Combining information augmentation aggregation and dual-granularity feature fusion for visible-infrared person re-identification

Visible-Infrared Person Re-Identification (VI-ReID), which aims to retrieve pedestrian images captured by visible and infrared cameras, presents a significant challenge in intelligent surveillance systems. VI-ReID should not only tackle the modality discrepancies between visible and infrared images, but also address intra-modality discrepancies caused by factors such as image occlusion, lighting changes, and background complexity. In this paper, we propose a novel VI-ReID network (IADGN) that effectively combines an Information Augmentation Aggregation Module (IAAM) with a Dual-Granularity Feature Module (DGFM) to balance the processing of cross-modality and intra-modality discrepancies. First, during the data processing stage, a random grayscale strategy is employed for both visible and infrared images to effectively minimize the interference of color information. Second, to address common intra-modality discrepancies such as occlusion, lighting changes, and background complexity, we design an Information Augmentation Aggregation Module (IAAM) based on a self-attention mechanism. This module accurately focuses on pedestrian features and aggregates them with the original features in a channel-adaptive manner, effectively ignoring the cluttered background and enhancing feature discriminability. Additionally, we propose a Dual-Granularity Feature Module (DGFM) that integrates global and local features, overcoming the limitations of single-granularity feature learning, and significantly enhancing the network's recognition accuracy. Finally, we propose the Center Distribution Consistency Loss function (CDCL), which reduces modality discrepancies and enhances the modality consistency of feature representation by aligning the inter-class distributions of visible and infrared images. Extensive experimental results on three publicly available datasets-SYSU-MM01, RegDB, and LLCM-demonstrate the effectiveness and superiority of the proposed method.