Research On License Plate Character Recognition Based On Deep Learning

With the development of economy and the progress of science and technology,the number of vehicles has gradually increased,this leads to the difficulty of vehicle management in society,how to manage vehicles effectively has become a popular trend.Our country's license plate technology has made some remarkable achievements for long time efforts.The traditional license plate recognition method has been widely applied to parking lots,toll stations,scenic spots,etc.The traditional license plate recognition method usually takes much time for feature extraction of license plate images,and often relies on manual feature extraction,which often fails to achieve good feature extraction results,at the same time,the license plate in the movement,the intensity of light is too dim,or the blurred definition of images caused by different angles,the license plate character tilt and so on,all these will affect the accuracy rate of license plate recognition.In this thesis,some traditional license plate character recognition methods based on machine learning are introduced,and the advantages and disadvantages of these machine learning algorithms are analyzed.Then,in view of the rapid development of AI and deep learning,deep learning has been widely used in computer vision,pattern recognition,data mining and so on.Therefore,this thesis uses the popular target detection method in deep learning to study the license plate character recognition.The main contents of this thesis include the following parts:First,we study the convolution neural network which is widely used in image processing direction in deep learning,and use the Faster R-CNN network model based on convolution neural network is used to carry out the experiment of license plate character recognition.In the process of training model,we use image annotation software to manually annotate two kinds of datasets.The dataset includes 10873 standard datasets and 4711 real scene datasets.In this license plate location experiment,the method of deep learning Faster R-CNN is used to compare with the traditional location method,such as color location,text location and SOBEL edge detection,it is verified that the localization effect of deep learning target detection method is higher than that of traditional methods.At the same time,in order to improve the accuracy of the character recognition of the license plate,the preprocessing of the license plate of the real scene,the size of magnification and the enhancement of the image are carried out.Finally,three different models in Faster R-CNN are used,and the best network model is chosen by training time efficiency on the recognition accuracy.Secondly,in order to further improve the accuracy of license plate character recognition,the method of deepening the number of network layers is used to extract the image features more completely,but if only the number of network layers is added and no other processing is done,the degradation will result in a decline in the accuracy of recognition.In view of this situation,we take another target detection network model R-FCN to identify and detect the license plate characters,and compare the experiments through different network layers,and discuss the influence of the number of network layers on the accuracy of character recognition.Finally,as the R-FCN model is the improvement of Faster R-CNN model,this thesis compares two different target detection network models to verify that R-FCN model is better than Faster R-CNN model in license plate character recognition.The experimental results show that the two different layer number residual network structure models of the R-FCN model are all higher than the three different models of the Faster R-CNN.