Research On Fine-grained Vehicle Recognition Method Based On Deep Learning

Fine-grained vehicle recognition technology is the key content of intelligent traffic system,which plays an important role in traffic management,traffic flow monitoring and vehicle statistics.Due to the small feature difference between the same car series in the problem of fine-grained vehicle recognition,it is difficult to identify it,how to heighten the recognition method to express the discriminative features of the vehicle has become a research difficulty in the area of fine-grained vehicle recognition.This paper studies and analyzes the residual network and dense convolution network in the deep convolution neural network,two fine-grained vehicle recognition methods based on IC-Res Net combining independent components and Res Net152 and PF-DenseNet based on part-focused,were proposed.1.Aiming at the problems of low accuracy of existing fine-grained vehicle recognition and difficulty in accurate identification between similar vehicles,Based on the Res Net152 deep learning network,an IC-Res Net vehicle recognition method combining independent components and Res Net152 was proposed.Reduce feature loss by improving the Res Net152 downsampling layer,an independent component is proposed to be embedded in Res Net152 to construct a relatively independent network structure and enhance the ability to acquire vehicle characteristics,combined training of central loss and Softmax loss function was used to Improves class cohesion.Experiments on the public data set of Stanford cars-196 show that the model has higher recognition accuracy among similar sub-cars,better than the original Res Net152 network,and has a higher recognition rate.2.Aiming at the problem that the distinguishing parts of fine-grained vehicle models cannot be effectively identified under the deep learning method,a PF-DenseNet vehicle recognition method based on part-focused is proposed.Combining the characteristics of vehicle component features and deep learning methods,DenseNet is improved,and a processing layer is used to repeatedly convolutionally extract vehicle features to obtain clearer feature maps and obtain more distinguishing information.At the same time,it is embedded in the network architecture Independent components further enhance the model’s ability to extract features of vehicle models.Experiments on the Stanford cars-196 data set showed that the recognition accuracy of this method reached95.0%,achieving the best effect of fine-grained car recognition.Aiming at the problem that the recognition accuracy of fine-grained vehicle is not high and the discrimination is mainly concentrated on the vehicle parts,the IC-Res Net combining independent components and Res Net152 and the PF-DenseNet recognition method based on part-focused are proposed.Through experiments on the Stanford cars-196 data set,the results are relatively high.The accuracy proves the helpful of the two methods.