Research Of Vehicle License Plate Recognition Based On Joint Optimization Network In Fog-haze Environments

Vehicle License Plate Recognition(VLPR)is a technology that can detect vehicles on monitored roads and automatically extract license plate information.This technology greatly facilitates the road traffic management and is one of the core technologies in intelligent transportation systems.It has been widely used,such as parking lot management,highway toll station monitoring,and urban traffic supervision.However,the stability of license plate recognition technology is typically broken by severe extreme weather(such as fog-haze,rainstorm,sandstorm),resulting in a significant reduction in recognition accuracy.In this thesis,aiming at the problem of license plate recognition in haze environments,the research based on a joint optimization and a cascade architecture is carried out.This thesis firstly investigates the relevant knowledge in the field of image dehazing and deep learning,and lays a theoretical foundation for the model design of the cascade architecture.Secondly,it analyzes the difficulties and challenges confronted by the industrial implementation,and provides some guidances for the optimization scheme in practical applications.Finally,based on the above theoretical knowledge and research conclusions,a technical framework for the license plate recognition in haze environments based on a joint optimization and a cascade architecture is proposed.In order to obtain the road images with higher resolution in haze environments,this thesis firstly places the image dehazing algorithm at the start of the cascaded architecture,and proposes a pre-processing dehazing algorithm based on a fog concentration refinement.The algorithm uses local atmospheric light values instead of global atmospheric light values,which can produce a finer transmission map.In addition,considering the accumulation of errors caused by parameter estimation in the pre-processing dehazing algorithm and in order to comprehensively optimize the license plate recognition reslults,this thesis also adopts a joint optimization which combines deep dehazing and license plate objectdetection,i.e.,the image refinement and dehazing process is embedded in the object detection module of the architecture,and a joint optimization model(A Joint Further-dehazing and Region-extracting Model,JFRM)based on convolutional neural network for deep dehazing and license plate region extraction is proposed.This model can significantly reduce the image restoration distortion caused by the cumulative error,and can synergistically improve the license plate detection results.Then,super-resolution reconstruction is performed on the detected license plate area to avoid the influence of motion blur on subsequent character recognition,and finally the license plate characters are obtained through the character recognition technology.Through experiments,we verify the feasibility of this model in the cascaded architecture and the positive effect on the accuracy of license plate recognition in haze environments.From the perspective of industrial implementation,the stability of the image dehazing module in the cascaded architecture is still poor,and the dehazing performance of the natural haze environments may be much worse than that of experimental results.For practical challenges,this thesis proposes Dehaze Cycle Net,a dehaze network based on the Cycle-GAN style transfer.Dehaze Cycle Net uses unpaired real natural images(including natural haze images and natural hazefree images)to replace the paired images used in experiments(including synthetic haze images and natural haze-free images)for unsupervised training,which improves the generalization ability of the dehazing network especially in the real environments,and thus greatly improves the practicability and environmental adaptability.Finally,based on the above research results,this thesis also implements a prototype system for license plate recognition in fog-haze environments.The system relies on the development of C/S architecture and Vue framework to realize the output function of license plate recognition in foghaze environments and the independent image dehazing function.The results show that when the prototype system provides a visual interactive interface,the data information corresponding to the users can be effectively managed,which is convenient for the queries and exports for users.This thesis combines a joint optimization and a cascade architecture to realize the license plate recognition in the fog-haze environments.However,from the perspective of network complexity and forward reasoning time,the cascaded architecture solution will face the challenges of industrial real-time requirements.Therefore,it is necessary to consider themodel compression in future work to facilitate the implementation of real-time applications.