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）：
# 将生成的车牌图片写入文件夹，对应的label写入label.txt
def genBatch(self, batchSize,pos,charRange, outputPath,size):
if (not os.path.exists(outputPath)):
os.mkdir(outputPath)
outfile = open('label.txt','w')
for i in xrange(batchSize):
plateStr,plate = G.genPlateString(-1,-1)
print plateStr,plate
img = G.generate(plateStr);
img = cv2.resize(img,size);
cv2.imwrite(outputPath + "/" + str(i).zfill(2) + ".jpg", img);
outfile.write(str(plate)+"\n")
生成好数据后，我们写一个reader来读取数据 ( reador.py )
def reader_creator(data,label):
def reader():
for i in xrange(len(data)):
yield data[i,:],int(label[i])
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：
# 读取训练数据
train_reader = paddle.batch(paddle.reader.shuffle(
reador.reader_creator(X_train,Y_train),buf_size=200),
batch_size=16) # 读取验证数据
val_reader = paddle.batch(paddle.reader.shuffle(
reador.reader_creator(X_val,Y_val),buf_size=200),
batch_size=16)
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。
def get_network_cnn(self):
# 加载data和label
x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(self.data))
y = paddle.layer.data(name='y', type=paddle.data_type.integer_value(self.label))
# 构建卷积-池化层-1
conv_pool_1 = paddle.networks.simple_img_conv_pool(
input=x,
filter_size=12,
num_filters=50,
num_channel=1,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
drop_1 = paddle.layer.dropout(input=conv_pool_1, dropout_rate=0.5)
# 构建卷积-池化层-2
conv_pool_2 = paddle.networks.simple_img_conv_pool(
input=drop_1,
filter_size=5,
num_filters=50,
num_channel=20,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
drop_2 = paddle.layer.dropout(input=conv_pool_2, dropout_rate=0.5) # 全连接层
fc = paddle.layer.fc(input = drop_2, size = 120)
fc1_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc1 = paddle.layer.fc(input = fc1_drop,size = 65,act = paddle.activation.Linear()) fc2_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc2 = paddle.layer.fc(input = fc2_drop,size = 65,act = paddle.activation.Linear()) fc3_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc3 = paddle.layer.fc(input = fc3_drop,size = 65,act = paddle.activation.Linear()) fc4_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc4 = paddle.layer.fc(input = fc4_drop,size = 65,act = paddle.activation.Linear()) fc5_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc5 = paddle.layer.fc(input = fc5_drop,size = 65,act = paddle.activation.Linear()) fc6_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc6 = paddle.layer.fc(input = fc6_drop,size = 65,act = paddle.activation.Linear()) fc7_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc7 = paddle.layer.fc(input = fc7_drop,size = 65,act = paddle.activation.Linear()) # 将训练好的7个字符的全连接层拼接起来
fc_concat = paddle.layer.concact(input = [fc21, fc22, fc23, fc24,fc25,fc26,fc27], axis = 0)
predict = paddle.layer.classification_cost(input = fc_concat,label = y,act=paddle.activation.Softmax())
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。
class NeuralNetwork(object):
def __init__(self,X_train,Y_train,X_val,Y_val):
paddle.init(use_gpu = with_gpu,trainer_count=1) self.X_train = X_train
self.Y_train = Y_train
self.X_val = X_val
self.Y_val = Y_val def get_network_cnn(self): x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(self.data))
y = paddle.layer.data(name='y', type=paddle.data_type.integer_value(self.label))
conv_pool_1 = paddle.networks.simple_img_conv_pool(
input=x,
filter_size=12,
num_filters=50,
num_channel=1,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
drop_1 = paddle.layer.dropout(input=conv_pool_1, dropout_rate=0.5)
conv_pool_2 = paddle.networks.simple_img_conv_pool(
input=drop_1,
filter_size=5,
num_filters=50,
num_channel=20,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
drop_2 = paddle.layer.dropout(input=conv_pool_2, dropout_rate=0.5) fc = paddle.layer.fc(input = drop_2, size = 120)
fc1_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc1 = paddle.layer.fc(input = fc1_drop,size = 65,act = paddle.activation.Linear()) fc2_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc2 = paddle.layer.fc(input = fc2_drop,size = 65,act = paddle.activation.Linear()) fc3_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc3 = paddle.layer.fc(input = fc3_drop,size = 65,act = paddle.activation.Linear()) fc4_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc4 = paddle.layer.fc(input = fc4_drop,size = 65,act = paddle.activation.Linear()) fc5_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc5 = paddle.layer.fc(input = fc5_drop,size = 65,act = paddle.activation.Linear()) fc6_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc6 = paddle.layer.fc(input = fc6_drop,size = 65,act = paddle.activation.Linear()) fc7_drop = paddle.layer.dropout(input = fc,dropout_rate = 0.5)
fc7 = paddle.layer.fc(input = fc7_drop,size = 65,act = paddle.activation.Linear()) fc_concat = paddle.layer.concact(input = [fc21, fc22, fc23, fc24,fc25,fc26,fc27], axis = 0)
predict = paddle.layer.classification_cost(input = fc_concat,label = y,act=paddle.activation.Softmax())
return predict # 定义训练器
def get_trainer(self): cost = self.get_network() #获取参数
parameters = paddle.parameters.create(cost) optimizer = paddle.optimizer.Momentum(
momentum=0.9,
regularization=paddle.optimizer.L2Regularization(rate=0.0002 * 128),
learning_rate=0.001,
learning_rate_schedule = "pass_manual") # 创建训练器
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=optimizer)
return trainer # 开始训练
def start_trainer(self,X_train,Y_train,X_val,Y_val):
trainer = self.get_trainer() result_lists = []
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 10 == 0:
print "\nPass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
if isinstance(event, paddle.event.EndPass):
# 保存训练好的参数
with open('params_pass_%d.tar' % event.pass_id, 'w') as f:
parameters.to_tar(f)
# feeding = ['x','y']
result = trainer.test(
reader=val_reader)
# feeding=feeding)
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics) result_lists.append((event.pass_id, result.cost,
result.metrics['classification_error_evaluator'])) # 开始训练
train_reader = paddle.batch(paddle.reader.shuffle(
reador.reader_creator(X_train,Y_train),buf_size=200),
batch_size=16) val_reader = paddle.batch(paddle.reader.shuffle(
reador.reader_creator(X_val,Y_val),buf_size=200),
batch_size=16)
# val_reader = paddle.reader(reador.reader_creator(X_val,Y_val),batch_size=16) 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