HYPERSPECTRAL IMAGE FUSION BASED ON NON-FACTORIZATION SPARSE
REPRESENTATION AND ERROR MATRIX ESTIMATION
Xiaolin Han
1
, Jiqiang Luo
2
, Jing Yu
3
, Weidong Sun
1
1. State Key Lab. of Intelligence Technology and Systems
Tsinghua National Lab. for Information Science and Technology
Dept. of Electronic Engineering, Tsinghua Univ., Beijing 100084, China
2. School of Opte-electronics, Beijing Institute of Technology, Beijing 100081, China
3. Colg. of Computer Science and Technology, Beijing Univ. of Technology, Beijing 100124, China
ABSTRACT
Matrix factorization with non-negative constrains is
widely used in hyperspectral image fusion. Nevertheless, the
non-negative restriction on the sparse coefficients limits the
efficiency of dictionary representation. To solve this
problem, a new hyperspectral image fusion method based on
non-factorization sparse representation and error matrix
estimation is proposed in this paper, for the fusion of
remotely sensed high-spatial multi-bands image with low-
spatial hyperspectral image in the same scene. Firstly, an
efficient spectral dictionary learning method is specifically
adopted for the construction of the spectral dictionary, which
avoids the procedure of matrix factorization. Then, the
sparse codes of the high-spatial multi-bands image with
respect to the learned spectral dictionary are estimated using
the alternating direction method of multipliers (ADMM)
without non-negative constrains. For improving the quality
of final fusion result, an error matrix estimation method is
also proposed, exploiting the spatial structure information
after non-factorization sparse representation. Experimental
results both on simulated and real datasets demonstrate that,
compared with the related state-of-the-art methods, our
proposed method achieves the highest quality of
hyperspectral image fusion, which can improve PSNR over
2.5844 and SAM over 0.3758.
Index Terms—hyperspectral image fusion, non-
factorization sparse representation, dictionary learning, error
matrix estimation.
1. INTRODUCTION
High-spatial hyperspectral images have been widely used
in various fields, such as environment monitoring, precision
agriculture, military detective and so on. Nevertheless, the
increasing of spectral bands in hyperspectal imaging, leads
This work was supported by the National Nature Science
Foundation (61501008) and the Natural Science Foundation of
Beijing (4172002).
to enlarging the size of the photosensitive elements, which
results in the limitations on spatial resolution [1]. Normally,
the spatial information lost in hyperspectral imaging, can be
obtained by fusing high-spatial multi-bands images through
some kind of post processing.
Recently, hyperspectral image fusion techniques have
been developed rapidly. Specifically, signal processing
based methods have been proposed to improve the spatial
resolution of hyperspectral image by fusing a high-spatial
panchromatic image with a low-spatial hyperspectral image
[2]. For further improvement of the fusion quality, relying
on the linear mixing model [3], unmixing based approach is
investigated in fusing multi-bands images, which assumes
each pixel as a mixture of a few materials corresponding to
the pure spectral signatures in spectral domain [4][5].
Besides, matrix factorization based approaches have played
a significant role in hyperspectral image fusion. Such as the
method for high-spatial hyperspectral imaging via matrix
factorization, proposed by Kawakami [6], decomposes the
image into a basis and a set of sparse coefficients. For
remotely sensed image, a similar method has been proposed
by Huang [7], where the pure spectral signatures are
conducted by a trained dictionary. A non-negative sparse
approach has also been proposed by Wyckoff [8], which
uses the alternating direction method of multipliers (ADMM)
for matrix factorization. Grohnfeldt [9] has considered a
joint sparse representation for the construction of the
dictionary pairs from the low-spatial hyperspectral and high-
spatial multi-bands image. Akhtar [1] has offered a sparse
spatial-spectral representation approach with non-negative
constrains, which assumes that the similar pixels have the
same abundance while estimating the sparse coefficients. A
non-negative structured sparse representation approach has
been proposed by Dong [10], which exploits the strong
spectral correlations in similar neighbors when the spectral
basis and sparse coefficients are estimated with the sparse
priors of the hyperspectral image jointly.
The non-negative matrix factorization, based on spectral
unmixing, is an effective technique for hyperspectral image
1155978-1-5090-5990-4/17/$31.00 ©2017 IEEE GlobalSIP 2017