MAHF —Net: Multi‐Scale Alignment and High‐Order Fusion Network for Unregistered Infrared‐Visible Image Fusion

ABSTRACT Infrared‐visible image fusion enhances perceptual quality by integrating the detailed textures of visible images with the thermal target saliency of infrared images, making it highly valuable for applications such as military reconnaissance and autonomous driving. However, existing methods encounter three major limitations: performance degradation in unregistered scenarios due to non‐rigid deformations, restricted cross‐modal feature interaction caused by low‐order operations, and fusion distortions resulting from variations in object scale. To overcome these challenges, this study introduces a multi‐scale alignment and high‐order fusion network ( MAHF‐Net ). Under a shared cross‐modal multi‐scale encoding‐decoding framework, three novel components are introduced: an Infrared‐Guided Spatial Deformable Alignment (ISDA) module, a High‐Order Spatial‐Channel Interaction Block (HOSCIB), and a Cross‐Scale Dynamic Aggregation (CSDA) module. The ISDA module exploits infrared salient regions to generate spatial attention, which regulates offset fields that guide deformable convolution for adaptive non‐rigid deformation correction. The HOSCIB applies a high‐order spatial–channel joint interaction mechanism, enabling iterative feature refinement through dynamic spatial attention gating combined with cascaded channel attention. The CSDA module incorporates dynamically gated attention to balance semantic and detail contributions. Extensive experiments have been conducted on the RoadScene , LLVIP and AVIID datasets. The results demonstrate that the proposed MAHF‐Net consistently outperforms state‐of‐the‐art registered and unregistered fusion methods, achieving mutual information improvements of 24.6% and 4.35%, respectively.