Design And Implementation Of Vehicle Sub-brand Classificaltion System Based On Fine-grained Image

With the development of the automobile industry,more and more vehicles are beginning to appear in the streets and lanes of the city.While facilitating people’s travel,it also brings management inconvenience to urban traffic management personnel.The classification of vehicle sub-brands helps to provide more detailed management of vehicles for urban traffic managers.The vehicle sub-brand classification includes two stages:vehicle detection and sub-brand classification.The vehicle detection stage uses the vehicle detection algorithm to extract the vehicle foreground.The sub-brand classification stage uses the image classification network to classify the extracted vehicle foreground to obtain sub-brand information.The RCNN series-based vehicle detection algorithm used in the detection phase may be missed when performing small-scale vehicle detection.The image classification network used in the classification stage cannot learn the local features of the sub-category vehicle for sub-brand classification.In view of the above problems,this paper designs a vehicle sub-brand classification system based on fine-grained images.The system includes a vehicle detection algorithm based on a residual block and an improved candidate network,and a sub-brand classification network based on local features.Based on the vehicle detection algorithm of residual block and improved candidate network,the target detection algorithm Faster RCNN is used as the framework to design a convolution feature extraction network based on residual block instead of the commonly used ZF network of Faster RCNN and VGG network to extract vehicle convolution characteristics.To improve the network’s ability to learn shallow texture information of vehicles;a regional recommendation network based on multi-convolution candidate region generation network instead of Faster RCNN is proposed to generate multi-scale candidate regions to reduce the occurrence of small-scale vehicle detection.Missing detection.The candidate region generation network based onmulti-layer convolution also includes regression-based detection frame refinement network,using the full connection layer training weight matrix,and then refining the detection frame through the weight matrix,so that the detection frame can accurately cover the vehicle.Reduce the impact of background on the classification stage.Based on the VGG19 network,the sub-brand classification network based on local features is used to generate the vehicle saliency map using the convolution feature of the last layer of the VGG19 convolution layer,and then use the pooling layer to filter the vehicle saliency map to obtain the feature point with the largest response..Combined with the receptive field,it can be considered that the region corresponding to the point with the greatest response is the local region where there is a sub-brand vehicle difference,and the local feature training sub-brand classification model is extracted.In order to verify the performance of the vehicle detection algorithm based on the residual block and improved candidate network proposed in this paper and the design of the sub-image classification system based on fine-grained image,the test is carried out on Car196,Comp Cars and self-made data sets,and the current The mainstream three target detection algorithms are compared with five vehicle sub-brand classification methods based on fine-grained images.The experimental results show that the proposed vehicle detection algorithm based on residual block and improved candidate network can effectively reduce the missed detection of small-scale vehicle detection and the average accuracy,recall rate and accuracy of the single category of the algorithm are higher than other.Three kinds of target detection algorithms;the vehicle sub-brand classification system based on fine-grained image is better than the other five fine-grained image classification methods,which proves that the design of this paper is based on The vehicle sub-brand classification system of fine-grained images can effectively realize the vehicle sub-brand classification.