CAIF: Cross-Attention Framework in Unaligned Infrared and Visible Image Fusion

Infrared and visible light image fusion has gained significant attention due to its advantages in challenging environments. While visible light cameras capture rich texture details under normal conditions, they perform poorly in situations like nighttime or smoke. Infrared sensors, however, can effectively highlight key targets even in extreme conditions by detecting thermal radiation. Most existing fusion methods assume perfectly aligned source images, overlooking geometric distortions and pixel shifts caused by varying camera parameters. To address this, we propose a unified framework for infrared and visible light image registration and fusion. This framework effectively handles misaligned images and introduces semantic constraints to enhance the performance in advanced vision tasks. Additionally, a cross-attention-based feature fusion module is integrated into the network. Experimental results show that our method not only improves fusion quality under misalignment but also enhances the performance of subsequent vision tasks.