Scholarlink

<p>Intelligent surveillance systems are increasingly vital for improving public safety and operational efficiency. The field of computer vision offers a robust platform for constructing 24/7 intelligent surveillance systems. A noteworthy technological aspect within this domain is person re-identification, a capability that facilitates intelligent surveillance with minimal human intervention. Visible-infrared person re-identification (VI-ReID) remains a highly challenging task due to substantial modality discrepancies and spatial misalignment between heterogeneous image pairs. This paper presents an end-to-end framework that integrates the Enhanced Holistic Scale Network (E-HSNet) with a novel Cross-Sensory Integration (CSI) module to address global identity learning and fine-grained cross-modal alignment jointly. E-HSNet enhances multiscale feature learning via a Holistic-Scale residual architecture, while the CSI module facilitates patch-level similarity modeling through attention-guided soft warping and person-mask-based background suppression. To ensure robust feature discrimination and alignment, the model is optimized using a composite loss that includes identity classification loss, identity consistency loss, and dense triplet loss. Extensive evaluations are conducted on the SYSU-MM01 and RegDB datasets. On SYSU-MM01 (all-search, single-shot), the proposed method achieves 65.1% Rank-1 accuracy and 64.3% mAP, significantly outperforming the strongest baseline DML (Rank-1: 58.4%, mAP: 56.1%). On the RegDB dataset, the model attains 82.1% Rank-1 and 74.7% mAP (infrared-to-visible), surpassing the leading method NFS (Rank-1: 80.54%, mAP: 72.10%). These results highlight the effectiveness of multi-scale modeling and cross-modal interaction in mitigating modality gaps, thereby establishing a new state-of-the-art for VI-ReID benchmarks.</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

Enhanced holistic scale feature learning for VI-ReID for intelligent surveillance systems: A lightweight CNN approach

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Deepthi Shree A.M.,
  • Brindha M

摘要

Intelligent surveillance systems are increasingly vital for improving public safety and operational efficiency. The field of computer vision offers a robust platform for constructing 24/7 intelligent surveillance systems. A noteworthy technological aspect within this domain is person re-identification, a capability that facilitates intelligent surveillance with minimal human intervention. Visible-infrared person re-identification (VI-ReID) remains a highly challenging task due to substantial modality discrepancies and spatial misalignment between heterogeneous image pairs. This paper presents an end-to-end framework that integrates the Enhanced Holistic Scale Network (E-HSNet) with a novel Cross-Sensory Integration (CSI) module to address global identity learning and fine-grained cross-modal alignment jointly. E-HSNet enhances multiscale feature learning via a Holistic-Scale residual architecture, while the CSI module facilitates patch-level similarity modeling through attention-guided soft warping and person-mask-based background suppression. To ensure robust feature discrimination and alignment, the model is optimized using a composite loss that includes identity classification loss, identity consistency loss, and dense triplet loss. Extensive evaluations are conducted on the SYSU-MM01 and RegDB datasets. On SYSU-MM01 (all-search, single-shot), the proposed method achieves 65.1% Rank-1 accuracy and 64.3% mAP, significantly outperforming the strongest baseline DML (Rank-1: 58.4%, mAP: 56.1%). On the RegDB dataset, the model attains 82.1% Rank-1 and 74.7% mAP (infrared-to-visible), surpassing the leading method NFS (Rank-1: 80.54%, mAP: 72.10%). These results highlight the effectiveness of multi-scale modeling and cross-modal interaction in mitigating modality gaps, thereby establishing a new state-of-the-art for VI-ReID benchmarks.