Research On License Plate Recognition Algorithm Based On Deep Learning

With the increase in the number of vehicles in the world,intelligent transportation technology has become an important means of intelligent management of urban traffic.Automatic License Plate Recognition(ALPR)has been a hot research object in intelligent transportation technology and is widely used in many traffic scenarios,such as vehicle management on expressways,vehicle recognition,parking lot access management,and even traffic scene understanding.ALPR involves in object detection,image processing,pattern recognition,and other fields,which provides great help for reducing traffic management burden,reducing traffic violations and traffic accidents,and building smart cities,and has a huge application prospect for public safety.In natural scenes,the geometric deformation,rotation,and distortion of the license plate caused by the side-view camera angles or the actual existence of the license plate,the weak or strong light in the shooting environment,the contaminated characters,and so on,which will affect the accuracy of ALPR.The deep neural network has made many achievements in the research of license plate recognition,and various deformation data will greatly affect the training and recognition accuracy of the deep network.To this end,an overall ALPR scheme is proposed.The main research contents and innovation points are as follows:To solve the problem in which the geometric deformation of the license plate affects the text recognition accuracy,a space transformation network(STN)with thin-plate-spline(TPS)transformation is introduced,and a new Inverse Compositional Spatial Transformer Network(ICSTN)is developed for rectifying license plate images.ICSTN can be trained by unfolding the architecture multiple times into the form of ICSTN＿s,where s represents the number of warp operation unfolded.Through a large number of experiments on the CCPD dataset and PKUData dataset,the model with the best performance of rectifying license plate image is determined,that is,the value of s is determined.A new deep neural network ICSTN-CRNN is constructed to solve the problem of license plate recognition.ICSTN-CRNN is a one-pass end-to-end neural network composed of a license plate rectification network ICSTN and a Convolutional Recurrent Neural Network(CRNN).This method significantly improves the accuracy of license plate recognition.The accuracy of license plate detection is crucial for subsequent recognition.Based on the existing object detection algorithm,a license plate detection algorithm based on deep learning is implemented,and a novel real-time ALPR solution is proposed.The solution integrates real-time license plate detection,license plate recognition network and data enhancement strategy.Numerous experiments conducted on existing benchmarks.Compared with other existing methods,the proposed solution can improve the accuracy by about 1%-16% and has the good real-time performance.