关键词: 跨模态, 行人重识别, 元学习, 特征增强, 全监督学习

Abstract: To address the difficulty of learning discriminative features caused by the substantial modality discrepancy between visible and infrared images, a visible-infrared person re-identification method based on modality alignment and feature enhancement was proposed. First, a Meta-learning-guided Modality-adaptive Alignment module (M²A) was developed to capture and quantify batch-wise distribution shifts. Grayscale weights and spectral perturbation parameters were dynamically generated to guide visible images toward infrared images in the color, spatial, and spectral domains, thereby producing modality-consistent augmented images, enriching training samples, and alleviating the cross-modality discrepancy. Second, a Dynamic Feature Enhancement Module (DFEM) was constructed. Guided by modality labels, independent channel-mapping spaces were established by parallel Multi-Layer Perceptron (MLP) branches, and globally shared spatial attention was incorporated to extract modality-consistent pedestrian features, achieving dynamic enhancement in both channel and spatial dimensions. Third, a discriminative feature mining module was designed by introducing a discriminative contrastive loss into the intermediate layers of the backbone network and jointly optimizing it with classification and metric learning losses, thereby strengthening the extraction of fine-grained invariant representations. Finally, a multi-task dynamic optimization strategy was developed by combining adaptive weighting of multiple losses with a stage-wise training scheme, so that the model was progressively guided from coarse-grained global alignment to fine-grained identity mining, improving its stability and generalization capability. Under the all-search single-shot setting on the SYSU-MM01 dataset, the proposed method achieves 85.20% Rank-1 accuracy and 82.92% mean Average Precision (mAP), outperforming Multi-level Cross-Modality Joint Alignment (MCJA) by 10.72 and 11.58 percentage points, respectively. Under the visible-to-infrared setting on the RegDB dataset, it achieves 94.85% Rank-1 accuracy and 95.14% mAP, with the latter improving by 5.19 percentage points over Low-light Invariant Representation Learning (IRL). Experimental results demonstrate the effectiveness of the proposed method.

Key words: cross-modality, person re-identification, meta-learning, feature enhancement, supervised learning