Research On License Plate Recognition Algorithm In Complex Scenarios

With the development of society,the number of domestic motor vehicles is increasing,which brings great challenges to traffic management.License plate is the most important attribute of vehicle,so license plate recognition(LPR)algorithm plays a key role in intelligent transportation system(ITS).In addition to being widely used in access control systems,license plate recognition algorithms have also been used to recognize license plates in images taken by Unmanned Aerial Vehicles(UAVs)and mobile phones in recent years,which means that the algorithm will face more complex scenarios than before.LPR algorithm in complex scenarios is studied.1.License plate generation algorithm based on GAN is studied,CycleGAN is used to transforming category equilibrium synthetic license plate data into real data domain.The problem of label errors of generated license plate data is solved.CycleGAN is used to generate a large number of class-balanced license plate data,it reduces the serious imbalance of character classes faced by the license plate sequence recognition algorithm,and brings considerable performance improvement to the license plate sequence recognition algorithm.2.License plate detection algorithm based on SSD is studied.The ratios of default box and the size of convolution kernels are improved according to the width-to-height ratio of license plate and default box is densified in vertical direction,which greatly improves the recall rate of license plate.Then feature pyramid network is used to fuse the deep layers' feature with the shallow layers' feature.The improved feature extraction network reduces the missing detection rate of small license plate and improves the performance of the detection algorithm at complex scales.In order to reduce the computational load of the feature extraction network,the feature extraction network is replaced by MobileNet and then a real-time detection model is obtained.3.License plate sequence recognition algorithm based on multi-label is studied.The advantages and disadvantages of various license plate sequence recognition algorithms based on deep learning is analyzed.Finally,multi-label network is used to design an end-to-end license plate sequence recognition algorithm.The algorithm has the potential to recognize various types of license plates.In order to recognize the license plate with complex deformation,Spatial Transformation Network(STN)is used to automatically correct the license plate.According to the problems encountered in the training process,the training skills of STN are summarized.Finally,Attention mechanism and Focal Loss are used to improve the recognition of confusing characters.