Research On Dynamic License Plate Recognition Algorithm With Convolutional Neural Network

With the improvement of living standards,cars have gradually entered people’s families.With the increase of vehicles,the management statistics of vehicles have become more complicated.Therefore,the license plate recognition system has appeared,which is widely used in various checkpoints,toll stations and criminal cases.It greatly reduces management costs,improves work efficiency,and eases traffic jams and other problems.However,the currently widely used license plate recognition system is mainly static recognition,which has higher requirements for recognition conditions.It is necessary to park the vehicle near a designated location and perform the recognition after shooting,which has greater limitations.This paper designs a dynamic license plate recognition algorithm based on convolutional neural network,which is mainly composed of three parts.The first part is the license plate location.The license plate location is completed by the YOLOv3-TINY algorithm twice,and GIo U is used to replace the Io U of the original algorithm.In order to be able to correct the license plate more effectively,the image area is enlarged by 15% when output.Experiments show that the accuracy and recall rate of this algorithm have reached about 95%,and the detection speed has reached 81 frames per second.The second part is the license plate correction.First,the position of the license plate is initially located through the HSV color filter,and then the white filter is performed in the expanded blue area.The license plate is screened by the proportion of blue and white to the image and the degree of dispersion of white characters.Afterwards,the convex hull detection is performed,the convex points are selected for straight line fitting to obtain the vertices,and finally the perspective transformation is performed according to the vertices.Experiments show that when comparing the Canny edge detection method and the projection method,this algorithm is superior to the other two in both the correction effect and the correction speed.The third part is character recognition.Firstly,every three corrected license plates in each frame of image are merged into one to improve the recognition efficiency.Then improve YOLOv3-TINY,add a character filtering algorithm to prevent a character from being recognized multiple times,and at the same time correct the output wrong license plate according to the license plate law.Experiments show that when the original resolution of characters is greater than 4,the correct rate reaches 95% and the recognition speed reaches 140 frames per second.Finally,the whole license plate system is tested,mainly from two parts: picture recognition and video recognition.Experiments show that both picture recognition and video recognition have a correct rate of more than 90%,and even a complex lane scene with 1920×1080 resolution has a correct rate of 90% and a recognition speed of 38 frames per second.Through many experiments,it is proved that this algorithm meets the real-time and correct rate requirements of dynamic recognition,and has good superiority and robustness to cope with various scenarios.