Self-supervised Segmentation Feature Alignment for Infrared and Visible Image Fusion

Existing deep learning-based methods for infrared and visible image fusion typically operate independently of other high-level vision tasks, overlooking the potential benefits these tasks could offer. For instance, semantic features from image segmentation could enrich the fusion results by providing detailed target information. However, segmentation focuses on target-level semantic feature information (e.g., object categories), while fusion focuses more on pixel-level detail feature information (e.g., local textures), creating a feature representation gap. To address this challenge, we propose a self-supervised segmentation feature alignment fusion network (SegFANet), which aligns target-level semantic features from segmentation tasks with pixel-level fusion features through self-supervised learning, thereby bridging the feature gap between the two tasks and improving the quality of image fusion. Extensive experiments on the WHU and Potsdam datasets show our method's effectiveness, outperforming the state-of-the-art methods.