Visible-infrared person re-identification (VI-ReID) is an urgent issue for realizing smart security. The task is designed to perform mutual retrieval of the same pedestrians in two modalities, facing the challenges caused by the huge intra-modal and inter-modal discrepancies. Existing methods often reduce modality gaps by introducing auxiliary modalities, but simple strategies like channel random selection or grayscale transformation yield limited accuracy gains. These approaches directly incorporate auxiliary modalities into training, leading to model oscillations. Therefore, the paper introduces a novel framework named Hybrid Fine-tuning Heterogeneous Fusion Network ( \(HF^{2}\text {-}Net\) ) for VI-ReID. \(HF^{2}\text {-}Net\) divides the dual-stream parameter-sharing model into four modules, including Random Hybrid Correction (RHC), Heterogeneous Fusion (HF), Fine-tuning Hybrid (FH) and United Learning (UL) to realize heterogeneous feature learning to homogeneous feature embedding. To enhance modality alignment, this paper designs Correction & Integration mechanisms (CIM A and B), and embeds them into different modules. CIM A and B are data-level and feature-level stochastic augmentation strategies, respectively. They perform feature fine-tuning and correction of the input information, further reducing the difficulties caused by modal discrepancies in the cross-modal task. Additionally, considering the favorable performance of global-local feature learning in ReID, the paper proposes a body fusion strategy for shared space alignment of global-local features. Combined with a multi-granularity supervision mechanism, this strategy improves intra-class and inter-class feature consistency. Experimental results on two public datasets demonstrate that the proposed method is effective in constructing pedestrian representations, thus improving the accuracy of VI-ReID. The codes are available at https://github.com/sherrysherry-song/hf2c.
