Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HSR-KAN: Efficient Hyperspectral Image Super-Resolution via Kolmogorov-Arnold Networks

Baisong Li, Xingwang Wang, Haixiao Xu

TL;DR

This work tackles hyperspectral image super-resolution by fusing low-resolution hyperspectral data with high-resolution multispectral imagery. It introduces HSR-KAN, a hybrid network that integrates Kolmogorov-Arnold Networks (KANs) with CNNs/MLPs through a KAN-Fusion module and a spectral channel attention block (KAN-CAB), complemented by a Restructure module and sparsity-based regularization to mitigate the Curse of Dimensionality. The method achieves state-of-the-art results across CAVE, Chikusei, and Harvard datasets, with strong generalization and improved efficiency due to spline-based activations and lightweight design. The approach provides a scalable, accurate, and efficient solution for high-quality HR-HSI reconstruction, enabling better spectral-spatial analysis in practical imaging tasks.

Abstract

Hyperspectral images (HSIs) have great potential in various visual tasks due to their rich spectral information. However, obtaining high-resolution hyperspectral images remains challenging due to limitations of physical imaging. Inspired by Kolmogorov-Arnold Networks (KANs), we propose an efficient HSI super-resolution (HSI-SR) model to fuse a low-resolution HSI (LR-HSI) and a high-resolution multispectral image (HR-MSI), yielding a high-resolution HSI (HR-HSI). To achieve the effective integration of spatial information from HR-MSI, we design a fusion module based on KANs, called KAN-Fusion. Further inspired by the channel attention mechanism, we design a spectral channel attention module called KAN Channel Attention Block (KAN-CAB) for post-fusion feature extraction. As a channel attention module integrated with KANs, KAN-CAB not only enhances the fine-grained adjustment ability of deep networks, enabling networks to accurately simulate details of spectral sequences and spatial textures, but also effectively avoid Curse of Dimensionality. Extensive experiments show that, compared to current state-of-the-art HSI-SR methods, proposed HSR-KAN achieves the best performance in terms of both qualitative and quantitative assessments. Our code is available at: https://github.com/Baisonm-Li/HSR-KAN.

HSR-KAN: Efficient Hyperspectral Image Super-Resolution via Kolmogorov-Arnold Networks

TL;DR

This work tackles hyperspectral image super-resolution by fusing low-resolution hyperspectral data with high-resolution multispectral imagery. It introduces HSR-KAN, a hybrid network that integrates Kolmogorov-Arnold Networks (KANs) with CNNs/MLPs through a KAN-Fusion module and a spectral channel attention block (KAN-CAB), complemented by a Restructure module and sparsity-based regularization to mitigate the Curse of Dimensionality. The method achieves state-of-the-art results across CAVE, Chikusei, and Harvard datasets, with strong generalization and improved efficiency due to spline-based activations and lightweight design. The approach provides a scalable, accurate, and efficient solution for high-quality HR-HSI reconstruction, enabling better spectral-spatial analysis in practical imaging tasks.

Abstract

Hyperspectral images (HSIs) have great potential in various visual tasks due to their rich spectral information. However, obtaining high-resolution hyperspectral images remains challenging due to limitations of physical imaging. Inspired by Kolmogorov-Arnold Networks (KANs), we propose an efficient HSI super-resolution (HSI-SR) model to fuse a low-resolution HSI (LR-HSI) and a high-resolution multispectral image (HR-MSI), yielding a high-resolution HSI (HR-HSI). To achieve the effective integration of spatial information from HR-MSI, we design a fusion module based on KANs, called KAN-Fusion. Further inspired by the channel attention mechanism, we design a spectral channel attention module called KAN Channel Attention Block (KAN-CAB) for post-fusion feature extraction. As a channel attention module integrated with KANs, KAN-CAB not only enhances the fine-grained adjustment ability of deep networks, enabling networks to accurately simulate details of spectral sequences and spatial textures, but also effectively avoid Curse of Dimensionality. Extensive experiments show that, compared to current state-of-the-art HSI-SR methods, proposed HSR-KAN achieves the best performance in terms of both qualitative and quantitative assessments. Our code is available at: https://github.com/Baisonm-Li/HSR-KAN.
Paper Structure (28 sections, 12 equations, 5 figures, 5 tables, 1 algorithm)

This paper contains 28 sections, 12 equations, 5 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Deep Learning-Based HSI-SR
  4. Kolmogorov-Arnold Networks (KANs)
  5. Proposed HSR-KAN Network
  6. Problem Formulation
  7. Architecture Overview
  8. KAN Layer
  9. KAN-Fusion
  10. KAN Channel Attention Block (KAN-CAB)
  11. Restructure Module and Loss Function
  12. Experiment
  13. Datasets and Benchmark
  14. Implementation Details
  15. Comparison with State of Arts Methods
  16. ...and 13 more sections

Figures (5)

  • Figure 1: The structure diagram of HSR-KAN. "Conv" denotes a convolutional layer with a 3$\times$3 kernel, "ReLU" denotes the ReLU activation function, "GAP" denotes the Global Average Pooling, "Spline" denotes the B-spline function, "SiLU" denotes SiLU activation function.
  • Figure 2: Visual quality comparison on the CAVE for $\times4$ SR, where the first row shows pseudo-color (R-20, G-30, B-2) images and second row shows corresponding heatmaps (mean squared error). (a) FUSE, (b) MHF-Net, (c) HSRnet, (d) Fusformer, (e) DCTransformer, (f) HSR-Diff, (g) HSR-KAN, (h) GT.
  • Figure 3: Visual quality comparison on Chikusei for $\times4$ SR, where first row shows pseudo-color (R-64, G-58, B-16) images and second row shows corresponding heatmaps. (a) FUSE, (b) MHF-Net, (c) HSRnet, (d) Fusformer, (e) DCTransformer, (f) HSR-Diff, (g) HSR-KAN, (h) GT.
  • Figure 4: Visual quality comparison on Havard for $\times4$ SR, where first row shows the pseudo-color (R-29, G-22, B-31) images and second row shows corresponding heatmaps. (a) FUSE, (b) MHF-Net, (c) HSRnet, (d) Fusformer, (e) DCTransformer, (f) HSR-Diff, (g) HSR-KAN, (h) GT.
  • Figure 5: The full-size HSR-KAN model (Baseline), is compared with the two ablation models 'Without Sparse Loss' and 'Without CAB' for their training performance on the CAVE dataset for $\times4$ SR. The left graph plots the variation in PSNR values, while the right graph shows the changes in loss values.