The Research Of Vehicle Model Fine-grained Recognition

With the increase of domestic car ownership,how to manage the traffic on the road in a timely and efficient way has become the primary challenge for the transportation sector.At this time,the technology of vehicle feature recognition based on video and image has gained wide attention.As one of the indispensable features of the vehicle,the vehicle model has a very important significance for its identification.A wide range of vehicle models,different types of vehicles have little difference in appearance and the vehicle image is easily affected by the environment,which greatly limits the effect of fine-grained car recognition.In view of these problems,two methods are proposed in this paper based on convolutional neural network:Fine-grained classification of car model using Fg-CarNet convolutional neural network and a multi view vehicle classification method based on embedding posture information model.Experiments on a large vehicle image dataset show that these two methods can effectively improve the precision of vehicle type identification.The main work of this paper is as follows:(1)Analyze the convolution neural network method:Analyze the algorithm principle and training process of the conventional neural network,and further introduce the principle of the convolution neural network forward process and backpropagation process.Then we introduced the vehicle fine-grained classification data set.(2)A vehicle classification model called Fg-CarNet is proposed.According to the fact that the angle of the image taken in the bayonet is positive,and the characteristics of the upper and lower parts of the vehicle image are unevenly distributed,a multi-branch and multi-dimension feature fusion CNN model Fg-CarNet is proposed.This model fuse the features of the upper and lower parts of the vehicle image in a multi-dimensional way,and achieve good results on the vehicle data set captured at the bayonet.(3)A vehicle fine-grained classification method based on embedded posture information is proposed.Firstly,the posture estimation network is used to estimate the posture of the vehicle,including the position information of the vehicle in the image and the shooting angle of the vehicle.Then,the perspective of the vehicle are used to recalibrate the vehicle features extracted from the classified sub-network,and the classification features and posture features are merged.All these enhances the features that contribute to fine-grained vehicle classification.Experiments on multi view vehicle data sets show that this method can achieve very good results.