Research On License Plate Recognition Using Hierarchical Feature Layers

In recent years,some methods based on convolutional neural network(CNN)have been applied in the field of license plate recognition,which has achieved better performance than traditional methods.However,the most of these methods do not adapt well when the application scenarios changed,especially in the aspect of small size license plate recognition.This paper mainly studies a license plate detection method based on CNN,which generate a series of proposal regions from different levels of feature maps,obtains the license plate area after the relevant image processing algorithm,and can detect the license plate with various sizes.On the other hand,based on the case of convolution neural network handwritten digit recognition,this paper presents a character recognition algorithm based on convolution neural network,and applies it to the vehicle license plate character recognition task.In this algorithm,the character sequence can be obtained by classifying the characters directly on the license plate without character segmentation.Meanwhile,this paper designs an automatic license plate generation algorithm,which is very close to the data collected in real scene,and reduces the heavy workload of data tagging.The innovation of this paper is:1.This paper constructs a method of license plate detection based on hierarchy feature layers.According to the characteristics of "the cell on hierarchy feature maps have different receptive field",different sizes of license plate can be detected by use this method,which can be used in different application scenarios.2.This paper improves the recognition algorithm of license plate character based on CNN.The algorithm can classify each character on the license plate without character segmentation,so that the sequence of license plate characters can be recognized quickly and accurately.3.This paper proposed an automatic license plate generation algorithm.Theoretically,the algorithm can generate an infinite number of training samples,which has the characteristics of high simulation and fast production,and saves lots of time and manpower.In this paper,the license plate detection algorithm obtains a 99.5%average precision(AP)on the OpenITS(an open test dataset),and obtains a high AP on the more complex test dataset.The comprehensive experiment shows that the license plate recognition system has the characteristics of high precision,strong stability and fast speed.