关系建模和频谱变换的跨模态行人重识别方法OA

Existing cross-modal person re-identification methods struggle to effectively address the signifi⁃cant modality gap between infrared and visible light images,often facing challenges in feature alignment and insufficient discriminative power,which severely limits recognition performance.To address this,this paper proposed a cross-modality person re-identification method based on relation modeling and spectrum transformation,starting from the perspectives of enhancing intrinsic feature correlations and mining com⁃mon information in the frequency domain.First,to address the difficulty in aligning local feature seman⁃tics,a segment relation modeling framework was introduced with a self-attention mechanism,strengthen⁃ing intra-modal local feature associations and establishing tighter contextual information.Second,to over⁃come the limitation of single-scale feature information,a multi-scale feature enhancement module was de⁃signed to improve the network's ability to capture subtle differences in people through multi-granularity per⁃ception.Finally,a channel spectral transformation process was designed to mine potential common spec⁃tral information in the frequency domain during feature extraction,further narrowing the modality gap.Ex⁃perimental results show that the proposed method achieves Rank-1 and mAP scores of 84.8％and 81.5％,respectively,in the all-search mode of the SYSU-MM01 dataset;92.6％and 87.1％on the RegDB dataset;and 58.0％and 64.5％on the LLCM dataset.These results demonstrate significant ad⁃vantages across multiple metrics,fully validating the effectiveness of the proposed method.