Research On License Plate Recognition Of Complex Scenes Based On Deep Learning

In recent years,with a significant improvement in people’s living standards,the total number of motor vehicles in our country has also increased rapidly.Traffic congestion and red-light running related to vehicles have become increasingly common.Therefore,how to scientifically and efficiently manage the vehicles involved in road traffic has become a problem.To effectively solve this problem,it is first necessary to identify the identity of the vehicles,that is,to recognize the license plates.Most traditional license plate recognition methods are designed for relatively fixed and restrictive scenarios such as lighting,shooting distance and shooting angle.However,with the decrease in the price of hardware and services,as well as the application of handheld devices in the traffic management system,and the ways and scenarios of obtaining license plates become more diverse.Therefore,the captured image may exhibit different distance,angle and light level of the license plate picture,and even the individual license plate characters may be obscured.Traditional methods are unable to handle license plate recognition tasks under these complex conditions.The emergence of deep learning technology provides a new approach for recognizing license plate images in complex scenarios.This paper researches license plate recognition tasks in complex scenes based on deep learning technology,improves two license plate recognition methods through deep learning technology,and increases the recognition accuracy and speed of license plate recognition algorithms in complex scenes.The main work is as follows:(1)An improved real-time license plate recognition algorithm with a multi-label classification structure is proposed.The algorithm enhances the accuracy of license plate recognition by improving the accuracy of character position feature extraction through channel attention mechanism and increases the speed of license plate recognition by reducing the computational complexity through lightweight functions.A shallow network is designed for feature extraction,which avoids feature redundancy and computational speed degradation from too deep a network while fully extracting license plate features.A lightweight channel character localization network is also designed,which locates license plate characters through squeeze and excitation attention mechanism and fuses the deep and shallow features of the network to improve character localization accuracy while extracting semantics.Additionally,the algorithm uses depth wise separable convolutions and lightweight functions to replace computationally expensive steps in the network,thereby improving license plate recognition speed.The algorithm achieves an average recognition accuracy of 98.0%and an average recognition speed of 93.8 FPS on CCPD,the largest open source license plate dataset in mainland China,outperforming the compared algorithms in terms of accuracy and speed,indicating significant advantages of the algorithm in license plate recognition tasks.(2)In order to further improve the accuracy and speed of license plate recognition,this paper proposes a license plate recognition algorithm based on an improved coding and decoding structure,using a parallel inference decoder to recognize license plates,changing the license plate recognition logic from serial to parallel to further improve the accuracy and speed of license plate recognition.In the feature extraction part,an improved feature extraction network is designed,which improves the integrity of feature extraction by adding a Focus structure and reducing the loss of feature information during downsampling,thus improving the accuracy of license plate recognition.In the decoding part,a parallel decoder is used,which only needs to infer once in the forward inference process,improving the computational efficiency of the algorithm.Moreover,the multihead attention in the parallel decoder can accurately locate the characters in the license plate,so the license plate image input into the network can be accurately recognized without correction,saving computing time.The algorithm achieved an average accuracy of 99.4%,93.1%,and 95.7%on the CCPD,CLPD,and PKUData datasets in mainland China,and an average accuracy of 98.5%,99.1%,and 96.1%on the AOLP-AC,AOLPLE,and AOLP-RP datasets in Taiwan,China,respectively.The average recognition speed on these datasets reached 159.8 FPS,surpassing the performance of compared algorithms in both accuracy and speed.