Abstract
To enhance scene perception and comprehension, infrared and visible image fusion (IVIF) integrates complementary data from two modalities. However, many existing methods fail to explicitly separate modality-specific and modality-shared features, which compromises fusion quality. To surmount this constraint, we introduce a novel hierarchical dual-branch fusion (HDF-Net) network. The network decomposes the source images into low-frequency components, which capture shared structural information, and high-frequency components, which preserve modality-specific details. Remarkably, we propose a pin-wheel-convolutional transformer (PCT) module that integrates local convolutional processing with directional attention to improve low-frequency feature extraction, thereby enabling more robust global-local context modeling. We subsequently introduce a hierarchical feature refinement (HFR) block that adaptively integrates multiscale features using kernel-based attention and dilated convolutions, further improving fusion accuracy. Extensive experiments on four public IVIF datasets (MSRS, TNO, RoadScene, and M3FD) demonstrate the high competitiveness of HDF-Net against 12 state-of-the-art methods. On the RoadScene dataset, HDF-Net achieves top performance across six key metrics-EN, SD, AG, SF, SCD, and SSIM-surpassing the second-best method by 0.67%, 1.85%, 17.67%, 5.26%, 3.33%, and 1.01%, respectively. These findings verify the generalization and efficacy of HDF-Net in practical IVIF scenarios.