Research And Implementation Of License Plate Recognition Technology Based On Deep Learning

License plate recognition is one of the core elements of intelligent transportation systems and has been widely used in various scenarios such as roads,car parks,communities and highway toll booths,greatly enhancing the efficiency of traffic management.However,with the continuous expansion of the application field of intelligent transportation system,the license plate recognition technology also faces various challenges,such as light intensity,rain and snow and other complex weather,camera shake or blurred license plates generated under high-speed movement and other factors will increase the difficulty of license plate recognition.Traditional licence plate recognition technology is mainly based on some subjective features of the licence plate,and its robustness and generalisation ability are relatively poor.With the rapid development of deep learning,its powerful feature extraction capability has brought a turnaround to the license plate recognition technology.Based on deep learning technology,this thesis conducts an in-depth study on license plate recognition algorithms in complex environments,and the main research work is summarised as follows:(1)A lightweight license plate detection model,GEG-YOLOv4,is proposed,using the target detection network YOLOv4 as the basic framework.Secondly,the ECA attention module,which can avoid dimensionality reduction and effectively capture cross-channel interaction information,is incorporated into the backbone network to increase the channel weights of the license plate information and reduce the interference of the complex environment background on the number plate information.Finally,the Ghost Module is used in the deeper network to retain the redundant information in the feature map,which further increases the detection accuracy of the model while reducing the number of model parameters.(2)A license plate character recognition technique is investigated and a license plate character recognition model GLPRNet is proposed.the model is based on the number plate recognition network LPRNet and firstly a spatial transformation network STNet is used to correct the skewed license plate before performing character recognition.Secondly the ECA attention module is firstly invoked in LPRNet and STNet to enhance the feature representation of the number plate characters.Next a deep over-parameterised convolution DO-Conv was used instead of partial traditional convolution,increasing the parameters that the network can learn and enabling the convolutional layers in the network to be enhanced.Finally the GELU activation function is used to further improve the feature representation of the network.(3)Enrichment of licence plate character data.Based on the characteristics of the licence plate characters,three aspects of CCPD dataset,network collection and manual generation are used to combine into licence plate character recognition data to alleviate the lack of some Chinese characters.Finally,this thesis integrates the improved license plate detection module and character recognition module to build a complete license plate recognition system,and the system performance is tested on number plate images in five complex scenarios.The experimental results show that the method proposed in this thesis has good robustness in complex environments and can meet the needs of practical scenarios.