Research On Vehicle Fine-grain Classification Algorithm Based On Convolutional Neural Network

Since the rapid development of intelligent transportation systems,vehicle classification based on image has received more and more attention.This paper focuses on the fine-grained classification of vehicle images in natural scenes.The characteristics of these images are that the background is complex and the shooting angle is not fixed.The vehicle images also have high intra-class differences and low inter-class differences.The extraction of traditional image features,such as HOG,SIFT and so on,need to be done manually,and the extraction process is complex.Some image features are designed according to specific scene and have no universality.In recent years,deep learning becomes more and more popular.As a branch of deep learning,convolutional neural networks have been widely used in image classification because of avoiding designing features by people.Therefore,this paper proposes a vehicle fine-grained classification algorithm based on convolutional neural network to improve the classification accuracy of vehicle images in natural scenes.Firstly,this paper introduces the basic theory of convolutional neural networks,including the development history of convolutional neural networks and the basic structure of convolutional neural networks,and introduces two common classification models: Softmax classifier and SVM classifier.Secondly,for the low classification accuracy of traditional SIFT features in vehicle image classification,this paper proposes a vehicle fine-grained classification algorithm which fuses FV-SIFT and deep convolutional features.This algorithm combines the FV-SIFT feature and the VGG-16 deep convolutional feature and uses one-versus-rest classification method based on SVM to classify the vehicle image.The experimental results show that the VGG-16 deep convolutional feature is more representative than traditional FV-SIFT feature,and the fused feature also improves the classification accuracy.Finally,for the insufficient expression ability of single deep convolutional feature,this paper also proposes a vehicle fine-grained classification algorithm based on improved bilinear convolutional neural network.This method uses two different convolutional neural networks instead of the two identical convolutional neural networks.It also improves the square root solution method of the feature matrix.In this paper,we use the out product to fuse the VGG-M deep convolutional feature and the VGG-16 deep convolutional feature.The experimental results show that the bilinear convolutional feature compared with single deep convolutional feature improves the classification accuracy in vehicle image classification.This paper proposes a vehicle fine-grained classification algorithm based on convolutional neural network,which overcomes some problems existing in traditional methods.It uses high-autonomous learning ability of convolutional neural networks to extract advanced features,improves the accuracy of vehicle image classification and has better robustness.