Wavelet-based multi-level information compensation learning for visible-infrared person re-identification

Abstract

The main challenge in cross-modal person re-identification (VI-ReID) is extracting discriminative features from different modalities. Most existing methods focus on minimizing modal differences but overlook the shallow modality-invariant information lost as network depth increases. To address this, we propose the Wavelet-based Multi-level Information Compensation (WMIC) learning method. At multiple network stages, we design an Information Compensation Block (ICB) that applies wavelet decomposition to deep features, producing four wavelet subbands to preserve modality-invariant details and enlarge the receptive field. These subbands are used to compute an attention matrix with shallow features, which is then applied to enhance shallow features' local information. Additionally, we represent each person image with two sets of embeddings by introducing a Wavelet Enhancement Block (WEB) to generate an additional embedding. Finally, we use a dual-branch center-guided loss to make the two embeddings complementary, thereby reducing the disparity between infrared and visible images. Extensive experiments on the SYSU-MM01, RegDB, and LLCM datasets demonstrate that WMIC outperforms existing methods.