基于斑块解耦-自协同对齐的可见光-红外行人重识别OA

Modality entanglement introduced by spectral differences and heterogeneity interference reduces the accuracy of cross-modality person re-identification.To suppress the degradation of discrimination characteristics and heterogeneity interference caused by modality entanglement,a Plaque-Decoupling Self-Collaborative Alignment model is proposed,which consists of a plaque-decoupling and self-collaborative alignment modules.To mine the intra-class feature discrimination information,the plaque-decoupling module decomposes the cross-modality features in the channel dimension and extracts patch features,which can suppress the imaging modality entanglement of the visible and infrared images.The self-collaborative alignment module includes diversity decoupling and alignment constraints that can filter abnormal features caused by lighting factors and eliminate inter-class differences.The diversity-decoupling constraints learn the diversity-decoupling features,and the decoupled information is used by the alignment constraints to match the features.Experimental results on the SYSU-MM01 dataset demonstrate the effectiveness of the method.