Scholarlink

<p>Infrared-visible person re-identification aims to map heterogeneous features into a unified space for person cross-modal retrieval. Existing methods overlook multi-scale frequency information and inherent spectral differences, leading to distortion of signal feature decomposition and degradation of essential discrimination ability. To address these issues, we proposes WaveRID, a framework leveraging wavelet representation for cross-modal matching. First, to decompose features into frequency subbands, the Wave-Adaptive Encoder (WAE) uses discrete wavelet transform. This module employs specialized perception units to process multi-scale information. Second, the Frequency Enhance Block (FEB) models global spectral characteristics using Fourier transform to complement wavelet limitations in capturing global information. Third, to construct robust feature spaces, Enhanced Similarity Distribution Clustering (ESDC) optimizes both cross-modal alignment and intra-modal discrimination. Experiments on RegDB and SYSU-MM01 datasets demonstrate state-of-the-art performance, validating the effectiveness of multi-scale wavelet representation in cross-modal recognition tasks.</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

WaveRID: wavelet representation for infrared-visible person re-identification

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Junyang Lai,
  • Xuefeng Tao,
  • Jun Kong

摘要

Infrared-visible person re-identification aims to map heterogeneous features into a unified space for person cross-modal retrieval. Existing methods overlook multi-scale frequency information and inherent spectral differences, leading to distortion of signal feature decomposition and degradation of essential discrimination ability. To address these issues, we proposes WaveRID, a framework leveraging wavelet representation for cross-modal matching. First, to decompose features into frequency subbands, the Wave-Adaptive Encoder (WAE) uses discrete wavelet transform. This module employs specialized perception units to process multi-scale information. Second, the Frequency Enhance Block (FEB) models global spectral characteristics using Fourier transform to complement wavelet limitations in capturing global information. Third, to construct robust feature spaces, Enhanced Similarity Distribution Clustering (ESDC) optimizes both cross-modal alignment and intra-modal discrimination. Experiments on RegDB and SYSU-MM01 datasets demonstrate state-of-the-art performance, validating the effectiveness of multi-scale wavelet representation in cross-modal recognition tasks.