HYPERSPECTRAL IMAGE DENOISING VIA SPECTRAL AND SPATIAL LOW-RANK
APPROXIMATION
Yi Chang, Luxin Yan, and Sheng Zhong
{yichang, luxinyan, zhongsheng}@hust.edu.cn
School of Automation, Huazhong University of Science and Technology, China
ABSTRACT
Hyperspectral images (HSI) unavoidably suffer from degra-
dations such as random noise, due to photon effects, calibra-
tion error, and so on. Most of existing HSI denoising methods
focus on utilizing the spectral correlation or the spatial non-
local self-similarity individually. In this paper, we propose an
unified low-rank recovery framework for HSI denoising, in
which taking both the underlying characteristics of high cor-
relation across spectra and non-local self-similarity over the
space cubic of HSI into consideration simultaneously. Our
work rely on a basic observation that both the multiple spec-
tral bands and similar spatial structures are lying on low-rank
subspaces and can facilitate to remove the noise jointly. Ex-
perimental results on both simulated and real HSI demon-
strate that the proposed method can significantly outperform
the state-of-the-art methods on several datasets in terms of
both visual and quantitative assessment.
Index Terms— Hyperspectral imaging, denoising, low-
rank.
1. INTRODUCTION
Most HSI classification/recognition algorithms assume that
the input HSI is of scene content that is clear and visible.
However, HSI often suffers from random noise in individual
bands, which badly limits the subsequent processing. There-
fore, it is natural for us to remove the noise as an important
preprocessing procedure.
As an classical yet hot research field, a variety of HSI
denoising methods have been proposed for the restoration of
HSIs. The HSI denoising methods have been classified into
different categories. In this work, we classify the HSI de-
noising methods into two categories according to the utiliza-
tion information: spectral methods [1, 2, 3, 4, 5, 6] and spa-
tial non-local self-similarity methods [7, 8, 9, 10]. The first
kind of methods mainly rely on the high spectral correlation
in HSI. In [1], by lexicographically ordering the 3D cube into
a 2D matrix, the authors proposed a low-rank matrix restora-
tion method for mixed noise removal in HSI. Further, Lu et
Thanks to National Natural Science Foundation of China under Grant
No. 61571207 for funding.
......
......
P
P
B
K
K
P*P*B
Noisy HSI Low-rank Matrix
P*P*K
B
Spatial
low-rank
Spectral
low-rank
P*P*BP*P*K
K
B
Cubic
matching
HSI
reconstruction
Estimated HSI
Spectral+Spatial
low-rank approximation
Aggregation
Fig. 1. Flowchart of the proposed HSI denoising algorithm.
We construct the low-rank matrix across spectral dimension
and spatial non-local similarity dimension, respectively.
al. [6] incorporated extra sparse constraint on the spectral in-
formation. Although these low-rank matrix recovery methods
have achieved impressive result in HSI denoising, they have
not fully exploit the abundantly spatial information in HSI.
Another research line follows the non-local self-similarity
perspective which has been used widely in single image de-
noising. The well-known BM3D method was also naturally
extended into BM4D [10] for volumetric data restoration. In
[7], the authors proposed a tensor dictionary learning model
via grouping similar patches for MSI denoising with hard
constraints on the rank of the core tensor. However, they have
neglected the exclusively spectral correlation property in HSI.
In this work, a unified low-rank approach is proposed to
simultaneously involve the spatial and spectral structure in-
formation to obtain a complete representation of HSI (Fig-