FUsionKAN: a Multi-focus Image Fusion Diffusion Model with KAN

Depth of Field (DoF) limitations in optical imaging often result in images with blurred regions, posing challenges for applications such as microscopy and remote sensing. This issue is addressed by multi-focus image fusion (MFIF) by merging images from different focal planes into a single, fully focused image. To enhance MFIF performance, we propose FUsionKAN, a diffusion model framework that incorporates Tok-KAN into the noise predictor, strengthening the model’s ability to perform nonlinear modeling of image features and significantly improving its performance. Additionally, an improved skip connection technique in U-Net is introduced, enabling better utilization of high-frequency information to preserve more edge details in the fused images. These innovations enhance the model’s ability to capture complex relationships and improve image generation quality. Comprehensive experiments on benchmark datasets demonstrate the effectiveness of FUsionKAN in producing high-quality fused images, supported by detailed ablation studies.