Scholarlink

<p>Visible-Infrared Person Re-Identification (VI-ReID) is an indispensable technology for intelligent surveillance in low-light or nighttime scenarios, yet it is severely constrained by three core challenges: inherent cross-modality discrepancies between visible (VIS) and infrared (IR) images, fine-grained semantic misalignment caused by static region division in feature learning, and limited feature diversity leading to insufficient discriminative power. Existing feature-level methods rely on fixed human body segmentation, failing to adapt to dynamic body part variations and resulting in semantic misalignment. Image-level methods, while attempting modality conversion via Generative Adversarial Networks (GANs), inevitably introduce noise and artifacts that degrade feature quality. To address these critical limitations, this paper proposes a Multi-Scale Prototype-Driven Representation Learning Network (MPCR), with the core motivation of achieving precise cross-modality alignment, enhancing fine-grained feature fusion, and enriching feature diversity for robust VI-ReID. MPCR integrates four novel and complementary modules: First, the Wavelet-Enhanced Multi-Feature Generation Module (WMFGM), which leverages multi-scale dilated convolutions and wavelet transform to generate diverse, frequency-aware features, mitigating data scarcity and strengthening cross-modality shared feature capture; Second, the Coordinate Attention Fusion Module (CAF), which adopts position-aware attention to realize dynamic, fine-grained feature fusion, resolving semantic misalignment from static division; Third, the Statistically Normalized Attention Prototype Module (SNAP), which combines statistical normalization and learnable prototypes to mine semantically consistent local features, promoting instance-level cross-modality alignment; Finally, the Cosine Diversity Loss (CDL), which fuses KL divergence, contrastive loss, and diversity loss to enhance prototype distinctiveness and feature diversity. Extensive experiments on the SYSU-MM01 and LLCM datasets validate MPCR’s superiority. On SYSU-MM01, under the All-search mode, MPCR achieves a Rank-1 accuracy of 77.27% and a mAP of 74.48%, outperforming the state-of-the-art DEEN by 2.57% and 2.68%, respectively. Under the Indoor-search mode, it attains Rank-1 and mAP of 86.06% and 88.05%, surpassing DEEN by 5.76% and 4.75%. These results confirm that MPCR effectively addresses the core challenges of VI-ReID, providing a robust and advanced solution for cross-modality person matching.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale prototype contrast and feature fusion for visible-infrared person re-identification

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Qiangqiang Xie,
  • Xudong Shi,
  • Dapeng Li,
  • Haitao Zhao,
  • Guang Han

摘要

Visible-Infrared Person Re-Identification (VI-ReID) is an indispensable technology for intelligent surveillance in low-light or nighttime scenarios, yet it is severely constrained by three core challenges: inherent cross-modality discrepancies between visible (VIS) and infrared (IR) images, fine-grained semantic misalignment caused by static region division in feature learning, and limited feature diversity leading to insufficient discriminative power. Existing feature-level methods rely on fixed human body segmentation, failing to adapt to dynamic body part variations and resulting in semantic misalignment. Image-level methods, while attempting modality conversion via Generative Adversarial Networks (GANs), inevitably introduce noise and artifacts that degrade feature quality. To address these critical limitations, this paper proposes a Multi-Scale Prototype-Driven Representation Learning Network (MPCR), with the core motivation of achieving precise cross-modality alignment, enhancing fine-grained feature fusion, and enriching feature diversity for robust VI-ReID. MPCR integrates four novel and complementary modules: First, the Wavelet-Enhanced Multi-Feature Generation Module (WMFGM), which leverages multi-scale dilated convolutions and wavelet transform to generate diverse, frequency-aware features, mitigating data scarcity and strengthening cross-modality shared feature capture; Second, the Coordinate Attention Fusion Module (CAF), which adopts position-aware attention to realize dynamic, fine-grained feature fusion, resolving semantic misalignment from static division; Third, the Statistically Normalized Attention Prototype Module (SNAP), which combines statistical normalization and learnable prototypes to mine semantically consistent local features, promoting instance-level cross-modality alignment; Finally, the Cosine Diversity Loss (CDL), which fuses KL divergence, contrastive loss, and diversity loss to enhance prototype distinctiveness and feature diversity. Extensive experiments on the SYSU-MM01 and LLCM datasets validate MPCR’s superiority. On SYSU-MM01, under the All-search mode, MPCR achieves a Rank-1 accuracy of 77.27% and a mAP of 74.48%, outperforming the state-of-the-art DEEN by 2.57% and 2.68%, respectively. Under the Indoor-search mode, it attains Rank-1 and mAP of 86.06% and 88.05%, surpassing DEEN by 5.76% and 4.75%. These results confirm that MPCR effectively addresses the core challenges of VI-ReID, providing a robust and advanced solution for cross-modality person matching.