Adaptive Regularized Low-Rank Tensor Decomposition for Hyperspectral Image Denoising and Destriping

TL;DR

This paper introduces an adaptive hyperLaplacian regularized low-rank tensor decomposition method for hyperspectral image denoising and destriping, effectively modeling stripe noise and preserving spatial-spectral structures.

Contribution

It proposes a novel tensor decomposition approach combined with adaptive hyper-Laplacian regularization to better handle stripe noise and spectral details in hyperspectral images.

Findings

Outperforms existing methods in denoising quality

Effectively separates stripe noise from hyperspectral images

Reduces spatial information loss and over-smoothing

Abstract

Hyperspectral images (HSIs) are inevitably degraded by a mixture of various types of noise, such as Gaussian noise, impulse noise, stripe noise, and dead pixels, which greatly limits the subsequent applications. Although various denoising methods have already been developed, accurately recovering the spatial-spectral structure of HSIs remains a challenging problem to be addressed. Furthermore, serious stripe noise, which is common in real HSIs, is still not fully separated by the previous models. In this paper, we propose an adaptive hyperLaplacian regularized low-rank tensor decomposition (LRTDAHL) method for HSI denoising and destriping. On the one hand, the stripe noise is separately modeled by the tensor decomposition, which can effectively encode the spatial-spectral correlation of the stripe noise. On the other hand, adaptive hyper-Laplacian spatial-spectral regularization is introduced to represent the distribution structure of different HSI gradient data by adaptively estimating the optimal hyper-Laplacian parameter, which can reduce the spatial information loss and over-smoothing caused by the previous total variation regularization. The proposed model is solved using the alternating direction method of multipliers (ADMM) algorithm. Extensive simulation and real-data experiments all demonstrate the effectiveness and superiority of the proposed method.