个数为60,图片总数为150张,则前面两次各输入60张,最后一次输入30张(余数30)

  167. batch_size = 60

  168. iterations = iterations

  169. batches_count = int(input_count / batch_size)

  170. remainder = input_count % batch_size

  171. print ("训练数据集分成 %s 批, 前面每批 %s 个数据,最后一批 %s 个数据" % (batches_count+1, batch_size, remainder))

  172.  
  173. # 执行训练迭代

  174. for it in range(iterations):

  175. # 这里的关键是要把输入数组转为np.array

  176. for n in range(batches_count):

  177. train_step.run(feed_dict={x: input_images[n*batch_size:(n+1)*batch_size], y_: input_labels[n*batch_size:(n+1)*batch_size], keep_prob: 0.5})

  178. if remainder > 0:

  179. start_index = batches_count * batch_size;

  180. train_step.run(feed_dict={x: input_images[start_index:input_count-1], y_: input_labels[start_index:input_count-1], keep_prob: 0.5})

  181.  
  182. # 每完成五次迭代,判断准确度是否已达到100%,达到则退出迭代循环

  183. iterate_accuracy = 0

  184. if it%5 == 0:

  185. iterate_accuracy = accuracy.eval(feed_dict={x: val_images, y_: val_labels, keep_prob: 1.0})

  186. print ('第 %d 次训练迭代: 准确率 %0.5f%%' % (it, iterate_accuracy*100))

  187. if iterate_accuracy >= 0.9999 and it >= 150:

  188. break;

  189.  
  190. print ('完成训练!')

  191. time_elapsed = time.time() - time_begin

  192. print ("训练耗费时间:%d秒" % time_elapsed)

  193. time_begin = time.time()

  194.  
  195. # 保存训练结果

  196. if not os.path.exists(SAVER_DIR):

  197. print ('不存在训练数据保存目录,现在创建保存目录')

  198. os.makedirs(SAVER_DIR)

  199. saver_path = saver.save(sess, "%smodel.ckpt"%(SAVER_DIR))

  200.  
  201.  
  202.  
  203. if __name__ =='__main__' and sys.argv[1]=='predict':

  204. saver = tf.train.import_meta_graph("%smodel.ckpt.meta"%(SAVER_DIR))

  205. with tf.Session() as sess:

  206. model_file=tf.train.latest_checkpoint(SAVER_DIR)

  207. saver.restore(sess, model_file)

  208.  
  209. # 第一个卷积层

  210. W_conv1 = sess.graph.get_tensor_by_name("W_conv1:0")

  211. b_conv1 = sess.graph.get_tensor_by_name("b_conv1:0")

  212. conv_strides = [1, 1, 1, 1]

  213. kernel_size = [1, 2, 2, 1]

  214. pool_strides = [1, 2, 2, 1]

  215. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  216.  
  217. # 第二个卷积层

  218. W_conv2 = sess.graph.get_tensor_by_name("W_conv2:0")

  219. b_conv2 = sess.graph.get_tensor_by_name("b_conv2:0")

  220. conv_strides = [1, 1, 1, 1]

  221. kernel_size = [1, 1, 1, 1]

  222. pool_strides = [1, 1, 1, 1]

  223. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  224.  
  225.  
  226. # 全连接层

  227. W_fc1 = sess.graph.get_tensor_by_name("W_fc1:0")

  228. b_fc1 = sess.graph.get_tensor_by_name("b_fc1:0")

  229. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  230. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  231.  
  232.  
  233. # dropout

  234. keep_prob = tf.placeholder(tf.float32)

  235.  
  236. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  237.  
  238.  
  239. # readout层

  240. W_fc2 = sess.graph.get_tensor_by_name("W_fc2:0")

  241. b_fc2 = sess.graph.get_tensor_by_name("b_fc2:0")

  242.  
  243. # 定义优化器和训练op

  244. conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

  245.  
  246. for n in range(1,2):

  247. path = "test_images/%s.bmp" % (n)

  248. img = Image.open(path)

  249. width = img.size[0]

  250. height = img.size[1]

  251.  
  252. img_data = [[0]*SIZE for i in range(1)]

  253. for h in range(0, height):

  254. for w in range(0, width):

  255. if img.getpixel((w, h)) < 190:

  256. img_data[0][w+h*width] = 1

  257. else:

  258. img_data[0][w+h*width] = 0

  259.  
  260. result = sess.run(conv, feed_dict = {x: np.array(img_data), keep_prob: 1.0})

  261. max1 = 0

  262. max2 = 0

  263. max3 = 0

  264. max1_index = 0

  265. max2_index = 0

  266. max3_index = 0

  267. for j in range(NUM_CLASSES):

  268. if result[0][j] > max1:

  269. max1 = result[0][j]

  270. max1_index = j

  271. continue

  272. if (result[0][j]>max2) and (result[0][j]<=max1):

  273. max2 = result[0][j]

  274. max2_index = j

  275. continue

  276. if (result[0][j]>max3) and (result[0][j]<=max2):

  277. max3 = result[0][j]

  278. max3_index = j

  279. continue

  280.  
  281. nProvinceIndex = max1_index

  282. print ("概率: [%s %0.2f%%] [%s %0.2f%%] [%s %0.2f%%]" % (PROVINCES[max1_index],max1*100, PROVINCES[max2_index],max2*100, PROVINCES[max3_index],max3*100))

  283.  
  284. print ("省份简称是: %s" % PROVINCES[nProvinceIndex])

城市代号训练+识别代码(保存文件名为train-license-letters.py):

 

 
  1. #!/usr/bin/python3.5

  2. # -*- coding: utf-8 -*-

  3.  
  4. import sys

  5. import os

  6. import time

  7. import random

  8.  
  9. import numpy as np

  10. import tensorflow as tf

  11.  
  12. from PIL import Image

  13.  
  14.  
  15. SIZE = 1280

  16. WIDTH = 32

  17. HEIGHT = 40

  18. NUM_CLASSES = 26

  19. iterations = 500

  20.  
  21. SAVER_DIR = "train-saver/letters/"

  22.  
  23. LETTERS_DIGITS = ("A","B","C","D","E","F","G","H","J","K","L","M","N","P","Q","R","S","T","U","V","W","X","Y","Z","I","O")

  24. license_num = ""

  25.  
  26. time_begin = time.time()

  27.  
  28.  
  29. # 定义输入节点,对应于图片像素值矩阵集合和图片标签(即所代表的数字)

  30. x = tf.placeholder(tf.float32, shape=[None, SIZE])

  31. y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASSES])

  32.  
  33. x_image = tf.reshape(x, [-1, WIDTH, HEIGHT, 1])

  34.  
  35.  
  36. # 定义卷积函数

  37. def conv_layer(inputs, W, b, conv_strides, kernel_size, pool_strides, padding):

  38. L1_conv = tf.nn.conv2d(inputs, W, strides=conv_strides, padding=padding)

  39. L1_relu = tf.nn.relu(L1_conv + b)

  40. return tf.nn.max_pool(L1_relu, ksize=kernel_size, strides=pool_strides, padding='SAME')

  41.  
  42. # 定义全连接层函数

  43. def full_connect(inputs, W, b):

  44. return tf.nn.relu(tf.matmul(inputs, W) + b)

  45.  
  46.  
  47. if __name__ =='__main__' and sys.argv[1]=='train':

  48. # 第一次遍历图片目录是为了获取图片总数

  49. input_count = 0

  50. for i in range(0+10,NUM_CLASSES+10):

  51. dir = './train_images/training-set/letters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  52. for rt, dirs, files in os.walk(dir):

  53. for filename in files:

  54. input_count += 1

  55.  
  56. # 定义对应维数和各维长度的数组

  57. input_images = np.array([[0]*SIZE for i in range(input_count)])

  58. input_labels = np.array([[0]*NUM_CLASSES for i in range(input_count)])

  59.  
  60. # 第二次遍历图片目录是为了生成图片数据和标签

  61. index = 0

  62. for i in range(0+10,NUM_CLASSES+10):

  63. dir = './train_images/training-set/letters/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  64. for rt, dirs, files in os.walk(dir):

  65. for filename in files:

  66. filename = dir + filename

  67. img = Image.open(filename)

  68. width = img.size[0]

  69. height = img.size[1]

  70. for h in range(0, height):

  71. for w in range(0, width):

  72. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  73. if img.getpixel((w, h)) > 230:

  74. input_images[index][w+h*width] = 0

  75. else:

  76. input_images[index][w+h*width] = 1

  77. #print ("i=%d, index=%d" % (i, index))

  78. input_labels[index][i-10] = 1

  79. index += 1

  80.  
  81. # 第一次遍历图片目录是为了获取图片总数

  82. val_count = 0

  83. for i in range(0+10,NUM_CLASSES+10):

  84. dir = './train_images/validation-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  85. for rt, dirs, files in os.walk(dir):

  86. for filename in files:

  87. val_count += 1

  88.  
  89. # 定义对应维数和各维长度的数组

  90. val_images = np.array([[0]*SIZE for i in range(val_count)])

  91. val_labels = np.array([[0]*NUM_CLASSES for i in range(val_count)])

  92.  
  93. # 第二次遍历图片目录是为了生成图片数据和标签

  94. index = 0

  95. for i in range(0+10,NUM_CLASSES+10):

  96. dir = './train_images/validation-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  97. for rt, dirs, files in os.walk(dir):

  98. for filename in files:

  99. filename = dir + filename

  100. img = Image.open(filename)

  101. width = img.size[0]

  102. height = img.size[1]

  103. for h in range(0, height):

  104. for w in range(0, width):

  105. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  106. if img.getpixel((w, h)) > 230:

  107. val_images[index][w+h*width] = 0

  108. else:

  109. val_images[index][w+h*width] = 1

  110. val_labels[index][i-10] = 1

  111. index += 1

  112.  
  113. with tf.Session() as sess:

  114. # 第一个卷积层

  115. W_conv1 = tf.Variable(tf.truncated_normal([8, 8, 1, 16], stddev=0.1), name="W_conv1")

  116. b_conv1 = tf.Variable(tf.constant(0.1, shape=[16]), name="b_conv1")

  117. conv_strides = [1, 1, 1, 1]

  118. kernel_size = [1, 2, 2, 1]

  119. pool_strides = [1, 2, 2, 1]

  120. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  121.  
  122. # 第二个卷积层

  123. W_conv2 = tf.Variable(tf.truncated_normal([5, 5, 16, 32], stddev=0.1), name="W_conv2")

  124. b_conv2 = tf.Variable(tf.constant(0.1, shape=[32]), name="b_conv2")

  125. conv_strides = [1, 1, 1, 1]

  126. kernel_size = [1, 1, 1, 1]

  127. pool_strides = [1, 1, 1, 1]

  128. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  129.  
  130.  
  131. # 全连接层

  132. W_fc1 = tf.Variable(tf.truncated_normal([16 * 20 * 32, 512], stddev=0.1), name="W_fc1")

  133. b_fc1 = tf.Variable(tf.constant(0.1, shape=[512]), name="b_fc1")

  134. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  135. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  136.  
  137.  
  138. # dropout

  139. keep_prob = tf.placeholder(tf.float32)

  140.  
  141. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  142.  
  143.  
  144. # readout层

  145. W_fc2 = tf.Variable(tf.truncated_normal([512, NUM_CLASSES], stddev=0.1), name="W_fc2")

  146. b_fc2 = tf.Variable(tf.constant(0.1, shape=[NUM_CLASSES]), name="b_fc2")

  147.  
  148. # 定义优化器和训练op

  149. y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

  150. cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

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

  152.  
  153. correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))

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

  155.  
  156. sess.run(tf.global_variables_initializer())

  157.  
  158. time_elapsed = time.time() - time_begin

  159. print("读取图片文件耗费时间:%d秒" % time_elapsed)

  160. time_begin = time.time()

  161.  
  162. print ("一共读取了 %s 个训练图像, %s 个标签" % (input_count, input_count))

  163.  
  164. # 设置每次训练op的输入个数和迭代次数,这里为了支持任意图片总数,定义了一个余数remainder,譬如,如果每次训练op的输入个数为60,图片总数为150张,则前面两次各输入60张,最后一次输入30张(余数30)

  165. batch_size = 60

  166. iterations = iterations

  167. batches_count = int(input_count / batch_size)

  168. remainder = input_count % batch_size

  169. print ("训练数据集分成 %s 批, 前面每批 %s 个数据,最后一批 %s 个数据" % (batches_count+1, batch_size, remainder))

  170.  
  171. # 执行训练迭代

  172. for it in range(iterations):

  173. # 这里的关键是要把输入数组转为np.array

  174. for n in range(batches_count):

  175. train_step.run(feed_dict={x: input_images[n*batch_size:(n+1)*batch_size], y_: input_labels[n*batch_size:(n+1)*batch_size], keep_prob: 0.5})

  176. if remainder > 0:

  177. start_index = batches_count * batch_size;

  178. train_step.run(feed_dict={x: input_images[start_index:input_count-1], y_: input_labels[start_index:input_count-1], keep_prob: 0.5})

  179.  
  180. # 每完成五次迭代,判断准确度是否已达到100%,达到则退出迭代循环

  181. iterate_accuracy = 0

  182. if it%5 == 0:

  183. iterate_accuracy = accuracy.eval(feed_dict={x: val_images, y_: val_labels, keep_prob: 1.0})

  184. print ('第 %d 次训练迭代: 准确率 %0.5f%%' % (it, iterate_accuracy*100))

  185. if iterate_accuracy >= 0.9999 and it >= iterations:

  186. break;

  187.  
  188. print ('完成训练!')

  189. time_elapsed = time.time() - time_begin

  190. print ("训练耗费时间:%d秒" % time_elapsed)

  191. time_begin = time.time()

  192.  
  193. # 保存训练结果

  194. if not os.path.exists(SAVER_DIR):

  195. print ('不存在训练数据保存目录,现在创建保存目录')

  196. os.makedirs(SAVER_DIR)

  197. # 初始化saver

  198. saver = tf.train.Saver()

  199. saver_path = saver.save(sess, "%smodel.ckpt"%(SAVER_DIR))

  200.  
  201.  
  202.  
  203. if __name__ =='__main__' and sys.argv[1]=='predict':

  204. saver = tf.train.import_meta_graph("%smodel.ckpt.meta"%(SAVER_DIR))

  205. with tf.Session() as sess:

  206. model_file=tf.train.latest_checkpoint(SAVER_DIR)

  207. saver.restore(sess, model_file)

  208.  
  209. # 第一个卷积层

  210. W_conv1 = sess.graph.get_tensor_by_name("W_conv1:0")

  211. b_conv1 = sess.graph.get_tensor_by_name("b_conv1:0")

  212. conv_strides = [1, 1, 1, 1]

  213. kernel_size = [1, 2, 2, 1]

  214. pool_strides = [1, 2, 2, 1]

  215. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  216.  
  217. # 第二个卷积层

  218. W_conv2 = sess.graph.get_tensor_by_name("W_conv2:0")

  219. b_conv2 = sess.graph.get_tensor_by_name("b_conv2:0")

  220. conv_strides = [1, 1, 1, 1]

  221. kernel_size = [1, 1, 1, 1]

  222. pool_strides = [1, 1, 1, 1]

  223. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  224.  
  225.  
  226. # 全连接层

  227. W_fc1 = sess.graph.get_tensor_by_name("W_fc1:0")

  228. b_fc1 = sess.graph.get_tensor_by_name("b_fc1:0")

  229. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  230. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  231.  
  232.  
  233. # dropout

  234. keep_prob = tf.placeholder(tf.float32)

  235.  
  236. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  237.  
  238.  
  239. # readout层

  240. W_fc2 = sess.graph.get_tensor_by_name("W_fc2:0")

  241. b_fc2 = sess.graph.get_tensor_by_name("b_fc2:0")

  242.  
  243. # 定义优化器和训练op

  244. conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

  245.  
  246. for n in range(2,3):

  247. path = "test_images/%s.bmp" % (n)

  248. img = Image.open(path)

  249. width = img.size[0]

  250. height = img.size[1]

  251.  
  252. img_data = [[0]*SIZE for i in range(1)]

  253. for h in range(0, height):

  254. for w in range(0, width):

  255. if img.getpixel((w, h)) < 190:

  256. img_data[0][w+h*width] = 1

  257. else:

  258. img_data[0][w+h*width] = 0

  259.  
  260. result = sess.run(conv, feed_dict = {x: np.array(img_data), keep_prob: 1.0})

  261.  
  262. max1 = 0

  263. max2 = 0

  264. max3 = 0

  265. max1_index = 0

  266. max2_index = 0

  267. max3_index = 0

  268. for j in range(NUM_CLASSES):

  269. if result[0][j] > max1:

  270. max1 = result[0][j]

  271. max1_index = j

  272. continue

  273. if (result[0][j]>max2) and (result[0][j]<=max1):

  274. max2 = result[0][j]

  275. max2_index = j

  276. continue

  277. if (result[0][j]>max3) and (result[0][j]<=max2):

  278. max3 = result[0][j]

  279. max3_index = j

  280. continue

  281.  
  282. if n == 3:

  283. license_num += "-"

  284. license_num = license_num + LETTERS_DIGITS[max1_index]

  285. print ("概率: [%s %0.2f%%] [%s %0.2f%%] [%s %0.2f%%]" % (LETTERS_DIGITS[max1_index],max1*100, LETTERS_DIGITS[max2_index],max2*100, LETTERS_DIGITS[max3_index],max3*100))

  286.  
  287. print ("城市代号是: 【%s】" % license_num)

车牌编号训练+识别代码(保存文件名为train-license-digits.py):

 
  1. #!/usr/bin/python3.5

  2. # -*- coding: utf-8 -*-

  3.  
  4. import sys

  5. import os

  6. import time

  7. import random

  8.  
  9. import numpy as np

  10. import tensorflow as tf

  11.  
  12. from PIL import Image

  13.  
  14.  
  15. SIZE = 1280

  16. WIDTH = 32

  17. HEIGHT = 40

  18. NUM_CLASSES = 34

  19. iterations = 1000

  20.  
  21. SAVER_DIR = "train-saver/digits/"

  22.  
  23. LETTERS_DIGITS = ("0","1","2","3","4","5","6","7","8","9","A","B","C","D","E","F","G","H","J","K","L","M","N","P","Q","R","S","T","U","V","W","X","Y","Z")

  24. license_num = ""

  25.  
  26. time_begin = time.time()

  27.  
  28.  
  29. # 定义输入节点,对应于图片像素值矩阵集合和图片标签(即所代表的数字)

  30. x = tf.placeholder(tf.float32, shape=[None, SIZE])

  31. y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASSES])

  32.  
  33. x_image = tf.reshape(x, [-1, WIDTH, HEIGHT, 1])

  34.  
  35.  
  36. # 定义卷积函数

  37. def conv_layer(inputs, W, b, conv_strides, kernel_size, pool_strides, padding):

  38. L1_conv = tf.nn.conv2d(inputs, W, strides=conv_strides, padding=padding)

  39. L1_relu = tf.nn.relu(L1_conv + b)

  40. return tf.nn.max_pool(L1_relu, ksize=kernel_size, strides=pool_strides, padding='SAME')

  41.  
  42. # 定义全连接层函数

  43. def full_connect(inputs, W, b):

  44. return tf.nn.relu(tf.matmul(inputs, W) + b)

  45.  
  46.  
  47. if __name__ =='__main__' and sys.argv[1]=='train':

  48. # 第一次遍历图片目录是为了获取图片总数

  49. input_count = 0

  50. for i in range(0,NUM_CLASSES):

  51. dir = './train_images/training-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  52. for rt, dirs, files in os.walk(dir):

  53. for filename in files:

  54. input_count += 1

  55.  
  56. # 定义对应维数和各维长度的数组

  57. input_images = np.array([[0]*SIZE for i in range(input_count)])

  58. input_labels = np.array([[0]*NUM_CLASSES for i in range(input_count)])

  59.  
  60. # 第二次遍历图片目录是为了生成图片数据和标签

  61. index = 0

  62. for i in range(0,NUM_CLASSES):

  63. dir = './train_images/training-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  64. for rt, dirs, files in os.walk(dir):

  65. for filename in files:

  66. filename = dir + filename

  67. img = Image.open(filename)

  68. width = img.size[0]

  69. height = img.size[1]

  70. for h in range(0, height):

  71. for w in range(0, width):

  72. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  73. if img.getpixel((w, h)) > 230:

  74. input_images[index][w+h*width] = 0

  75. else:

  76. input_images[index][w+h*width] = 1

  77. input_labels[index][i] = 1

  78. index += 1

  79.  
  80. # 第一次遍历图片目录是为了获取图片总数

  81. val_count = 0

  82. for i in range(0,NUM_CLASSES):

  83. dir = './train_images/validation-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  84. for rt, dirs, files in os.walk(dir):

  85. for filename in files:

  86. val_count += 1

  87.  
  88. # 定义对应维数和各维长度的数组

  89. val_images = np.array([[0]*SIZE for i in range(val_count)])

  90. val_labels = np.array([[0]*NUM_CLASSES for i in range(val_count)])

  91.  
  92. # 第二次遍历图片目录是为了生成图片数据和标签

  93. index = 0

  94. for i in range(0,NUM_CLASSES):

  95. dir = './train_images/validation-set/%s/' % i # 这里可以改成你自己的图片目录,i为分类标签

  96. for rt, dirs, files in os.walk(dir):

  97. for filename in files:

  98. filename = dir + filename

  99. img = Image.open(filename)

  100. width = img.size[0]

  101. height = img.size[1]

  102. for h in range(0, height):

  103. for w in range(0, width):

  104. # 通过这样的处理,使数字的线条变细,有利于提高识别准确率

  105. if img.getpixel((w, h)) > 230:

  106. val_images[index][w+h*width] = 0

  107. else:

  108. val_images[index][w+h*width] = 1

  109. val_labels[index][i] = 1

  110. index += 1

  111.  
  112. with tf.Session() as sess:

  113. # 第一个卷积层

  114. W_conv1 = tf.Variable(tf.truncated_normal([8, 8, 1, 16], stddev=0.1), name="W_conv1")

  115. b_conv1 = tf.Variable(tf.constant(0.1, shape=[16]), name="b_conv1")

  116. conv_strides = [1, 1, 1, 1]

  117. kernel_size = [1, 2, 2, 1]

  118. pool_strides = [1, 2, 2, 1]

  119. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  120.  
  121. # 第二个卷积层

  122. W_conv2 = tf.Variable(tf.truncated_normal([5, 5, 16, 32], stddev=0.1), name="W_conv2")

  123. b_conv2 = tf.Variable(tf.constant(0.1, shape=[32]), name="b_conv2")

  124. conv_strides = [1, 1, 1, 1]

  125. kernel_size = [1, 1, 1, 1]

  126. pool_strides = [1, 1, 1, 1]

  127. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  128.  
  129.  
  130. # 全连接层

  131. W_fc1 = tf.Variable(tf.truncated_normal([16 * 20 * 32, 512], stddev=0.1), name="W_fc1")

  132. b_fc1 = tf.Variable(tf.constant(0.1, shape=[512]), name="b_fc1")

  133. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  134. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  135.  
  136.  
  137. # dropout

  138. keep_prob = tf.placeholder(tf.float32)

  139.  
  140. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  141.  
  142.  
  143. # readout层

  144. W_fc2 = tf.Variable(tf.truncated_normal([512, NUM_CLASSES], stddev=0.1), name="W_fc2")

  145. b_fc2 = tf.Variable(tf.constant(0.1, shape=[NUM_CLASSES]), name="b_fc2")

  146.  
  147. # 定义优化器和训练op

  148. y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

  149. cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

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

  151.  
  152. correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))

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

  154.  
  155. sess.run(tf.global_variables_initializer())

  156.  
  157. time_elapsed = time.time() - time_begin

  158. print("读取图片文件耗费时间:%d秒" % time_elapsed)

  159. time_begin = time.time()

  160.  
  161. print ("一共读取了 %s 个训练图像, %s 个标签" % (input_count, input_count))

  162.  
  163. # 设置每次训练op的输入个数和迭代次数,这里为了支持任意图片总数,定义了一个余数remainder,譬如,如果每次训练op的输入个数为60,图片总数为150张,则前面两次各输入60张,最后一次输入30张(余数30)

  164. batch_size = 60

  165. iterations = iterations

  166. batches_count = int(input_count / batch_size)

  167. remainder = input_count % batch_size

  168. print ("训练数据集分成 %s 批, 前面每批 %s 个数据,最后一批 %s 个数据" % (batches_count+1, batch_size, remainder))

  169.  
  170. # 执行训练迭代

  171. for it in range(iterations):

  172. # 这里的关键是要把输入数组转为np.array

  173. for n in range(batches_count):

  174. train_step.run(feed_dict={x: input_images[n*batch_size:(n+1)*batch_size], y_: input_labels[n*batch_size:(n+1)*batch_size], keep_prob: 0.5})

  175. if remainder > 0:

  176. start_index = batches_count * batch_size;

  177. train_step.run(feed_dict={x: input_images[start_index:input_count-1], y_: input_labels[start_index:input_count-1], keep_prob: 0.5})

  178.  
  179. # 每完成五次迭代,判断准确度是否已达到100%,达到则退出迭代循环

  180. iterate_accuracy = 0

  181. if it%5 == 0:

  182. iterate_accuracy = accuracy.eval(feed_dict={x: val_images, y_: val_labels, keep_prob: 1.0})

  183. print ('第 %d 次训练迭代: 准确率 %0.5f%%' % (it, iterate_accuracy*100))

  184. if iterate_accuracy >= 0.9999 and it >= iterations:

  185. break;

  186.  
  187. print ('完成训练!')

  188. time_elapsed = time.time() - time_begin

  189. print ("训练耗费时间:%d秒" % time_elapsed)

  190. time_begin = time.time()

  191.  
  192. # 保存训练结果

  193. if not os.path.exists(SAVER_DIR):

  194. print ('不存在训练数据保存目录,现在创建保存目录')

  195. os.makedirs(SAVER_DIR)

  196. # 初始化saver

  197. saver = tf.train.Saver()

  198. saver_path = saver.save(sess, "%smodel.ckpt"%(SAVER_DIR))

  199.  
  200.  
  201.  
  202. if __name__ =='__main__' and sys.argv[1]=='predict':

  203. saver = tf.train.import_meta_graph("%smodel.ckpt.meta"%(SAVER_DIR))

  204. with tf.Session() as sess:

  205. model_file=tf.train.latest_checkpoint(SAVER_DIR)

  206. saver.restore(sess, model_file)

  207.  
  208. # 第一个卷积层

  209. W_conv1 = sess.graph.get_tensor_by_name("W_conv1:0")

  210. b_conv1 = sess.graph.get_tensor_by_name("b_conv1:0")

  211. conv_strides = [1, 1, 1, 1]

  212. kernel_size = [1, 2, 2, 1]

  213. pool_strides = [1, 2, 2, 1]

  214. L1_pool = conv_layer(x_image, W_conv1, b_conv1, conv_strides, kernel_size, pool_strides, padding='SAME')

  215.  
  216. # 第二个卷积层

  217. W_conv2 = sess.graph.get_tensor_by_name("W_conv2:0")

  218. b_conv2 = sess.graph.get_tensor_by_name("b_conv2:0")

  219. conv_strides = [1, 1, 1, 1]

  220. kernel_size = [1, 1, 1, 1]

  221. pool_strides = [1, 1, 1, 1]

  222. L2_pool = conv_layer(L1_pool, W_conv2, b_conv2, conv_strides, kernel_size, pool_strides, padding='SAME')

  223.  
  224.  
  225. # 全连接层

  226. W_fc1 = sess.graph.get_tensor_by_name("W_fc1:0")

  227. b_fc1 = sess.graph.get_tensor_by_name("b_fc1:0")

  228. h_pool2_flat = tf.reshape(L2_pool, [-1, 16 * 20*32])

  229. h_fc1 = full_connect(h_pool2_flat, W_fc1, b_fc1)

  230.  
  231.  
  232. # dropout

  233. keep_prob = tf.placeholder(tf.float32)

  234.  
  235. h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

  236.  
  237.  
  238. # readout层

  239. W_fc2 = sess.graph.get_tensor_by_name("W_fc2:0")

  240. b_fc2 = sess.graph.get_tensor_by_name("b_fc2:0")

  241.  
  242. # 定义优化器和训练op

  243. conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

  244.  
  245. for n in range(3,8):

  246. path = "test_images/%s.bmp" % (n)

  247. img = Image.open(path)

  248. width = img.size[0]

  249. height = img.size[1]

  250.  
  251. img_data = [[0]*SIZE for i in range(1)]

  252. for h in range(0, height):

  253. for w in range(0, width):

  254. if img.getpixel((w, h)) < 190:

  255. img_data[0][w+h*width] = 1

  256. else:

  257. img_data[0][w+h*width] = 0

  258.  
  259. result = sess.run(conv, feed_dict = {x: np.array(img_data), keep_prob: 1.0})

  260.  
  261. max1 = 0

  262. max2 = 0

  263. max3 = 0

  264. max1_index = 0

  265. max2_index = 0

  266. max3_index = 0

  267. for j in range(NUM_CLASSES):

  268. if result[0][j] > max1:

  269. max1 = result[0][j]

  270. max1_index = j

  271. continue

  272. if (result[0][j]>max2) and (result[0][j]<=max1):

  273. max2 = result[0][j]

  274. max2_index = j

  275. continue

  276. if (result[0][j]>max3) and (result[0][j]<=max2):

  277. max3 = result[0][j]

  278. max3_index = j

  279. continue

  280.  
  281. license_num = license_num + LETTERS_DIGITS[max1_index]

  282. print ("概率: [%s %0.2f%%] [%s %0.2f%%] [%s %0.2f%%]" % (LETTERS_DIGITS[max1_index],max1*100, LETTERS_DIGITS[max2_index],max2*100, LETTERS_DIGITS[max3_index],max3*100))

  283.  
  284. print ("车牌编号是: 【%s】" % license_num)

保存好上面三个python脚本后,我们首先进行省份简称训练。在运行代码之前,需要先把数据集解压到训练脚本所在目录。然后,在命令行中进入脚本所在目录,输入执行如下命令:

python train-license-province.py train

训练结果如下:

 

然后进行省份简称识别,在命令行输入执行如下命令:

 

python train-license-province.py predict

 

 

执行城市代号训练(相当于训练26个字母):

 

python train-license-letters.py train


识别城市代号:

 

 

python train-license-letters.py predict

 

执行车牌编号训练(相当于训练24个字母+10个数字,我国交通法规规定车牌编号中不包含字母I和O):

 

 

python train-license-digits.py train


识别车牌编号:

 

 

python train-license-digits.py predict

 

可以看到,在测试图片上,识别准确率很高。识别结果是闽O-1672Q。

 

下图是测试图片的车牌原图:

评论 5
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值