Fine-grained Vehicle Model Recognition Research Based On Deep Learning

With the number of urban cars continues to rise,the problem of traffic and environment has become more conspicious.In order to address these problems,Intelligent Transportation System has become the focus of urban development research.The fine-grained vehicle recognition technology has been proved to be the key point in the Intelligent Transportation System which has many advantaged superority to deal with the traffic problem,including the improvement of charging efficiency,the determination of traffic responsibility and the tracking of the escapes,and so on.It is of great significance to the construction of Intelligent Transportation System.However,the existing vehicle recognition methods mainly adopt low-level features which have large limitations in dealing with fine-grained recognition tasks.Nowadays,deep learning has advantage in high-level semantic feature extractiong as well as learning effective feature representation aiming at specific tasks from big data.Therefore,we propose an architetecture using convolution neural network to identify fine-grained vehicle in surveillance scene.The main work is as follows:Firstly,vehicles in surveillance scene images are always multi-target and multi-scale.In addition,the collection of pictures has the problem of disortion.After comparing the existing deep learning detection framework,we completed the vehicle detction and segmentaion in the use of Faster R-CNN.In this method,VGG16 is used as the network model to extract the feature,and finally the corresponding detection bounding box coordinates is obtained by linear regression.The related experiments show that this method has both strong robustness and good detection accuracy for vehicle detection in surveillance scene.Secondly,multitask deep learing neural network is employed to identify the vehicle.According to the complex properties of the vehicle,the network not only identify fine-grained vehicle model,but also recognize the vehicle color and type as two auxiliary tasks.After comparing with single task network,it can be seen that the auxiliay tasks make main task more accuratelv.During the experiment,we improve some deep neural network models which are pretrained and finetuned to a certain degree.The results of analysis show that ResNet-50 has faster convergence rate and higher recognition accuracy than other networks in the experiment,and the recognition rate of the 281 database models is 97.8%which shows its great model expression ability.