Some Key Issues Of Vehicle Model Recognition Based On Deep Learning

With the rapid progress of artificial intelligence technologies such as deep learning,the intelligent transportation system(ITS)has made breakthroughs.As the one of the most basic parts of the ITS,Vehicle Model Recognition(VMR)plays an important role in intelligent transportation,intelligent security and intelligent charging,et al.The accurate recognition of vehicle models is different from the classification of car brands and types,and it belongs to the category of fine-grained classification.Thus how to improve recognition accuracy and efficiency is an urgent problem to be solved.Based on the deep learning method,this paper studies the influence of dataset attributes on vehicle identification accuracy and efficiency,then we propose the concept of angle compactness of vehicle model recognition training set,and create the MTV-1638 s dataset for the Chinese market.Based on ResNet-50,the multi-stage learning network MS-CNN model identification method is proposed.Numerical experiments show that the algorithm can obtain better identify results.The research contents are as follows:(1)Analyze and propose the angle compactness of the vehicle model recognition training set and create a corresponding VMR database.First of all,for the identification of the Chinese market vehicle models,a more reasonable car dataset MTV-1638 s was created,including the 1638 models commonly found in the Chinese market.Through classification analysis,we found that the training set number and angle of samples of a single vehicle in the training set have a great influence on the recognition effect.To this end,this paper creates an average angle sampling model data set ASMTV120 s,on the basis of which the influence of the vehicle angle distribution on the recognition rate is studied.The experiment shows that the single model of the training set contains only sample images of no more than 18 specified angles.Using our result,the VMR method can achieve the same recognition result and can reduce the training time by 41.18% as well.(2)A multi-stage neural network MS-CNN model identification method for enhancing high-order information is proposed.There is a big gap between the inner-class and the small gap between inter-class.If it is possible to make full use of the high-level network features which rich in semantic information,it will be of great benefit for vehicle identification.Based on ResNet-50,this paper proposes a multi-stage neural network MS-CNN that enhances high-order information for vehicle model recognition.Numerical experiments show that the network can extract richer semantic features and geometric information,and can obtain higher recognition accuracy than VGG-16,ResNet-50 and B-CNN,et al.