Research On Vehicle Recognition Based On Convolutional Neural Networks

With the continuous development of urban transportation system,intelligent transportation system has become an important research direction of traffic management system.Vehicle recognition,as an important branch of intelligent transportation system and key technology,has also achieved widespread attention.Previous vehicle recognition mainly focused on rough recognition of vehicles,but in many practical applications,we need vehicle make and model automatic recognition.So the fine automatic vehicle recognition technology is also more and more important.This paper mainly studies the fine vehicle make and models recognition based on image content,including the research of vehicle image foreface detection and fine vehicle recognition.Firstly,the vehicle detection method based on convolution neural networks(CNNs)is studied and realized.On this basis,study of the precise vehicle make and model recognition in this paper is carried out in the following aspects:(1)In order to learn representative features and improve the recognition accuracy,a method based on multiscale layer-skipping CNNs is proposed.Construct multiscale layer-skipping CNNs and use an adaptive fusion method to adjust the contribution of different networks.The recognition results of some single scale layer-skipping CNNs are fused.The experimental results show that the multiscale layer-skipping CNNs can improve the accuracy of recognition.(2)Focused on the issue that in the latter part of training the recognition accuracy of CNNs increases slowly and the training process is time-consuming,a reinforcement learning error training samples and error prone training samples CNNs recognition algorithm is proposed.The algorithm uses the advantage that the network error rate decreases rapidly in the initial epochs of training.In the initial stage of training,reduce the training epochs and after each training,error samples and error prone samples are updated to the training set.Experimental results show that in the process of training learning error samples and error prone samples repeatedly improves the recognition accuracy and reduces the training time of CNNs.