Research On Key Technologies Of License Plate Recognition In Low Light

With the continuous growth of car ownership,intelligent transportation has penetrated into our daily lives.As an important part of intelligent transportation,license plate detection has also been developed rapidly,bringing great convenience to urban public transportation planning and travel..Due to the extremely complex environment of the car,it has brought great difficulties to the license plate detection and recognition.By analyzing a large number of license plate data,it is concluded that the light is the main factor that affects the license plate detection and recognition.Through the analysis of the actual problems in the license plate recognition scene and the data characteristics of the license plate image,this paper uses the convolutional neural network as the basic framework to focus on the low-light enhancement,license plate location and license plate recognition issues in the license plate recognition process,aiming to achieve The task of license plate recognition in a more complex(low-light)situation.Considering that the model used in this article is mainly composed of convolutional neural network or its deformation,the principle,structure and optimization method of convolutional neural network are introduced to lay the foundation for subsequent research methods.The main work of this paper is as follows:(1)In order to solve the problem of low license plate recognition rate in the case of insufficient light,low-light enhancement is performed on the license plate to ensure that the license plate obtains a better effect in the subsequent positioning and recognition.First,considering that the dark light enhancement model is a supervised network model,the license plate image that meets the requirements is processed with low brightness;then,a low brightness enhancement model based on convolutional neural network is constructed.The model is divided into brightness recognition module and noise The recognition module and the image enhancement module are three parts;finally,the CCPD data set is used to train the model and compare with other models to verify the effectiveness of the proposed method.(2)Aiming at the problem of license plate location,integrating the robustness of the algorithm and the generalization ability of the network,it is proposed to use the convolutional neural network to locate the license plate,and improve the network loss function,increase the aspect ratio information in the license plate frame,and make the position of the license plate more accurate.Precise positioning.The experimental data uses 80,000 license plate images in natural scenes in the CCPD dataset.The experimental results show that the improved convolutional network has a very high accuracy and recall rate for license plate positioning.After analyzing the wrong positioning data,low-light preprocessing of the license plate is introduced.The operation greatly improves the positioning accuracy of the low-light license plate,improves the robustness of the model and enhances the anti-interference ability of the environment.(3)In view of the problem of license plate recognition,firstly,the inclination correction of the license plate after positioning is performed,and then by analyzing the traditional character recognition algorithm and the convolutional neural network character recognition method,comprehensively considering the real life scenes used in the license plate,the end-to-end type of DenseNet network structure is selected.License plate character recognition.Through the unique dense connections in the network structure,the depth of the network can be deepened,over-fitting can be avoided,and the accuracy of model character recognition can be improved.The experimental results show that DenseNet can achieve 99.23% of the license plate character recognition accuracy,with extremely high accuracy and network generalization capabilities.Finally,each experiment is integrated,and the experiment is streamlined through the program to realize the automatic output of the car image input to the license plate character in the natural scene.The experiment shows that the accuracy of the license plate character recognition in the actual scene can reach 98.37%,which greatly meets the application in the real scene and verifies the feasibility of the experiment.