IEEE TRANSACTIONS ON CYBERNETICS, VOL. 49, NO. 4, APRIL 2019 1417
Image Representation and Learning With
Graph-Laplacian Tucker Tensor Decomposition
Bo Jiang , Chris Ding, Jin Tang, and Bin Luo, Senior Member, IEEE
Abstract—Tucker tensor decomposition (TD) is widely used
for image representation, reconstruction, and learning tasks.
Compared to principal component analysis (PCA) models, ten-
sor models retain more 2-D characteristics of images whereas
PCA models linearize images. However, traditional TD involves
attribute information only and thus does not consider the pair-
wise similarity information between images. In this paper, we
propose a graph-Laplacian tucker tensor decomposition (GLTD)
which explores both attributes and pairwise similarity informa-
tion simultaneously. Generally, GLTD has three main benefits:
1) GLTD reconstruction shows clear robustness against image
occlusions/outliers. We provide analysis to show that Laplacian
regularization is mainly responsible to this robustness via an
out-of-sample GLTD model. To the best of our knowledge, this
Laplacian regularization induced robustness of TD has not been
studied or emphasized before; 2) GLTD representation per-
forms more regularity, which improves both unsupervised and
supervised learning results; and 3) an effective algorithm is
derived to solve GLTD problem. Although GLTD is a non-
covex problem, the proposed algorithm is shown experimentally
to provide a stable/unique solution starting from different
random initializations. Experimental results on image reconstruc-
tion, data clustering, and classification tasks show the benefits
of GLTD.
Index Terms—Classification, clustering, dimension reduction,
Laplacian regularization, tucker tensor decomposition (TD).
I. INTRODUCTION
I
MAGE representation and feature extraction is an important
research topic in computer vision and machine learning
area. In typical image representation, an image is first rep-
resented as a 1-D long feature vector which denotes one
data point in a high dimensional space. Then, many subspace
learning methods [1]–[5], such as principal component anal-
ysis (PCA) [6], [7], nonnegative matrix factorization [4], and
Manuscript received February 9, 2017; revised October 10, 2017, December
15, 2017, and January 22, 2018; accepted January 26, 2018. Date of publica-
tion February 19, 2018; date of current version February 22, 2019. This work
was supported by in part by the National Natural Science Foundation of China
under Grant 61602001, Grant 61572030, and Grant 61671018, in part by the
Natural Science Foundation of Anhui Province under Grant 1708085QF139,
and in part by the Natural Science Foundation of Anhui Higher Education
Institutions of China under Grant KJ2016A020. This paper was recommended
by Associate Editor A. Cichocki. (Corresponding author: Bo Jiang.)
B. Jiang, J. Tang, and B. Luo are with the School of Computer
Science and Technology, Anhui University, Hefei 230601, China (e-mail:
jiangbo@ahu.edu.cn; ahhftang@gmail.com; luobin@ahu.edu.cn).
C. Ding is with the Department of Computer Science and Engineering,
University of Texas at Arlington, Arlington, TX 76019 USA (e-mail: chqd-
ing@uta.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCYB.2018.2802934
linear discriminant analysis [6] etc., can be used for image
representation and feature extraction.
However, as we all know, the 1-D vector denotation of
image as a whole ignores the 2-D neighborhood feature
information within one image, and an image should be nat-
urally represented as a 2-D matrix. Therefore, many tensor
decomposition (TD) techniques have been studied [8]–[14].
Shashua and Levine [15] proposed an image representation
method by using a rank-one decomposition method. Ye [16]
proposed a generalized low rank approximation for matrices.
Ding and Ye [17] proposed a 2-D singular value decompo-
sition method. Zhang and Ding [18] proposed a robust TD
by using a
1
-norm loss function. Recently, some other ten-
sor methods have also been proposed [19]–[24]. The above
tensor methods can naturally retain the desired 2-D neighbor-
hood feature information within each image. However, one
limitation is that they generally fail to consider the geomet-
rical (manifold) structure of the image set X (e.g., distances
or similarities between images). This geometrical relationship
information has been shown importantly for data learning tasks
and thus has been widely used for data representation via
regularized method [2], [25]–[30].
Motivated by recent works on manifold learning and reg-
ularization, in this paper we propose graph-Laplacian tucker
tensor decomposition (GLTD) for image representation and
learning. GLTD aims to learn a low dimensional repre-
sentation for image data that incorporates the geometrical
structure information between images while maintains the
good data reconstruction property of Tucker TD method.
The idea of graph manifold learning has been employed
for TD method [31]–[33]. Wang et al. [32] developed a
graph Laplacian regularized nonnegative tensor factorization
method (LRNTF) for image representation. Wang et al. [34]
proposed a neighborhood preserving nonnegative tensor fac-
torization method (NPNTF) for image representation. The
aim of NPNTF is to incorporate LLE [3] into nonnegative
TD. Li et al. [33] also proposed a Laplacian regularized
nonnegative TD method for image representation and dimen-
sion reduction. For Tucker TD, similar idea has also been
proposed in work [31] with applications in tensor completion.
Differently, in this paper, we focus on image data represen-
tation with application in image data learning tasks. Also,
we derive a new algorithm to solve the proposed model by
adapting the higher-order orthogonal iteration (HOOI) [14]
algorithm. We show experimentally that the converged solu-
tion of GLTD is stable/unique with different initializations
although GLTD is a nonconvex model. This indicates the
2168-2267
c
2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Anhui University. Downloaded on April 08,2020 at 09:42:09 UTC from IEEE Xplore. Restrictions apply.