【深度学习系列】车牌识别实践（二）
上节我们讲了第一部分，如何用生成简易的车牌，这节课中我们会用PaddlePaddle来识别生成的车牌。
数据读取
在上一节生成车牌时，我们可以分别生成训练数据和测试数据，方法如下（完整代码在这里）：
1 # 将生成的车牌图片写入文件夹，对应的label写入label.txt 2 def genBatch(self, batchSize,pos,charRange, outputPath,size): 3 if (not os.path.exists(outputPath)): 4 os.mkdir(outputPath) 5 outfile = open('label.txt','w') 6 for i in xrange(batchSize): 7 plateStr,plate = G.genPlateString(-1,-1) 8 print plateStr,plate 9 img = G.generate(plateStr); 10 img = cv2.resize(img,size); 11 cv2.imwrite(outputPath + "/" + str(i).zfill(2) + ".jpg", img); 12 outfile.write(str(plate)+"\n")
生成好数据后，我们写一个reader来读取数据 ( reador.py )
1 def reader_creator(data,label): 2 def reader(): 3 for i in xrange(len(data)): 4 yield data[i,:],int(label[i]) 5 return reader
灌入模型时，我们需要调用paddle.batch函数，将数据shuffle后批量灌入模型中：
1 # 读取训练数据 2 train_reader = paddle.batch(paddle.reader.shuffle( 3 reador.reader_creator(X_train,Y_train),buf_size=200), 4 batch_size=16) 5 6 # 读取验证数据 7 val_reader = paddle.batch(paddle.reader.shuffle( 8 reador.reader_creator(X_val,Y_val),buf_size=200), 9 batch_size=16) 10 trainer.train(reader=train_reader,num_passes=20,event_handler=event_handler)
构建网络模型
因为我们训练的是端到端的车牌识别，所以一开始构建了两个卷积-池化层训练，训练完后同步训练7个全连接层，分别对应车牌的7位字符，最后将其拼接起来，与原始的label计算Softmax值，预测训练结果。
1 def get_network_cnn(self): 2 # 加载data和label 3 x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(self.data)) 4 y = paddle.layer.data(name='y', type=paddle.data_type.integer_value(self.label)) 5 # 构建卷积-池化层-1 6 conv_pool_1 = paddle.networks.simple_img_conv_pool( 7 input=x, 8 filter_size=12, 9 num_filters=50, 10 num_channel=1, 11 pool_size=2, 12 pool_stride=2, 13 act=paddle.activation.Relu()) 14 drop_1 = paddle.layer.dropout(input=conv_pool_1, dropout_rate=0.5) 15 # 构建卷积-池化层-2 16 conv_pool_2 = paddle.networks.simple_img_conv_pool( 17 input=drop_1, 18 filter_size=5, 19 num_filters=50, 20 num_channel=20, 21 pool_size=2, 22 pool_stride=2, 23 act=paddle.activation.Relu()) 24 drop_2 = paddle.layer.dropout(input=conv_pool_2, dropout_rate=0.5) 25 26 # 全连接层 27 fc = paddle.layer.fc(input = drop_2, size = 120) 28 fc1_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 29 fc1 = paddle.layer.fc(input = fc1_drop,size = 65,act = paddle.activation.Linear()) 30 31 fc2_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 32 fc2 = paddle.layer.fc(input = fc2_drop,size = 65,act = paddle.activation.Linear()) 33 34 fc3_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 35 fc3 = paddle.layer.fc(input = fc3_drop,size = 65,act = paddle.activation.Linear()) 36 37 fc4_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 38 fc4 = paddle.layer.fc(input = fc4_drop,size = 65,act = paddle.activation.Linear()) 39 40 fc5_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 41 fc5 = paddle.layer.fc(input = fc5_drop,size = 65,act = paddle.activation.Linear()) 42 43 fc6_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 44 fc6 = paddle.layer.fc(input = fc6_drop,size = 65,act = paddle.activation.Linear()) 45 46 fc7_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 47 fc7 = paddle.layer.fc(input = fc7_drop,size = 65,act = paddle.activation.Linear()) 48 49 # 将训练好的7个字符的全连接层拼接起来 50 fc_concat = paddle.layer.concact(input = [fc21, fc22, fc23, fc24,fc25,fc26,fc27], axis = 0) 51 predict = paddle.layer.classification_cost(input = fc_concat,label = y,act=paddle.activation.Softmax()) 52 return predict
训练模型
构建好网络模型后，就是比较常见的步骤了，譬如初始化，定义优化方法， 定义训练参数，定义训练器等等，再把第一步里我们写好的数据读取的方式放进去，就可以正常跑模型了。
1 class NeuralNetwork(object): 2 def __init__(self,X_train,Y_train,X_val,Y_val): 3 paddle.init(use_gpu = with_gpu,trainer_count=1) 4 5 self.X_train = X_train 6 self.Y_train = Y_train 7 self.X_val = X_val 8 self.Y_val = Y_val 9 10 11 def get_network_cnn(self): 12 13 x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(self.data)) 14 y = paddle.layer.data(name='y', type=paddle.data_type.integer_value(self.label)) 15 conv_pool_1 = paddle.networks.simple_img_conv_pool( 16 input=x, 17 filter_size=12, 18 num_filters=50, 19 num_channel=1, 20 pool_size=2, 21 pool_stride=2, 22 act=paddle.activation.Relu()) 23 drop_1 = paddle.layer.dropout(input=conv_pool_1, dropout_rate=0.5) 24 conv_pool_2 = paddle.networks.simple_img_conv_pool( 25 input=drop_1, 26 filter_size=5, 27 num_filters=50, 28 num_channel=20, 29 pool_size=2, 30 pool_stride=2, 31 act=paddle.activation.Relu()) 32 drop_2 = paddle.layer.dropout(input=conv_pool_2, dropout_rate=0.5) 33 34 fc = paddle.layer.fc(input = drop_2, size = 120) 35 fc1_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 36 fc1 = paddle.layer.fc(input = fc1_drop,size = 65,act = paddle.activation.Linear()) 37 38 fc2_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 39 fc2 = paddle.layer.fc(input = fc2_drop,size = 65,act = paddle.activation.Linear()) 40 41 fc3_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 42 fc3 = paddle.layer.fc(input = fc3_drop,size = 65,act = paddle.activation.Linear()) 43 44 fc4_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 45 fc4 = paddle.layer.fc(input = fc4_drop,size = 65,act = paddle.activation.Linear()) 46 47 fc5_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 48 fc5 = paddle.layer.fc(input = fc5_drop,size = 65,act = paddle.activation.Linear()) 49 50 fc6_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 51 fc6 = paddle.layer.fc(input = fc6_drop,size = 65,act = paddle.activation.Linear()) 52 53 fc7_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5) 54 fc7 = paddle.layer.fc(input = fc7_drop,size = 65,act = paddle.activation.Linear()) 55 56 fc_concat = paddle.layer.concact(input = [fc21, fc22, fc23, fc24,fc25,fc26,fc27], axis = 0) 57 predict = paddle.layer.classification_cost(input = fc_concat,label = y,act=paddle.activation.Softmax()) 58 return predict 59 60 # 定义训练器 61 def get_trainer(self): 62 63 cost = self.get_network() 64 65 #获取参数 66 parameters = paddle.parameters.create(cost) 67 68 69 optimizer = paddle.optimizer.Momentum( 70 momentum=0.9, 71 regularization=paddle.optimizer.L2Regularization(rate=0.0002 * 128), 72 learning_rate=0.001, 73 learning_rate_schedule = "pass_manual") 74 75 76 # 创建训练器 77 trainer = paddle.trainer.SGD( 78 cost=cost, parameters=parameters, update_equation=optimizer) 79 return trainer 80 81 82 # 开始训练 83 def start_trainer(self,X_train,Y_train,X_val,Y_val): 84 trainer = self.get_trainer() 85 86 result_lists = [] 87 def event_handler(event): 88 if isinstance(event, paddle.event.EndIteration): 89 if event.batch_id % 10 == 0: 90 print "\nPass %d, Batch %d, Cost %f, %s" % ( 91 event.pass_id, event.batch_id, event.cost, event.metrics) 92 if isinstance(event, paddle.event.EndPass): 93 # 保存训练好的参数 94 with open('params_pass_%d.tar' % event.pass_id, 'w') as f: 95 parameters.to_tar(f) 96 # feeding = ['x','y'] 97 result = trainer.test( 98 reader=val_reader) 99 # feeding=feeding) 100 print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics) 101 102 result_lists.append((event.pass_id, result.cost, 103 result.metrics['classification_error_evaluator'])) 104 105 # 开始训练 106 train_reader = paddle.batch(paddle.reader.shuffle( 107 reador.reader_creator(X_train,Y_train),buf_size=200), 108 batch_size=16) 109 110 val_reader = paddle.batch(paddle.reader.shuffle( 111 reador.reader_creator(X_val,Y_val),buf_size=200), 112 batch_size=16) 113 # val_reader = paddle.reader(reador.reader_creator(X_val,Y_val),batch_size=16) 114 115 trainer.train(reader=train_reader,num_passes=20,event_handler=event_handler)
输出结果
上一步训练完以后，保存训练完的模型，然后写一个test.py进行预测，需要注意的是，在预测时，构建的网络结构得和训练的网络结构相同。
#批量预测测试图片准确率 python test.py /Users/shelter/test ##输出结果示例 output: 预测车牌号码为:津 K 4 2 R M Y 输入图片数量:100 输入图片行准确率:0.72 输入图片列准确率:0.86
如果是一次性只预测一张的话，在终端里会显示原始的图片与预测的值，如果是批量预测的话，会打印出预测的总准确率，包括行与列的准确率。
总结
车牌识别的方法有很多，商业化落地的方法也很成熟，传统的方法需要对图片灰度化，字符进行切分等，需要很多数据预处理的过程，端到端的方法可以直接将原始的图片灌进去进行训练，最后出来预测的车牌字符的结果，这个方法在构建了两层卷积-池化网络结构后，并行训练了7个全连接层来进行车牌的字符识别，可以实现端到端的识别。但是在实际训练过程中，仍然有一些问题，譬如前几个训练的全连接层的准确率要比最后一两个的准确率高，大家可以分别打印出每一个全连接层的训练结果准确率对比一下，可能是由于训练还没有收敛导致的，也可能有其他原因，如果在做的过程中发现有什么问题，或者有更好的方法，欢迎留言~
参考文献：
1.我的github：https://github.com/huxiaoman7/mxnet-cnn-plate-recognition
