Research On Person Re-identification Method Based On Multi-branch Network

With the fast development of deep learning and the widespread application of person re-identification in the fields of video surveillance,autonomous driving,and public safety,person re-identification based on deep network has become a research hotspot in the field of computer vision.The person re-identification tasks always have the problems of pose,occlusion,and view angle change.A new research direction which builds multi-branch network to learn a comprehensive,robust and discriminative feature representation can better solve these problems.The paper studied the person re-identification method based on multi-branch network from loss function and network structure.The main three works are as follows:1.Improve the multi-loss function method of person re-identification.The triplet loss of traditional multi-loss fusion method only increases the inter-class distance and ignores the absolute distance on inter-class distance.To solve this problem,this paper proposed use center loss improve the multi-loss fusion method.This method effectively solves the problem of that triplet loss ignores the absolute distance on inter-class distance when converging features,and center loss further improves the discrimination of features.This work was experimentally verified in the Market1501 dataset.Experimental results show that the proposed method has better advantages than the traditional multi-loss fusion method.2.Improve the neck structure in multi-branch network with the multi-loss.The branches in Batch DropBlock Network have different cha racteristics,and multiple loss functions interfere with each other during training,so it is necessary to find the suitable loss function and neck structure for each branch.To solve this problem,this thesis proposes to use the center loss and BNNeck structure to impro ve Batch DropBlock Network,and explores the impact of BNNeck and center loss on different branches of Batch DropBlock Network.BNNeck separates the triplet loss and Softmax loss into different and suitable feature spaces,it makes the triplet loss and Softmax loss don't interfere with each other during training,and the center loss further enhances the discriminative of the model.This work was experimentally verified in the Market1501 dataset,CUHK03 dataset and DukeMTMC dataset.Experimental results show that the proposed method improves the robustness and classification accuracy than Batch DropBlock Network.3.A new multi-branch network structure is proposed with comprehensive and discriminative feature representation.Person features extracted from traditional network only contain global information,and ignore the local fine-grained information.To solve this problem,this thesis proposes to use Batch DropBlock Network and Multiple Granularities Network to extract the local fine-grained features.This network not only contains edge information of person by Batch DropBlock branch,but also extracts the local features of different granularities through the branches of the Multiple Granularities Network,finally it obtains a more comprehensive,robust and discriminative feature representation.This work was experimentally verified in the Market1501 dataset.Experimental results show that the proposed method has better performance than the existing methods.