GROUP-BASED HYPERSPECTRAL IMAGE DENOISING USING LOW RANK
REPRESENTATION
Mengdi Wang
1
, Jing Yu
2
, Weidong Sun
1
1.State Key Lab. of Intelligent Technology & Systems
Tsinghua National Lab. for Information Science & Technology
Dept. of Electronic Engineering, Tsinghua University, Beijing 100084, China
2.College of Computer Science and Technology, Beijing University of Technology, Beijing 100124, China
Email: wangmd12@mails.tsinghua.edu.cn, jing.yu@bjut.edu.cn, wdsun@tsinghua.edu.cn
ABSTRACT
For the hyperspectral image(HSI) denoising problem, we pro-
pose a group-based low rank representation (GLRR) method.
A corrupted HSI is divided into overlapping patches and the
similar patches are combined into a group. The group is de-
noised as a whole using low rank representation(LRR). Our
method can employ both the local similarity within the patch
and the nonlocal similarity across the patches within a group
simultaneously, while nonlocal similar patches within the
group can bring extra structure information for the corrupted
patch, which makes the noise more significant to be detected
as outliers. Since the uncorrupted patches have an intrinsic
low-rank structure, LRR is employed for the denoising of the
patch group. Both simulated and real data are used in the
experiments. The effectiveness of our method is proved both
qualitatively and quantitatively.
Index Terms— Denoising, Hyperspectral image, Low
rank representation, Nonlocal similarity.
1. INTRODUCTION
In recent years, hyperspectral images(HSIs) have been widely
used in various fields, such as environment monitoring, agri-
culture, military and so on. However, corruption by various
noises degrades the image quality greatly, leading to the low
accuracy of classification, object segmentation, unmixing and
subpixel mapping. Therefore, the denoising procedure is an
essential preprocessing step before the following applications
of HSIs.
Over the past several decades, many kinds of denoising
algorithms have been proposed using different framework.
Othman[1] reduces the noise using a wavelet-shrinkage fil-
ter in the hybrid spatial-spectral derivative domain. Zhang
[2] proposed to employ a cubic total variation regularization
This work was supported in part by the National Nature Science
Foundation (No.61171117), National Science & Technology Pillar Program
(No.2012BAH31B01) and Key Project of the Science & Technology Devel-
opment Program of BEC (No.kz201310028035) of China.
and Yuan [3] extended the regularization item into a spectral-
spatial adaptive one. In [4] and [5], tensor analysis is em-
ployed, with Tucker decomposition and rank-1 decomposi-
tion applied respectively. Zhang [6] introduced a denoising
method based on low rank representation (LRR). The method
received overwhelmingly good performance upon other meth-
ods as the uncorrupted HSI complies with the assumption of
low rank structure highly. However, within the corrupted area,
the noisy patches have a high percentage of different kinds of
noises, which makes the specific noise is not significant for
detection.
In this paper, we propose a novel algorithm for HSI de-
noising using group-based low rank representation(GLRR),
in which local and nonlocal similarity of the HSI can be si-
multaneously considered under the united framework of LR-
R, while the traditional LRR method just employs local sim-
ilarity. In GLRR, similar patches are combined into a group,
while the similarity across the patches within a group we de-
note as nonlocal similarity. The reconstructed unit with LRR
is the group consisting of similar patches, instead of the indi-
vidual patches. Within the group, the nonlocal similar patches
bring auxiliary structure information for the corrupted patch,
which makes the noise more significant to be detected.
The remainder of this paper is organized as follows. Sec-
tion 2 describes the proposed method in detail. Experimental
results and discussion are shown in Section 3. Section 4 draws
the conclusion.
2. PROPOSED METHOD
2.1. Low Rank Representation
Low rank representation is originally proposed in [7] by
Wright et al. The model is under the assumption that the
uncorrupted data has a low-rank structure. And the low-rank
property of HSIs has been demonstrated in [8, 9] from the
perspective of a linear spectral unmixing model.
A corrupted matrix X ∈ R
m×n
is modelled as combina-