D-SELACFP: Visible-Infrared Person Re-Identification via Local Attention and Cross-Modal Feature Perception Enhancement

Visible-Infrared Person Re-Identification (VI-ReID) faces significant challenges such as difficult feature matching and large pose variations, leading to low cross-modality image matching accuracy. Existing approaches typically extract features directly from raw images and embed features from different modalities into a common space to learn shared representations. However, they often neglect the interference of noise and fail to fully utilize identity-discriminative information within modality-specific features, resulting in weak cross-modality invariant feature extraction and interference from irrelevant noise during feature matching. To address the issues of modal noise interference and feature matching difficulty in VI-ReID, this paper proposes a dual-stream neural network framework designed to reduce modal noise and enhance perceptual features. The framework incorporates a Local Attention Module (LAM) and an Inter-modal Feature Perception Enhancement Module (I-MFPE). These modules work synergistically to mine salient features both within and across modalities, while effectively leveraging the identity-discriminative information inherent in modality-specific features. Specifically, the LAM focuses on discriminative part-level features and suppresses background noise interference, thereby effectively extracting both modality-shared and modality-specific features while reducing the impact of irrelevant information. The I-MFPE module enhances the ability to extract shareable features from heterogeneous images by optimizing fine-grained feature representations in both channel and spatial dimensions, simultaneously mitigating the influence of modal differences on matching. The proposed method effectively alleviates noise introduced by factors such as viewpoint variations and background clutter, enhancing the discriminative power of cross-modality pedestrian features and providing a more robust feature representation for VI-ReID. Experimental results demonstrate the significant advantages of the proposed method in suppressing noise interference and improving feature matching performance.