Research On License Plate Recognition Based On Convolution Neural Network

Because our country’s transportation develop quickly,the number of cars is increasing,and the development of efficient vehicle management system has become an important problem that needs to be solved urgently.In the traditional way,the license plate recognition system realized by image processing technology has poor anti-interference ability,and cannot perform accurate recognition under the influence of various factors.The research of deep learning has made great progress,the application of neural network to realize license plate recognition has become the current mainstream.Compared with traditional methods,neural network has high recognition rate,fast recognition speed and strong anti-interference ability.However,in order to gain higher performance,the model grows more and more complicated and has a growing number of parameters,so it cannot be applied to embedded devices.So,this paper mainly studies two directions.One is to use convolutional neural network to achieve efficient license plate recognition;the other is to optimize the neural network in a lightweight way to reduce the amount of parameters and complexity of the network,so that the license plate recognition network can be used in embedded devices.Firstly,this paper studies the application of convolutional neural network in license car discriminate.The paper uses the YOLOv3 algorithm to realize license plate recognition,and introduces two representative target detection algorithms,Faster R-CNN and SSD,for comparative analysis.The final license plate recognition rate is 99.66%,and the recognition speed is 0.074s,which meets the market requirements for accuracy and recognition speed.In view of the problem of network lightweight,this paper uses the skeleton network of MobileNetV2 to replace the skeleton network of YOLOv3,and further improves the network by using depthwise separable convolution,residual structure,and network clipping.Finally,the trial achievement suggest that the ultimate recognition rate of the lightweight network is 99.16%,the recognition speed is 0.054s,and the number of network parameters is greatly reduced to only 3.37M,which proves that the lightweight method in this paper has a good effect.This paper also introduces YOLOv3-Tiny and the lightweight YOLOv3 network for comparison.In embedded devices,the lightweight YOLOv3 network has better recognition speed and recognition accuracy than YOLOv3-Tiny,which is more suitable for license plates in embedded devices.