Analysis And Research On Vehicle Recognition Based On Deep Learning

With the deepening of the complexity of traffic situation,the workload of public security andtraffic police becomes more and more heavy,therefore,the further development of intelligent transportation system becomes particularly important in urban areas.Vehicle recognition system is very popular and has been widely applied in the market of intelligent transportation system,it depends on license plate recognition,vehicle-logo recognition,vihicle model recognition,etc..Among them,the research on license plate recognition has been mature and widely used to carport,uptown and other places.For the vehicle-logo recognition,as the logo area is too small,the image we needed to research usually should be of very high quality,if not,vehicle-logo recognition often cannot achieve a good recognition effect.And the traditional vehicle identification can only divide the vehicle into several major categories,like trucks,buses,cars and so on.With the faster development of intelligent,the identification of specific models of vehicles will be of great value in practical applications.Analogy to face recognition,this paper do a series of studies around the vehicle-face image.First of all,due to the lack of the vehicle-face image data which could be directly used for the study,we cut out the vehicle-face part of the vehicle image and normalize the scale of the vehicleface image.Use some operations,such as shift rotation,brightness adjustment and motion blur,to simulate the possible factors reality for data amplification,and ensure the diversity of data.Finally,we built a vehicle-face image database contains 31 sub-models.The detection algorithm of image feature is the core of image recognition method based on the image processing mode.Through analyzing and comparing various algorithms for feature extraction,our paper chose the SIFT feature extraction algorithm which is the most common and effective in the field of vehicle identification.We expounded its working principle in detail and used it to extract image feature.In this paper,the SUSAN corner detection operator also was introduced to improve the shortcomings of SIFT who generated too complicated descriptors.Through experiments,we successfully extracted simple and accurate features and the k-adjacent classifier is used to classify.Finally,this paper introduced the most popular method,convolution neural network,in the field of deep learning to study and extract the feature of vehicle images independently.To build a convolution neural network model,we starts form the aspects of network layer,size of convolution kernel,down-sampling method and type of activation function,to analyze the change of training time,the feature dimension and the feature extraction time of ?he convolution neural network under different network settings.Selecet the optimal value and determine the final convolution neural network model.In this paper,the convolution neural network model was combined with the classifier to experiment.Compared with SIFT algorithm and SUSAN-SIFT algorithm,the recognition method based on convolution neural network was improved in terms of speed and accuracy.