Wavelet-Based Frequency Feature Learning for Visible-Infrared Person Re-Identification

Visible-infrared person re-identification (VI-ReID) is a challenging person retrieval task under two different modality cameras, due to the large gap between visible and infrared images and intra-class variations within the same individuals. Existing works mainly focus on extracting discriminative spatial features to achieve matching purposes, while overlooking the difference between the low-frequency and high-frequency components in different modalities. To solve this problem, we propose a Wavelet-based Frequency Feature Learning network called WFLNet, which tackles the VI-ReID task from a frequency perspective by filtering out the irrelevant modality-specific components and leveraging the frequency-domain cues to learn more comprehensive and discriminative features. Firstly, we introduce a Processed Low-Frequency Mask (PLFM) to filter out the irrelevant low-frequency components, guiding the model to focus more on modality-invariant features and suppress the negative impact of modality differences caused by the brightness and color information in VIS and thermal radiation information in IR. Secondly, a Multi-scale Wavelet Query Attention module (MWQA) is designed to extract the modality-invariant feature by integrating low- and high-frequency features in cross-scales to enhance global structure and detail information. Finally, the dual frequency loss is proposed to reduce the modality gap by pulling one modality feature closer to another modality center. Extensive experiments on the SYSU-MM01, RegDB, and LLCM datasets demonstrate that the proposed WFLNet outperforms most existing works.