Design And Optimization Of Multi-branch Cooperative Deep Neural Networks For Person Re-identification

Deep Neural Networks(DNNs)have been successfully applied in person ReIdentification(Re-ID)due to its powerful expression capability for feature representation,which can learn the distinguishable feature of pedestrian.Therefor,how to design DNNs for achieving their best potential for the purpose of person Re-ID has important theoretical and practical significance.In recent years,it incurs considerable interests in how to design the multi-branch DNNs for learning rich feature representation in person Re-ID.By focusing on the application of deep learning algorithm in person Re-ID,the main work can be summarized as follows:(1)We propose a novel Asymmetrical Network(As Net)structure consisting of a global branch and a part branch.A simple asymmetrical structure between two branches ensures that features extracted by them are complementary for final matching.We also propose a lightweight version of attention module.The insertion of this lightweight attention module into backbone network further prompts the performance with only trivial increase of parameters.We validate the efficacy of our proposed mechanisms with extensive experiments over standard datasets of person Re-ID.In particular,As Net achieves the m AP accuracy of 89.13% on the Market1501,and 81.1% on the Duke MTMC-re ID,respectively.(2)We propose a branch-cooperative network,termed BC-Net,for person Re-ID.By stacking four cooperative branches,namely,a global branch,a local branch,a relational branch and a contrastive branch,we obtain powerful feature representation for person Re-ID.The analysis and experiment show that the person characteristics extracted by the four branches of BC-Net have good diversity and complementarity.BC-Net can be applied to different backbone networks,and OSNet and Res Net are considered as backbone networks for BC-Net verification in this thesis.Competitive performance has been achieved on several popular person re-identification databases.In particular,BC-Res Net achieves the rank-1 accuracy of 90.6% on the CUHK03-Labeled,and 86.6% on the CUHK03-Detected,respectively.The proposed BC-Net was used in the PRCV2020 mass person re-identification competition,which finished second in the competition.(3)On the basis of multi-branch structure,this thesis continues to study the further optimization of BC-Net with some adjustments on various micro-structures,including generalized-mean pooling,continuous Gaussian Dropout,attention modules of Batch Drop Block(BDB)and Relation-Aware Global Attention(RGA),etc,OSNet and Res Net are considered as backbone networks for BC-Net verification in this thesis.Experimental results also show that the optimized BC-OSNet achieves 89.9%,82.1%,84.2%,and 81.5% m AP on the four person reidentification datasets,including Market1501,Duke,CUHK03_Labeled,and CUHK03_Detected,respectively.This means that the optimized BC-OSNet surpasses BCOSNet about 0.6%,1.4%,1.1% and 1.7% in m AP for these datasets.