基于动态特征融合与特征感知的可见光-红外行人重识别方法OA

Visible-infrared person re-identification(VI-ReID)remains challenging due to cross-modal discrepancies and intra-class variations.While existing methods focus on feature space alignment for cross-modal shared representations,they often neglect modality-specific discriminative features such as visible textures and infrared thermal patterns.To address this limitation,a visible-infrared person re-identification framework via dynamic feature fusion and feature-awareness(DFFA)is proposed.Leveraging Transformer's long-range modeling capabilities,DFFA dynamically integrates local visible textures with global infrared thermal features through cross-modal attention,effectively preserving complementary modality-specific and shared characteristics.Our framework features a multi-dimensional loss design:a modality-fusion contrastive loss aligns local visible features,global infrared features,and auxiliary representations,an intra-modality consistency loss stabilizes feature learning,a global classification loss enhances identity discrimination.Additionally,a cross-modal optimized k-reciprocal re-ranking algorithm to mitigate modality bias in neighborhood similarity assessment is developed.Extensive experiments on three datasets(SYSU-MM01,RegDB,and LLCM)demonstrate state-of-the-art performance:achieving Rank-1 accuracies of 76.35％(SYSU-MM01,All-search)and 92.83％(RegDB,VIS to IR)with mAP scores of 79.56％and 88.68％,respectively.In complex low-light scenarios of LLCM,Rank-1 reaches 68.2％(VIS to IR)and 56.5％(IR to VIS)with mAP of 66.3％and 63.4％.These results validate the synergistic effectiveness of our dynamic fusion strategy,multi-loss design,and enhanced re-ranking approach.