Research On Person Re-identification Algorithm Based On Local Information Interaction Enhancement And Inter-domain Fusion And Intra-domain Style Normalization

As the central technology of intelligent surveillance,person re-identification plays an increasingly important role in society.Person re-identification technology aims to retrieve other images with the same identity as a specific query image at different times and under non-overlapping cameras,and is one of the research topics in the field of artificial intelligence.However,supervised person re-identification mainly faces challenges such as large differences in pedestrian gestures,camera style variations,and occlusion of pedestrian body parts.In addition,unsupervised person re-identification faces inter-domain differences caused by inconsistent data distribution between source domain and target domain and insufficient feature extraction ability caused by intra-domain style differences between samples.To address the above issues,this paper proposes two models to improve the performance of supervised and unsupervised person reidentification.The main work is as follows:(1)In order to simultaneously mine the salient information and potential significant information of pedestrian images,we propose a Local Information Interaction Enhancement Network(LIEN).The network integrates attention mechanism with local-based strategy to achieve mutually facilitating and complementary effects.Specifically,LIEN is composed of three parts: a backbone network,a Dual Attention Module(DAM),and a Local Information Interaction Module(LIM).DAM is used to guide the backbone network to extract the local salient information,and LIM is used to mine the potential significant information that DAM has ignored by exploring the correlations between different channel-level local features.With the combination of DAM and LIM,LIEN can mine the salient features while obtaining more potential effective features,and effectively avoid pedestrian misalignment caused by uniformly dividing pedestrian images.(2)To simultaneously weaken the effects of inter-domain differences and intra-domain style differences between samples,we propose an Inter-domain Fusion and Intra-domain Style Normalization Network(DFDSN-Net).Specifically,DFDSN-Net mainly consists of a backbone network,an Inter-domain Fusion Module(DFM),and an Intra-domain Style Normalization Module(DSNM).DFM helps the network to gradually adapt to the data distribution between source domain and target domain and weakens the impact of inter-domain differences by deeply fusing the source domain features and target domain features.DSNM improves the feature extraction ability of the network by eliminating the intra-domain style differences between samples while focusing on the useful information eliminated by instance normalization.With the combination of DFM and DSNM,DFDSN-Net can simultaneously weaken the effects of inter-domain differences and intra-domain style differences between samples,and improve the performance of unsupervised person re-identification.(3)To verify the effectiveness and advancedness of LIEN and DFDSN-Net,we conducted extensive ablation experiments and comparison experiments on the Market-1501,Duke MTMC-Re ID,MSMT17,CUHK03 and Person X dataset.The experimental results show that both LIEN and DFDSN-Net achieve high accuracy compared with the state-of-the-art person re-identification methods.