KCUNET: Multi-Focus Image Fusion via the Parallel Integration of KAN and Convolutional Layers

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Multi-focus image fusion (MFIF) integrates images from different focal planes to create a fully focused image.
• The defocus spread effect (DSE) causes blurred boundaries in fused images, degrading visual quality.
• Existing methods struggle with effectively reducing DSE and preserving fine details.

Purpose of the Study:

• To propose a novel deep learning model, KCUNet, for enhanced multi-focus image fusion.
• To address the challenge of defocus spread effect (DSE) in fused images.
• To improve the quality and edge preservation of fused images.

Main Methods:

• Developed KCUNet, a U-Net architecture integrating Kolmogorov-Arnold networks with parallel convolutional layers.
• Maintained spatial dimensions and increased channel depth for multi-level feature extraction.
• Incorporated a content-guided attention mechanism for edge information processing.
• Utilized a hybrid loss function evaluating edge alignment, mask prediction, and image quality.

Main Results:

• KCUNet demonstrated superior performance in reducing DSE and preserving edge details.
• Qualitative and quantitative evaluations showed significant improvements over 15 state-of-the-art methods.
• The model effectively maintained high-resolution details and captured multi-level features.

Conclusions:

• KCUNet offers a robust solution for multi-focus image fusion, significantly mitigating DSE.
• The proposed architecture and hybrid loss function contribute to enhanced image clarity and detail preservation.
• KCUNet represents a significant advancement in the field of image fusion technology.