Zero-Shot Hyperspectral Image Denoising Using Self-Completion With 3D Random Patterned Masks

Abstract

Hyperspectral images (HSIs) have higher spectral resolution than RGB images and are used in various tasks. However, HSIs are prone to degradation due to noise generated during imaging, making it difficult to obtain non-degraded images. Additionally, supervised learning, which relies on pairs of degraded and non-degraded images, is often challenging to apply to HSI restoration because of the high cost of imaging and the need to prepare large amounts of data. To overcome these limitations, recent advances in self-supervised learning have led to the development of learning-based image restoration methods that do not require non-degraded images. However, these methods have limitations, including low accuracy and the need to estimate the noise distribution. In this paper, we propose a zero-shot HSI deep denoising method based on self-supervised image restoration. The proposed method achieves zero-shot recovery by repeatedly predicting blind-spots in 3D blocks during the learning process. Notably, our method does not require training or clean images, nor does it rely on noise distribution information. Numerical experiments and ablation studies confirmed that the restoration accuracy of the proposed method is comparable to or better than that of conventional zero-shot methods.

Keywords

• Hyperspectral image
• denoising
• zero-shot learning
• self-supervised learning