License Plate Recognition And Vehicle Type Identification Based On Deep Learning

In practical monitoring scenarios such as intelligent transportation and unmanned driving,license plate recognition and vehicle model recognition are a very common demand,and also one of the key challenges.With the development of national modernization and the needs of traffic supervision,these issues have become the research hotspot in China.On the basis of studying the basic theoretical algorithm of license plate recognition,this paper mainly studies the application of deep learning in intelligent transportation,and improves some traditional algorithms and neural network framework to obtain the recognition and classification of low-quality pictures.The main research contents of this paper are as follows:Firstly,the basic theory of image processing in license plate recognition is studied.In order to improve picture quality and facilitate later character recognition,an improved adaptive binarization method is adopted in this paper.Through adaptive threshold selection,this method has better binarization performance when the local illumination of vehicle pictures changes greatly.In addition,based on the analysis of license plate location algorithm,a joint positioning method based on texture features and OTSU method is used for coarse license plate location and the connected domain analysis based on morphology for precise positioning.In order to verify the effectiveness of these license plate preprocessing methods,the recognition results are verified by MATLAB simulation.Secondly,in the problem of license plate recognition based on BP neural network,the momentum factor is introduced to improve BP neural network to avoid the possibility of the network falling into local optimization.A vehicle license plate recognition method based on deep convolutional neural network is used,which can distinguish Numbers,letters and background images.Based on lenet-5 deep convolutional neural network,this paper USES an improved deep convolutional neural network structure.The network has five convolution layers and two full connection layers.The results show that compared with BP neural network,the improved convolutional extraction neural network can effectively reduce the error rate and recognition time.Finally,the genetic algorithm is used to detect the vehicles in the crosswalk.Through the analysis of the fitness function,when the fitness is greater than 0.1,there is an illegal stop of the vehicle.Simulation results show that this method has the advantages of fast identification time and real-time monitoring system.In the model recognition,this paper USES two neural network joint recognition method.The first convolutional neural network is used for vehicle location to process images of multiple vehicles and smaller vehicles with complex backgrounds.The second convolutional neural network is used for vehicle identification,and two loss functions are used to train the network jointly.This method has good results for vehicle identification of complex background.