MSFDN: multi-scale spatial-spectral-frequency joint denoising network for hyperspectral images

Hyperspectral images often suffer from various types of noise, such as Gaussian noise, impulse noise, stripe noise, and deadline, which are caused by weather conditions and sensor equipment. Hyperspectral denoising aims to remove noise and obtain clear images. Therefore, in order to more accurately...

Full description

Saved in:
Bibliographic Details
Main Authors: Kai Ren, Weiwei Sun, Xiangchao Meng, Gang Yang
Format: Article
Language:English
Published: Taylor & Francis Group 2025-03-01
Series:Geo-spatial Information Science
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/10095020.2025.2477552
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Hyperspectral images often suffer from various types of noise, such as Gaussian noise, impulse noise, stripe noise, and deadline, which are caused by weather conditions and sensor equipment. Hyperspectral denoising aims to remove noise and obtain clear images. Therefore, in order to more accurately describe and extract noise, and obtain better denoising results, we propose a novel approach called the multi-scale spatial-spectral-frequency joint denoising network for hyperspectral images (MSFDN). Considering the spatial-spectral distribution characteristics of noisy signals, MSFDN first converts the hyperspectral image into a low-dimensional subspace by constructing a subspace projection network. Then, the frequency distribution characteristics of noise signals in subspaces are described, and the noise signals are removed by designing a noise separation network. Finally, a subspace back projection network is constructed to reconstruct the initialized denoising image, and the conditional diffusion model is used to refine the initialized denoising results. It is worth noting that the above framework is embedded in the UNet network. Extensive experiments are conducted on simulated and real datasets. The experimental results demonstrate that MSFDN outperforms other state-of-the-art methods and exhibits robustness.
ISSN:1009-5020
1993-5153