强逆光极端环境下基于动态事件相机的车牌识别OA

In view of the the problems of the decline in license plate recognition performance caused by the lim-ited dynamic range of traditional frame-based cameras under strong backlight,an end-to-end license plate recogni-tion method based on dynamic event cameras is proposed.Dynamic event cameras have the advantages of a dynamic range greater than 126 dB,only outpacing brightness changes,and overcoming the failure of frame-based imaging.The event stream of license plates with strong backlight is collected by using the self-developed camera.The number of events is mapped to grayscale images through the cumulative sliding time window to construct the first event dataset DVS-PD(Dynamic Vision Sensor Plate Dataset)for license plate recognition with strong backlight.The recognition network adopts the STN(Spatial Transformer Network)to correct the tilted license plate,and the back end uses the LPRNet(License Plate Recognition Network)to output the character sequence.In the training section,the CTC(Connectionist Temporal Classification)loss and data augmentation are introduced,and the two denoising preprocess-ing methods of STCF(Spatio-Temporal Correlation Filter)and BES(Background Event Suppression)are compared.The experimental results show that the recognition accuracy rates of the proposed model in the three types of event frames,namely the original,STCF and BES,are 93.4%,94.5%and 95.1%respectively,which are superior to the LPDRNet(License Plate Detection and Recognition Network),the RCNN(Recurrent Convolutional Neural Network)and the SLPNet(Sequence License Plate Recognition Network).It has been proved that the dynamic event camera can effectively solve the problem of license plate recognition in strong backlight.The constructed dataset provides a benchmark for subsequent research.The proposed model combines lightweight and high inference speed,and can ef-fectively meet the scene requirements of real-time license plate detection and recognition.