PWTSI-Net: a progressive wavelet-refined and transformer-enhanced semantic interaction framework for infrared and visible image fusion

Infrared (IR) and visible (VIS) image fusion focuses on merging complementary thermal and structural information into a single informative fused image. IR images effectively highlight salient thermal targets, whereas VIS images preserve abundant texture and structural information. However, existing fusion methods often suffer from insufficient frequency representation, loss of fine details, and limited semantic understanding. To address these challenges, this paper proposes PWTSI-Net, a novel framework for IR and VIS image fusion. The framework integrates frequency-aware feature enhancement and semantic-guided contextual refinement within a unified encoder–decoder architecture. Initially, a coarse fused representation is generated from the source modalities and progressively refined using the proposed adaptive wavelet attention refinement block (AWARB). The AWARB consists of an adaptive wavelet enhancement block (AWEB) for multi-frequency feature enhancement and an attention-guided feature refinement block (AGFRB) for adaptive feature recalibration. Furthermore, a context aware transformer fusion module (CATFM) is introduced to enhance semantic representation by incorporating segmentation-guided semantic priors, modelling long-range feature dependencies through transformer-based attention, and adaptively regulating feature interactions via gating mechanisms. Experimental results demonstrate that PWTSI-Net effectively preserves thermal targets, structural details, and texture information while producing visually consistent and information-rich fused images with improved contrast and semantic consistency.