文献类型：期刊文献

英文题名：KANFuse: Enhancing infrared and visible image fusion through nonlinear representation modeling

作者：Zhang, Yongzi[1];Li, Shengshi[2];Fang, Aolin[1];He, Xinglong[1];Zhu, Daoheng[2];Wu, Keer[2];Xiao, Xiuchun[2]

机构：[1]Guangdong Ocean Univ, Coll Math & Comp Sci, Zhanjiang 524088, Peoples R China;[2]Guangdong Ocean Univ, Coll Elect & Informat Engn, Zhanjiang 524088, Peoples R China

年份：2026

卷号：153

外文期刊名：INFRARED PHYSICS & TECHNOLOGY

收录：SCI-EXPANDED(收录号：WOS:001648897000001)、、EI(收录号：20255219772324)、Scopus(收录号：2-s2.0-105025349974)、WOS

基金：This work was supported by the program for Scientific Research Start-up Funds of Guangdong Ocean University (060302112503, 060302112317, 060302112405), the National Natural Science Foundation of China (62472107), the Natural Science Foundation of Guangdong Province, China (2023A1515011477), the Demonstration Bases for Joint Training of Postgraduates of the Department of Education of Guangdong Province (202205), the Innovation Team Project of General University in Guangdong Province of China (2024KCXTD042), the Science and Technology Plan Project of Zhanjiang City (2022A01063, 2025B01103), the Postgraduate Education Innovation Plan Project of Guangdong Ocean University (202440, 202520, 202544), the Undergraduate Innovation Team Project of Guangdong Ocean University (CXTD2021019), the Guangdong University Student Science and Technology Innovation Cultivation Special Fund Support Project (pdjh2023 a0243), the Innovation and Entrepreneurship Training Program for College Students of Guangdong Ocean University (S202410566052), and the Guangdong Ocean University Postgraduate Practical Innovation Ability Improvement Plan Cultivation Project (202544) .

语种：英文

外文关键词：KAN; Deep learning; Image fusion; Infrared image; Visible image

外文摘要：Infrared and Visible Image Fusion (IVIF) is a technique used to integrate thermal information from infrared images with the fine details and textures of visible images to achieve comprehensive scene perception. It has broad applications in night vision, surveillance, and autonomous driving, where highlighting thermal targets while retaining scene details is essential. However, existing methods often struggle to effectively preserve modality-specific features and suffer from limited nonlinear modeling capacity, which hinders their ability to fully exploit the complementary information across modalities. In this research, we proposed KANFuse, a novel fusion network in which Kolmogorov-Arnold Networks (KAN) are incorporated to model complex nonlinear cross-modal interactions. To further enhance representation, Wavelet Convolution Blocks (WCBs) are employed for edge-aware and noise-suppressing feature extraction, while Dynamic Fusion Modules (DFMs) are integrated into skip connections to balance multi-source contributions. Additionally, a spectral guided fidelity loss (SFL) is designed for second-phase training to better retain realistic visual information. Extensive evaluations on TNO, M3FD, and LLVIP demonstrate that KANFuse consistently outperforms fourteen state-of-the-art methods in both qualitative and quantitative evaluations.