Research On Weakly Supervised Learning Fine-grained Vehicle Recognition Method

One research direction in intelligent transportation systems is vehicle model identification.Traditional vehicle model recognition cannot represent high-level semantic information,and coarse-grained vehicle model recognition cannot meet actual needs,while fine-grained vehicle model recognition aims to identify vehicle brand,series,model,and year information through data,so it is widely used in industry demand.At present,the fine-grained vehicle model recognition algorithm mainly has the following difficulties and challenges: On the one hand,this type of algorithm relies heavily on information such as the target bounding box and part marking points,which brings a lot of labeling costs;On the other hand,fine-grained vehicle image recognition generally has many types,small differences between categories,and classification accuracy needs to be improved.Therefore,related research needs to be further improved.In view of the above problems,this article carries out the following research work:1.In order to improve the effectiveness of the network for feature extraction,this paper combines the advantages of SENet and Dense Net121-BC and proposes SE-Dense Net-BC.It uses feature re-calibration strategy to strengthen the useful information for classification.Based on the combination of SENet and Dense Net121-BC,three variant structures were proposed,which can improve the classification accuracy to a certain extent.2.In order to reduce the cost of manual labeling,this paper proposes a method for the overall and local detection of vehicles under weak supervision.Only the image label information is used to obtain the target’s overall and discriminative part detection information on the convolutional activation map.3.In order to further obtain the discriminative local features of the vehicle,two-level attention data augmentation is used to train the network.Bilinear pooled attention maps can achieve weakly supervised target localization,and attention cropping can enlarge image detail areas.In addition,contrast loss and classification loss are designed to constrain parameter learning to improve fine-grained vehicle identification tasks.Experiments show that the proposed SE-Dense Net-BC as a basic network for feature extraction can enhance the extraction of useful information for classification.Among them,SE-Dense Net-BC variant two obtained the highest on the Stanford Cars-196 and Cifar10 dataset.The accuracy rates are 88.56% and 95.08%,which are better than the Dense Net121-BC benchmark network.In addition,When the SE-Dense Net-BC model is consistent,this paper proposes a weakly supervised attention target detection and attention cropping data enhancement strategy.The classification rate on the Car＿Bayonet dataset is increased from 84.48% to 89.86%,verifying that the strategy proposed in this paper can further improve the classification results.However,further research is needed on the problem of fine-grained vehicle identification in complex environments and small sample data.