Design And Algorithm Implementation Of License Plate Recognition System Based On Convolutional Neural Network

With the development of information technology and intelligent technology,the informationization and intelligence of traffic management systems are the general trend.The license plate recognition system is a core component of the intelligent traffic system and has been widely applied to many places..In the current license plate recognition process,there are many problems such as large background interference,rapid illumination change,and unfixed license plate size,which leads to the slow recognition speed and low recognition rate.This paper makes use of the convolutional neural network to recognize the superiority of twodimensional images,and designs a license plate recognition system based on convolutional neural network.In response to the problems existing in the license plate recognition process for each module,the following five tasks have been completed:1)In vehicle detection,in view of the existence of two kinds of noise in video surveillance,median and mean integrated filter is used to reduce the noise.In order to ensure high real-time performance,a vehicle detection method based on background difference method is designed,and the image after the difference is detected by morphological way.Finally,the performance of the vehicle detection algorithm is tested under a variety of road conditions and light conditions.2)In license plate positioning,this paper presents a license plate location method based on morphology and edge statistics.The vehicle foreground image first locates the license plate area by morphology and then locates the license plate according to the unique edge sparseness of the license plate,to exclude the similar non-license plate area.3)In character segmentation,a bi-directional scanning character segmentation method based on vertical projection is proposed in this paper,which is specific to the "." area of Chinese license plate.Through a horizontal scan and two vertical scanning,the influence of the left and right borders on the characters is avoided and the characters are effectively separated.4)In character recognition,due to the traditional convolution neural LeNet-5 network used in Chinese license plate recognition exists a series of problems.In this paper,we design a global median pooling(GMP-LeNet)network,elaborating the detailed structure and parameters of the network.This paper introduces the training and measuring method of Chinese license plate within the caffe framework and proposes a residual stop-early training algorithm.5)In terms of system design,we construct a license plate recognition system based on convolutional neural network.A video surveillance system is designed and a license plate recognition software is implemented,which integrates each submodule.After testing and analyzing,the recognition rate of each sub-module is over 96%;the recognition rate of the whole system is 93%;and the recognition speed is about 350 milliseconds.It performs well in the same type of license plate recognition system,which can basically meet the practical application requirements.