Research And Implementation Of License Plate Recognition Algorithm In Natural Scene Based On Deep Learning

With the continuous development of intelligent transportation system,China’s traffic management is also in continuous progress,has been widely installed in many cities integrated automatic license plate recognition system.License plate recognition is an important part of intelligent transportation and often plays a decisive role.The current application of license plate recognition is mainly carried out under fixed conditions,such as license plate recognition of vehicles in and out of the gate of the community.With the expansion of the scope of applicable scenarios,license plate recognition technology is facing more problems.At present,the image acquisition technology of camera has been quite mature,but when the demand of the application scene expands,the license plate location and recognition technology still needs to be further developed.With the rapid development of artificial intelligence and deep learning,computer processing speed and computer vision problems have been a great breakthrough,but also to the traditional license plate recognition technology has been transformed.The vehicle license plate detection and recognition through deep learning technology has higher robustness and adaptive ability.The research scene of this thesis is the license plate recognition of vehicles parked on the side of the street,focusing more on the license plate recognition in the natural scene of the street under the high level camera.License plate recognition is divided into three modules: license plate detection,license plate correction and license plate recognition.Specific contents of the research on street parking identification with high-elevation cameras in this thesis are as follows:1.The main characteristics of the domestic license plate are studied,and the data set of the license plate is completed by combining three parts of the data,which is mainly composed of the large license plate data set of CCPD and the collected street scene data set.In these two parts of the data,there are problems of uneven distribution of Chinese characters and lack of characters in most provinces.To solve this problem,the license plate data is synthesized to increase the proportion of Chinese characters and alleviate the lack of Chinese characters.2.The related technology of license plate detection is studied.The feasible detection algorithm based on Open CV cascading classifier and the improved detection algorithm based on YOLOv2 are determined according to the requirements of recognition scenes.The YOLOv2 network is improved to better adapt to the single target detection task,and a multi-scale fusion mechanism is added to improve the detection rate of small-scale license plates.The two algorithms were compared and analyzed,and a better license plate detection method based on YOLOv2 was selected to realize the license plate detection algorithm.3.In terms of license plate character recognition,a one-stage method is adopted to realize the recognition of license plate of variable length.In this thesis,based on CRNN network and LPRNET network,combined with the characteristics of these two network structures,the license plate recognition network in this thesis is constructed,and the variable length license plate recognition is realized by full convolutional network plus CTC loss function.The improved CNN+CTC network adopts long volume kernel instead of RNN,simplifies the network structure,and uses 1*1 convolution kernel to reduce dimension to character class number,which can realize input multi-scale image recognition.In the performance test,the improved network has achieved a good balance in the recognition accuracy and speed,which is higher than LPRNET in accuracy and faster than CRNN in speed,and can meet the requirements of street scenes in this thesis.In order to solve the problem that the license plate is inclined,the possible tilt Angle of the license plate is analyzed,according to the characteristics of three-dimensional space deformation,the perspective transformation method is used to correct the license plate,and the experimental results show that the correction algorithm can improve the sequence recognition accuracy of 4.6% for the comprehensive type of blue plate pictures.Finally,this thesis integrates three algorithms of license plate detection,license plate correction and license plate recognition to achieve a complete license plate recognition process.By verifying the working process of the whole system algorithm,it proves the feasibility of its task in street scene recognition.