Neurocomputing 204 (2016) 198–210

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Face recognition using class specific dictionary learning for sparse

representation and collaborative representation
Bao-Di Liu a,n, Bin Shen b, Liangke Gui c, Yu-Xiong Wang c, Xue Li e, Fei Yan d,
Yan-Jiang Wang a
a
College of Information and Control Engineering, China University of Petroleum, Qingdao 266580, China
b
Department of Computer Science, Purdue University, West Lafayette, IN 47907, USA
c
School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, USA
d
Lijin County Party Committee Office, Dongying 257499, China
e
Department of Electronic Engineering, Tsinghua University, Beijing 100084, China

art ic l e i nf o a b s t r a c t

Article history: Recently, sparse representation based classification (SRC) and collaborative representation based clas-
Received 2 March 2015 sification (CRC) have been successfully used for visual recognition and have demonstrated impressive
Received in revised form performance. Given a test sample, SRC or CRC formulates its linear representation with respect to the
7 July 2015
training samples and then computes the residual error for each class. SRC or CRC assumes that the
Accepted 20 August 2015
Available online 8 April 2016
training samples from each class contribute equally to the dictionary in the corresponding class, i.e., the
dictionary consists of the training samples in that class. This, however, leads to high residual error and
Keywords: instability. To overcome this limitation, we propose a class specific dictionary learning algorithm. To be
Class specific dictionary learning specific, by introducing the dual form of dictionary learning, an explicit relationship between the basis
Sparse representation
vectors and the original image features is represented, which also enhances the interpretability. SRC or
Collaborative representation
CRC can be thus considered as a special case of the proposed algorithm. Blockwise coordinate descent
Face recognition
algorithm and Lagrange multipliers are then adopted to optimize the corresponding objective function.
Extensive experimental results on five benchmark face recognition datasets show that the proposed
algorithm achieves superior performance compared with conventional classification algorithms.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction
Face recognition is a classical yet challenging research topic in
computer vision and pattern recognition [33]. Two stages are
usually considered for effective face recognition: (1) feature
extraction, (2) classifier construction and label prediction. For the
first stage, Turk and Pentland [25] proposed eigenfaces by per-
forming principal component analysis (PCA). He et al. [8] proposed
laplacianfaces to preserve local information. Belhumeur et al. [2]
suggested fisherfaces to maximize the ratio of between-class
scatter to within-class scatter. Yan et al. [28] proposed a multi-
subregion based correlation filter bank algorithm to extract both
the global-based and local-based face features. For the latter stage,
Richard et al. [19] proposed a nearest neighbor method to predict
the label of a test image using its nearest neighbors in the training
n
Corresponding author. Tel.: þ 86 15764217948.
E-mail addresses: thu.liubaodi@gmail.com (B.-D. Liu),
stanshenbin@gmail.com (B. Shen), guiliangke@gmail.com (L. Gui),
yuxiongw@cs.cmu.edu (Y.-X. Wang), lixue421@gmail.com (X. Li),
232842281@qq.com (F. Yan), yjwang@upc.edu.cn (Y.-J. Wang).
samples. Ho et al. [9] and Tao et al. [23] proposed a nearest sub-
space method to assign the label of a test image by comparing its
reconstruction error for each category.
Under the nearest subspace [44, 45] framework, Wright et al.
[27] proposed a sparse representation based classification (SRC)
system and achieved impressive performance. Given a test sample,
sparse representation techniques represent it as a sparse linear
combination of the training samples. The predicted label is
determined by the residual error from each class. Different from
traditional decomposition frameworks like PCA, non-negative
matrix factorization [39], and low-rank factorization [40], SRC
allows coding under over-complete bases, and thus makes the
attained sparse codes capable of representing the data more
adaptively and flexibly. To analyze SRC, Zhang et al. [31] proposed
collaborative representation based classification (CRC) as an
alternative approach. CRC represents a test sample as the linear
combination of almost all the training samples. They found that it
is the collaborative representation rather than the sparse repre-
sentation that makes the nearest subspace method powerful for
classification. SRC, CRC, and their variants have been also used in
other visual data sensing and analysis tasks, such as image

http://dx.doi.org/10.1016/j.neucom.2015.08.128
0925-2312/& 2016 Elsevier B.V. All rights reserved.
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210
classification [12], image inpainting [41], object detection [42],
image annotation [1], and transfer learning [43].
Despite their promise, both SRC and CRC algorithms directly use
the training samples as the dictionary for each class. By contrast, a
well learned dictionary, especially by enforcing some discriminative
criteria, can reduce the residual error greatly and achieve superior
performance for classification tasks. Existing discriminative dic-
tionary learning approaches are mainly categorized into three
types: shared dictionary learning, class specific dictionary learning,
and hybrid dictionary learning. In shared dictionary learning, each
basis is associated to all the training samples. Mairal et al. [16]
proposed to learn a discriminative dictionary with a linear classifier
of coding coefficients. Liu et al. [15] learned a Fisher discriminative
dictionary. Liu et al. [35, 36, 37] and Yu et al. [38] presented a graph
embedded dictionary learning method. Zhang and Li [32] proposed
a joint dictionary learning algorithm for face recognition. In class
specific dictionary learning, each basis only corresponds to a single
class so that the class specific reconstruction error could be used for
classification. Yang et al. [30] learned a dictionary for each class with
sparse coefficients and applied it for face recognition. Sprechmann
and Sapiro [22] also learned a dictionary for each class with sparse
representation and used it in signal clustering. Castrodad and Sapiro
[4] learned a set of action specific dictionaries with non-negative
penalty on both dictionary atoms and representation coefficients.
Wang et al. [26] introduced mutual incoherence information to
promote class specific dictionary learning in action recognition.
Yang et al. [29] embedded the Fisher discriminative information
into class specific dictionary learning.
The shared dictionary learning approaches usually lead to a
dictionary of small size and the discriminative information (i.e.,
the label information corresponding to coding coefficients) is
embedded into the dictionary learning framework. The class spe-
cific dictionary learning approaches usually focus on the classifier
construction aspect since each basis vector is fixed to a single class
label. The combination of shared basis vectors and class specific
basis vectors is then learned in hybrid dictionary learning. Zhou
et al. [34] learned a hybrid dictionary with Fisher regularization on
the coding coefficient. Gao et al. [6] learned a shared dictionary to
encode common visual patterns and learned a class specific dic-
tionary to encode subtle visual differences among different cate-
gories for fine-grained image representation. Liu et al. [12] pro-
posed a hierarchical dictionary learning method to produce a
shared dictionary and a cluster specific dictionary. In spite of the
Fig. 1. The framework of our proposed class specific dictionary learning algorithm.
199
demonstrated performance of hybrid dictionary learning, it is still
a challenge to balance between the shared dictionary and the class
specific dictionary.
In this paper, motivated by the superior performance of the SRC
and CRC algorithms and the class specific dictionary learning
method, we propose class specific dictionary learning (CSDL) for
both sparse representation based classifier (CSDL-SRC) and colla-
borative representation based classifier (CSDL-CRC). Fig. 1 shows
the framework of our proposed CSDL. The major distinction
between our approach and the existing class specific dictionary
learning methods is that the existing methods directly optimize
the dictionary basis vectors (the “primal” form), whereas we
leverage a “dual” reformulation of dictionary learning and opti-
mize the weights instead. Compared with the “primal” form
widely used by the existing methods, our novel “dual” form offers
several benefits:
 It provides an explicit relationship between the basis vectors
and the original image features, thus enhancing the interpret-
ability of the learned dictionary.
 It is easy to be kernelized due to the separation of original data,
which is difficult for the existing methods. The generalization to
kernel spaces will be elaborated in Section 5.7.3.
 Most of the existing class specific dictionary learning methods
focus on introducing additional regularization terms, which
could be easily incorporated into our dual formulation of class
specific dictionary learning to further improve the performance.
Our main contributions are threefold:
 We propose a novel class specific dictionary learning scheme
that considers the weight of each sample when generating the
dictionary (i.e., subspace). The traditional CRC and SRC methods
perform face recognition without training procedures (i.e., the
training samples are directly used for predicting the labels). By
contrast, our proposed method compensates this deficiency by
introducing class specific dictionary learning. It is applicable to
both CRC and SRC. Furthermore, it is reasonable to assume
additionally that different samples contribute unevenly in
constructing the corresponding subspace. CRC or SRC can be
thus viewed as special cases of our proposed algorithm.
 We propose the dual form of dictionary learning to enhance the
interpretability.
200 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210

 We propose a simple optimization procedure that combines 2: for c ¼ 1; c rC; c þ þ do

blockwise coordinate descent and Lagrange multipliers to solve
the minimization problem.
A preliminary version of this work appeared as Liu et al. [13].
We clarify many technical details omitted in the previous version,
present results on substantially extended datasets (e.g., Yale
dataset and ORL dataset), and provide an in-depth analysis of how,
why, and when the proposed algorithm is superior to the standard
approaches. These extensive experimental evaluations demon-
strate that our proposed algorithm achieves impressive perfor-
mance for face recognition tasks.
The rest of the paper is organized as follows. Section 2 reviews
sparse representation and collaborative representation based
classifier algorithms. Section 3 explains the proposed class specific
dictionary learning for sparse representation and collaborative
representation methods. The solution to the corresponding opti-
mization problem together with its convergence analysis is ela-
borated in Section 4. The overall algorithm is also summarized.
The application of the proposed CSDL algorithm in face recognition
is shown in Section 5. Finally, discussions and conclusions are
drawn in Section 6.
2. Overview of SRC and CRC
In this section, two classical classifier algorithms, sparse
representation based classifier and collaborative representation
based classifier, are briefly reviewed. Sparse representation and
collaborative representation algorithms can be considered as
methods of rearranging the structure of the original data in order
to make the representation compact under over-complete and
non-orthogonal bases. Hence, the data vector is represented as a
linear combination of active basis vectors.
2.1. Sparse representation based classification
Wright et al. [27] proposed the sparse representation based
classification (SRC) algorithmh for robust
i face recognition. Given
the training samples X ¼ X 1 ; X 2 ; …; X C A RDN , X c A RDNc repre-
sents the training samples from the cth class, C represents the
number of classes, Nc represents the number of training samples in
P
the cth class (N ¼ Cc ¼ 1 N c ), and D represents the dimension of the
samples. Supposing that y A RD1 is a test sample, the sparse
representation algorithm aims to solve the following objective
function:
n 2 o sentation or collaborative representation.
s^ ¼ arg min y  Xs2 þ 2αksk1 : ð1Þ
s
Here, Eqn. (1) is the ℓ1-norm regularizedleast-squares(ℓ1-ℓs)
minimization problem. s A RN1 is the sparse code of y. α is the
 c 2
3: Compute the residuals ec ðyÞ ¼ y  X c s^ 2
4: end for
5: idðyÞ ¼ arg minc fec g
6: return id(y)
2.2. Collaborative representation based classification
Zhang et al. [31] proposed the collaborative representation
based classification (CRC) by replacing the objective function in
SRC with ℓ2 regularized minimization problems as
n 2 o
s^ ¼ arg min y  Xs2 þ βksk22 : ð3Þ
s
Here, s A RN1 is the fitting coefficient of y. β is the regularization
parameter to control the tradeoff between fitting goodness and
collaborative property (i.e., multiple entries in X participating into
representing the test sample).
The residual error ec in Algorithm 1 is associated with most of
the images in class c. Both SRC and CRC algorithms directly use the
training samples as the dictionary and encode the test sample y as
y  XWs; ð4Þ
where W A RNN is an identity matrix. This means that the training
samples contribute equally for constructing the dictionary B ¼ XW
when representing the test sample y.
3. Proposed class specific dictionary learning algorithm
From Eqns. (1), (3) and (4), we could observe that W is pre-
defined as an identity matrix I. However, for different classification
tasks and sample distribution, one would favor a data-driven for-
mulation of W. That is, to make W more adaptive, it would be of
great benefit to impose that the training samples of the same class
have different weights when constructing bases in the corre-
sponding dictionary while the training samples have no con-
tribution when constructing bases in the different classes of dic-
tionary. Hence, we generalize the weight coefficient matrix from
W¼I in the cases of SRC and CRC to a block-diagonal matrix as
given in Fig. 2.
W can be obtained effectively by dictionary learning. SRC and
CRC can be thus considered as a special case of our proposed class
specific dictionary  learning. Now, ec is calculated by
ec ðyÞ ¼ y  X c s^ c 2 . The with-in class residual error will further
decrease via learning an adaptive dictionary for sparse repre-
These motivate our dictionary learning for sparse representa-
tion, whose objective function is
( )
  XNc

regularization parameter to control the tradeoff between fitting f W c ; Sc ¼ kX c  X c W c Sc ‖2F þ 2α ‖Sc•n ‖1

n¼1
goodness and sparseness. The sparse representation based classi-  c c 2
s:t: X W  r 1; ∀k ¼ 1; 2; …; K: ð5Þ
fier is to find the minimum value of the residual error for each •k F

class:
 c 2
idðyÞ ¼ arg minc y  X c s^ 2 : ð2Þ ⎡W 1 0 0 L ⎤ 0
c
s^ represents the partial sparse codes related to images in class c. ⎢ 2 ⎥
The procedure of SRC is shown in Algorithm 1. The impressive ⎢0 W 0 L 0 ⎥
results of SRC have been reported in Wright et al. [27]. ⎢0 0 W3 L 0 ⎥
Algorithm 1. Algorithm for SRC.
⎢ ⎥
⎢L L L L L ⎥
Require: Training samples X A RDN , α, and test sample y ⎢ ⎥
Code y with the dictionary X via ℓ1-minimization. ⎢⎣0 0 0 L WC⎥
1:
n 2 o ⎦
s^ ¼ arg mins y  Xs þ 2αksk1
2
Fig. 2. The learned weight coefficient matrix W for constructing dictionary.
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210 201

Similarly, the objective function of learning dictionary for col- Based on the convexity and monotonic property of the parabolic
laborative representation is function, it is easy to know that f ðSckn Þ reaches the minimum at the
       unique point
f W c ; Sc ¼ X c  X c W c Sc 2F þ βSc 2F
 c c 2
Sckn ¼ minf½W c X c X c kn  ½ES~c kn ;  αg
kn
s:t: X W •k F r 1; ∀k ¼ 1; 2; …; K: ð6Þ T T

þ maxf½W c X c X c kn  ½ES~c kn ; αg;

T T kn
In Eqns. (5) and (6), c is the cth class and K is the size of the ð10Þ
learned dictionary. In the paper, K is set to be twice of the number ( c
S ; p ak J q a n
where E ¼ W c X c X c W c , S~c ¼
T T kn pq
of the training samples in c. Wc is thus the learned weight coef-
0; p ¼ k&q ¼ n:
ficient for constructing the dictionary and Sc is the corresponding
sparse representation or collaborative representation. Eqns. Furthermore, given that the optimal value for Sckn does not
depend on the other entries in the same row, each entire row of Sc
(5) and (6) can be intuitively viewed as the “dual” reformulations
can be optimized simultaneously. That is,
of the “primal” form of the sparse representation and collaborative
Sck ¼ minf½W c X c X c k  ½ES~c k ;  αg
representation problems, respectively. This is better understood if T T k

one draws an analogy with linear SVM training: one can formulate
þ maxf½W c X c X c k  ½ES~c k ; αg;
T T k
the training problem as an optimization over (1) either directly the ð11Þ
( c
weights of a linear classifier vector of the same dimension as the Sp ; p a k
where S~c ¼
k
input signal (2) or the support vector weights, that is a vector of
0; p ¼ k:
dimension equal to the size of the training set that is used to
linearly combine the training inputs. In the context of the problem 4.1.2. ℓ2–ℓs minimization subproblem
here, the support vector weights are analogous to the weights W. With Sc fixed , the objective function of ℓ2–ℓs subproblem
The atoms now capture a notion of centroids similar to K-means, becomes
which explicitly expresses what happens during dictionary      
f W c ¼ X c  X c W c Sc 2F s:t: X c W c•k 2F r 1; ∀k ¼ 1; 2; …; K: ð12Þ
learning, leading to enhanced interpretability.
Here, the Lagrange multipliers are used to solve the ℓ2-norm
constrained minimization subproblem.
4. Optimization of the objective function W c can be optimized in a column-wise manner. Specifically,
T
ignoring the constant term trfX c X c g, its Lagrangian is
In this section, we focus on solving the optimization problem of K h
X i
LðW c ; λk ; μk Þ ¼  2
T
the two types of CSDL algorithms proposed above. Specifically, Sc X c X c W c•k
k
similar to the optimization strategy used in Lee et al. [11] and Liu k¼1

et al. [14], we cast the problem as two subproblems via alternating X

K h i
T T T
þ W c•k X c X c W c Sc Sc
minimization: ℓ1-norm or ℓ2-norm regularized least-squares k
k¼1
minimization subproblem with W c fixed, and ℓ2-norm con-  h i
strained least-squares minimization subproblem with Sc fixed. þ λk 1  W X X W cT cT c c
; ð13Þ
kk

where λk is the Lagrange variable.

4.1. Class specific dictionary learning for sparse representation
According to the Karush–Kuhn–Tucker (KKT) conditions, the
optimal solution W c•k should satisfy the following criteria:
The sparse representation based class specific dictionary
learning problem can be decomposed into ℓ1-norm regularized ∂LðW c ; λk Þ
ðaÞ : ¼ 0; ð14Þ
least-squares (ℓ1-ℓs) minimization subproblem with W c fixed and ∂W c•k
ℓ2-norm constrained least-squares (ℓ2-ℓs) minimization sub-
problem with Sc fixed.
T T
ðbÞ : ð1  ½W c X c X c W c kk Þ ¼ 0
4.1.1. ℓ1–ℓs minimization subproblem λk a 0: ð15Þ
With W c fixed, the objective function of ℓ1-ℓs subproblem is Hence, the solution to W c ð:; kÞ becomes
  X
Nc 
W c•k ¼ Sc•k  ½W~ c F•k sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T k
f Sc ¼ ‖X c  X c W c Sc ‖2F þ 2α Sc k1 : ð7Þ      ;
•n
Sc  W~ c F ~ c kF
k
n¼1 T X c T X c Sc  W
T T
•k •k •k
Eqn. (7) can be simplified as •k

ð16Þ
  n o N h
X i (
T T
f Sc ¼ tr X c X c  2 X c X c W c Sc•n W•pc; p≠k
where F ¼ Sc Sc and W~ c ¼
T k
n•
n¼1
0; p ¼ k:
X
N h i X
K X
N 
þ
T T T
Sc•k W c X c X c W c Sc•n þ 2α Sc j: ð8Þ Algorithm 2. CSDL algorithm for sparse representation.
kn
n¼1 k¼1n¼1
Require Data matrix X c A RDN , α, and K
T
Ignoring the constant term trfX c X c g, the objective function of Sckn 1: W c ’randðN c ; KÞ; Sc ’zerosðK; N c Þ
c 
reduces
 to Eqn. (9) with W and S1n ; S2n ; …; SKn ⧹Sckn fixed. Here,
c c c
2: for k ¼ 1; k r K; k þ þ do
( Sc1n ; Sc2n ; …; ScKn ⧹Sckn stands for all elements in Sc except the qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3: T
element in the kth row and nth column): W c•k ¼ W c•k = W c•k W c•k
( ) 4: end for
  nh i o XK h i
T T T T
f Sckn ¼ Sckn 2 W c X c X c W c kk þ2Sckn W c X c X c W c kl Scln 5: iter ¼ 0
l ¼ 1;l≠k 6: while ðf ðiterÞ  f ðiter þ 1ÞÞ=f ðiterÞ 4 1e  5 do
nh i o  7: iter’iter þ 1
2Sckn cT
W X X cT c
þ 2αSckn j: ð9Þ
kn 8: Update Sc :
9: T T
T T
Here, ½W c X c X c W c kk ¼ 1. f ðSckn Þ is a piece-wise parabolic function. E ¼ W c Xc XcW c
202 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210

10: for k ¼ 1; k rK; k þ þ do For ℓ2–ℓs regularized minimization subproblem, the objective
11: function is
¼ minf½W X X k  ½ES~c k ;  αg
cT k
cT
Sck c
 2  2
12:
þ maxf½W c X c X c k  ½ES~c k ; αg
T T k
f ðSc Þ ¼ X c  X c W c Sc F þ βSc 2 : ð18Þ
13: end for Eqn. (18) can be easily solved by derivation and its analytical
14: Update W c : solutions is
15: T
Compute F ¼ Sc Sc , G ¼ F  ð1 IÞ  1
16: for k ¼ 1; k rK; k þ þ do Sc ¼ W Tc X Tc X c W c þ βI W Tc X Tc X c : ð19Þ
17: T
W c•k ¼ Sc•k  W c G•k
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi For ℓ2–ℓs constrained minimization subproblem, the objective
18: T T
W c•k ¼ W c•k = W c•k X c X c W c•k function is the same as Eqn. (12), and the solution thus remains as
19: end for Eqn. (16).
20: Update the objective function:
21: P
Nc 4.3. CSDL for label prediction
f ¼ ‖X c  X c W c Sc ‖2F þ 2α ‖Sc•n k1
n¼1
22: end while Algorithm 3 shows our proposed CSDL-SRC or CSDL-CRC algo-
23: return W c , and Sc rithm for predicting the label of a test sample.

Algorithm 3. Algorithm for label prediction of our proposed CSDL

4.1.3. Convergence analysis and discussion
The objective function in Eqn. (5) is nondifferentiable and
nonconvex. Under our optimization framework, it is partitioned as
2K blocks in which each block is a column of Wc or a row of Sc:
  X
K 3:
f W c1 ; …; W cK ; Sc1 ; …; ScK ¼ ‖X c  X c W ck Sck ‖2F þ 2αkSck k1
k¼1
 c c 2
s:t: X W  r 1; 8 k ¼ 1; 2; …; K:
k 2
Since the exact minimization point is obtained by Eqn. (11) or
Eqn. (16), each operation updates Sc1 ; …; ScK ; W c1 ; …; W cK alter-
nately, and it monotonically decreases the objective function in
Eqn. (5). Considering that the objective function is obviously
bounded below, it converges.
Now consider the convergence of the variables W c and Sc in the
function. Each subproblem with respect to W c1 ; W c2 ; …; W cK , Sc1 ;
Sc2 ; …; ScK becomes convex. It thus satisfies the separability and
regularity properties proposed in Tseng [24], and the optimal value
for a given block of variables is uniquely achieved by the solution
Eqns. (11) and (16) due to the strict convexity of subproblems at
each iteration. These properties make the subsequence (W c and
Sc ) via alternating minimization by block coordinate descent
converge to a stationary point. The detailed proof of the con-
vergence of (blockwise) coordinate descent for functions satisfying
some mild conditions is referred to as Theorem 4.1 in Tseng [24]
and also in Bertsekas [3].
The time complexity of one iteration round for sparse coding
(obtaining S) and weight learning (learning W) are OðK 2 D þ2KDN
þ K 2 NÞ and OðKD þ KDN þ 2K 2 NÞ, respectively. Here, K represents
the number of bases, D represents the dimensionality of features,
and N represents the number of features.
4.1.4. Overall algorithm
Our algorithm for sparse representation based class specific
dictionary learning is shown in Algorithm 2.
Here, 1 A RKK is a square matrix with all elements 1, I A RKK is
the identity matrix, and  indicates the Hadamard product. By
iterating Sc and W c alternately, the sparse codes are obtained, and
the corresponding bases are learned.
4.2. Class specific dictionary learning for collaborative representation
Similar to the above procedure, the collaborative representa-
tion based class specific dictionary learning problem can be
decomposed into ℓ2-norm regularized least-squares (ℓ2–ℓs)
minimization subproblem with W c fixed and ℓ2-norm constrained
least-squares (ℓ2–ℓs) minimization subproblem with Sc fixed.
algorithm.
Require Training samples X A RDN , α, β, and test sample y
1: for c ¼ 1; c r C; c þ þ do
2: Code y with the learned Wc via Eqn. (11) or Eqn. (19).
 
Compute the residuals ec ðyÞ ¼ y X c W c sc 2
4: end for
5: idðyÞ ¼ arg minc fec g
ð17Þ
6: return id(y)
5. Experimental results
In this section, we use five standard benchmark datasets,
including the Yale dataset [2], ORL dataset [20], Extended YaleB
dataset [7], CMU PIE dataset [21], and AR dataset [17], to evaluate
the performance of our proposed class specific dictionary learning
(CSDL) algorithm. We first describe experimental settings, and
then present the corresponding experimental results on these
datasets. Finally, we give some additional analysis and discussion
with respect to dictionary size, visualization and generalization to
kernel spaces.
5.1. Experimental settings
For the dataset, each face image is cropped to 32  32, pulled
into column vector, and ℓ2 normalized to form the raw ℓ2 nor-
malized feature. After that, power normalization is also performed
to produce the power normalized features. To eliminate the ran-
domness of the experiment, the data are randomly split into the
training set and the testing set 10 times, respectively. The mean
and standard deviation of the face recognition rate are reported.
To show the performance of our proposed CSDL algorithm, four
classical algorithms are used as the baseline. They are nearest
neighbor classification (NN), collaborative representation based
classification (CRC) [31], sparse representation based classification
(SRC) [27], and SVM [5]. Given a test sample, CRC and SRC compute
its linear representation with respect to all the training sample,
and use the fitting error as criterion for classification. For SVM, one
against all multi-class classification strategy is used by LIBSVM [5].
For the Yale dataset and ORL dataset, we randomly select
5 samples per class for training and the rest samples per class for
testing. For the AR dataset, 7 samples per class are used for
training and the rest samples per class are used for testing. For the
Extended YaleB dataset and CMU PIE dataset, 10 and 20 samples
per class are used for training and testing, respectively.
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210 203

Fig. 3. Examples of the Yale dataset.

Fig. 4. Influence of α for SRC & CSDL-SRC and β for CRC & CSDL-CRC on the Yale dataset, respectively.

Three parameters need tuning in our proposed CSDL algorithm.

α is used to adjust the tradeoff between the reconstruction error Table 1
and the sparsity for CSDL-SRC. β is used to adjust the tradeoff Recognition rate on the Yale dataset (%).
between the reconstruction error and the collaborative property
for CSDL-CRC. K is the dictionary size for each class. α is varied on Methods Features
the grid of f10  5 ; 10  4 ; 10  3 ; 10  2 ; 10  1 g. β is varied on the grid
ℓ2 norm Power norm
of f10  5 ; 10  4 ; 10  3 ; 10  2 ; 10  1 ; 100 g. K is set to be twice of the
number of the training samples per class. The detailed parameter NN 55.78 7 4.15 60.447 4.97
tuning procedure is shown in the following subsection. SVM 76.00 7 3.93 78.22 7 2.47
CRC 79.677 1.82 82.117 1.10
SRC 77.677 2.37 81.337 2.45
5.2. Yale dataset
CSDL-CRC 80.897 1.37 83.56 7 2.21
CSDL-SRC 79.00 72.89 84.567 1.61
The Yale dataset contains 165 grayscale face images of 15
individuals. For each individual, the images are captured under
variant facial expression or configuration, such as center-light,
wear glasses, happy, left-light, w/no glasses, normal, right-light,
81.33%, and for CSDL-SRC α is 0.01 with recognition rate 84.56%.
sad, sleepy, surprised, and wink. Fig. 3 shows some sample
Table 1 compares the recognition rate of NN, SVM, CRC, SRC,
images from the dataset. The performance with different values
CSDL-CRC, and CSDL-SRC. From Table 1, our proposed CSDL
of α or β is reported in Fig. 4. From Fig. 4, using ℓ2 normalized
algorithm achieves superior performance to the other four
features, for CRC the optimal β is 0.01 with recognition rate
classical classification methods.
79.67%, and for CSDL-CRC β is 0.1 with recognition rate 80.89%;
for SRC the optimal α is 0.001 with recognition rate 77.67%, and
for CSDL-SRC α is 0.001 with recognition rate 79%. Using power 5.3. ORL dataset
normalized features, for CRC the optimal β is 0.1 with recogni-
tion rate 82.11%, and for CSDL-CRC β is 0.1 with recognition rate For the ORL dataset, there are 10 different images of each of
83.56%; for SRC the optimal α is 0.01 with recognition rate 40 distinct subjects. For some subjects, the images were taken at
204 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210

Fig. 5. Examples of the ORL dataset.

Fig. 6. Influence of α for SRC & CSDL-SRC and β for CRC & CSDL-CRC on the ORL dataset, respectively.

different times, varying lighting, facial expressions (open/closed

Table 2
eyes, smiling/not smiling), and face facial details (glasses/no Recognition rate on the ORL dataset (%).
glasses). All the images were taken against a dark homogeneous
background with the subject in an upright, frontal position. Methods Features

Fig. 5 shows some sample images from the dataset. The per- ℓ2 norm Power norm
formance with different values of α or β is reported in Fig. 6.
Table 2 shows the recognition rate of NN, SVM, CRC, SRC, CSDL- NN 85.8 73.18 84.25 72.82
SVM 94.65 72.57 93.7 72.95
CRC, and CSDL-SRC. From Table 2, our proposed CSDL algorithm CRC 94.6 72.39 93.75 72.29
achieves superior performance to the other four classical clas- SRC 93.65 72.22 92.157 1.55
sification methods. CSDL-CRC 96.457 1.61 95.45 71.96
CSDL-SRC 95.55 71.50 95.507 1.45

5.4. Extended YaleB dataset

For the Extended YaleB dataset, there are 2,414 frontal face
images and 38 individuals in total. All the images are captured
under varying illumination conditions. Fig. 7 shows some sample
images from the dataset. The performance with different values of
α or β is reported in Fig. 8. Table 3 shows the recognition rate of
NN, SVM, CRC, SRC, CSDL-CRC, and CSDL-SRC. From Table 3, our
proposed CSDL algorithm achieves superior performance to the
other four classical classification methods.
5.5. CMU PIE dataset
The CMU PIE dataset contains 41,368 images of 68 indivi-
duals in total. Each individual is under 13 different poses, 43
different illumination conditions, and with 4 different expres-
sions. Each individual thus may lie on multiple manifolds. Five
near frontal poses (C05, C07, C09, C27, C29) and all different
illuminations and expressions are used in our experiment. There
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210 205

Fig. 7. Examples of the Extended YaleB dataset.

Fig. 8. Influence of α for SRC & CSDL-SRC and β for CRC & CSDL-CRC on the Extended YaleB dataset, respectively.

are about 170 images for each individual and 11,554 images in
total. Fig. 9 shows some sample images from the dataset. The Table 3
performance with different values of α or β is reported in Fig. 10. Recognition rate on the Extended YaleB dataset (%).
Table 4 shows the recognition rate of NN, SVM, CRC, SRC, CSDL-
Methods Features
CRC, and CSDL-SRC. From Table 4, our proposed CSDL algorithm
achieves superior performance to the other four classical clas- ℓ2 norm Power norm
sification methods.
NN 53.147 1.24 51.46 71.49
SVM 85.767 0.80 94.82 70.76
5.6. AR dataset CRC 89.96 7 1.04 97.377 0.74
SRC 89.92 7 1.09 97.13 70.56
For the AR dataset, there are over 4,000 frontal faces for 126 CSDL-CRC 90.36 7 0.90 97.747 0.70
individuals. A subset consisting of 50 male and 50 female cate- CSDL-SRC 91.467 0.64 98.287 0.57
gories is used here. There are 26 face images for each class.
Compared with the two former datasets, the AR dataset contains
more facial variations, such as illumination change, various comparison with state-of-the-art dictionary learning based face
expressions, and facial disguises. Fig. 11 shows some samples face recognition methods is presented.
images from the dataset. The performance with different values of
α or β is reported in Fig. 12. Table 5 shows the recognition rate of 5.7.1. The size of bases K per class
NN, SVM, CRC, SRC, CSDL-CRC, and CSDL-SRC. From Table 5, our Fig. 13 shows the influence of the size of bases K per class for
proposed CSDL algorithm achieves superior performance to the our proposed CSDL-CRC and CSDL-SRC. From Fig. 13, when K is
other four classical classification methods. small, e.g., K ¼5, the performance of CSDL is relative poor. When K
is larger than twice of the number of the training samples per
5.7. Analysis and discussions class, the performance becomes high and steady.

In this section, first, the size of bases per class is briefly ana- 5.7.2. Visualization of the learned dictionary
lyzed. Second, the learned dictionary is visualized and analyzed. Fig. 14 shows the comparison of the original face images and
Third, the generalization to kernel space is illustrated. Finally, the class specific learned dictionary with the size 20 (the learned
206 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210

Fig. 9. Examples of the CMU PIE dataset.

Fig. 10. Influence of α for SRC & CSDL-SRC and β for CRC & CSDL-CRC on the CMU PIE dataset.

Table 4 ¼ ½ϕðX 1 Þ; ϕðX 2 Þ; …; ϕðX N Þ. Then, the objective function of

Recognition rate on the CMU PIE dataset (%). learning dictionary for sparse representation, Eqn. (5), can be
generalized to reproducing kernel Hilbert spaces as
Methods Features
( )
      X
Nc
ℓ2 norm Power norm f W c ; Sc ¼ ‖ϕ X c  ϕ X c W c Sc ‖2H þ 2α ‖Sc•n ‖1
n¼1
NN 44.78 71.48 44.187 1.41   c  c 2
s:t: ϕ X W  r 1; ∀k ¼ 1; 2; …; K: ð20Þ
SVM 85.067 0.61 85.54 7 0.79 •k H
CRC 86.54 7 0.76 88.0770.85
SRC 86.517 0.85 84.3570.79
Similarly, the objective function of learning dictionary for col-
laborative representation, Eqn. (6), can be generalized to repro-
CSDL-CRC 88.91 70.81 89.747 0.86
ducing kernel Hilbert spaces as
CSDL-SRC 90.157 0.72 91.32 70.56
          
f W c ; Sc ¼ ϕ X c  ϕ X c W c Sc 2H þ βSc 2H
  c  c 2
dictionary for each class is obtained by XWc. Fig. 14 demonstrates s:t: ϕ X W •k H r 1; ∀k ¼ 1; 2; …; K: ð21Þ
one representative class of training face images and the corre-
Now, the Frobenius norm has been replaced by the inner-
sponding dictionary from the Extended YaleB dataset). From
product norm of that Hilbert space, such that J ϕðX c Þ J 2H ¼ κ ðX c ; X c Þ,
Fig. 14, both CSDL-SRC and CSDL-CRC efficiently extend the origi-
with kernel function κ ðX ci ; X cj Þ ¼ ϕðX ci ÞT ϕðX cj Þ. The dictionary
nal face subspace with more choices of bases. A larger subspace
becomes a set of K arbitrary functions in that Hilbert space. Using
would reduce the residual error more adaptively when recon-
the “kernel trick”, we obtain Eqn. (22) for class specific dictionary
structing the samples from the same class.
learning for sparse representation as
  c        
5.7.3. Generalization to kernel spaces ϕ X  ϕ X c W c Sc 2 þ 2αSc 1 ¼ tr κ X c ; X c
H
    n o
Another important property of our proposed class specific T  
 2 tr κ X c ; X c W c Sc þ tr Sc W c κ X c ; X c W c Sc
T
dictionary learning is that it is easy to be kernelized due to the
separation of original data. Suppose that there exists a feature X
Nc 
þ 2α Sc k1 : ð22Þ
mapping function ϕ : RD -Rt . It maps the original feature space to •n
n¼1
a high dimensional kernel space: X ¼ ½X 1 ; X 2 ; …; X N -ϕðXÞ
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210 207

Fig. 11. Examples of the AR dataset.

Fig. 12. Influence of α for SRC & CSDL-SRC and β for CRC & CSDL-CRC on the AR dataset, respectively.

Table 5 Similarly, we obtain Eqn. (23) for class specific dictionary

Recognition rate on the AR dataset (%). learning for collaborative representation as
  c        
Methods Features ϕ X  ϕ X c W c Sc 2 þ 2αSc 2 ¼ tr κ X c ; X c
H H

    n o  
ℓ2 norm Power norm T  
 2 tr κ X c ; X c W c Sc þ tr Sc W c κ X c ; X c W c Sc þ βSc 2H :
T

NN 36.89 71.34 42.23 7 1.55

ð23Þ
SVM 88.52 70.85 88.357 0.98
CRC 94.92 70.43 95.157 0.44
Therefore, we can now search an optimal dictionary directly in
SRC 92.487 0.49 90.82 7 0.71
the kernel space through optimizing W c instead of Bc (the latter is
CSDL-CRC 96.057 0.42 96.217 0.32
an extremely hard optimization problem). Since Eqns. (22) and
CSDL-SRC 95.54 70.46 96.287 0.42
(23) only depend on the kernel function κ ðX c ; X c Þ ¼ ϕðX c ÞT ϕðX c Þ,
which can be pre-computed before sparse representation and
collaborative representation, we can now handle arbitrary kernels
with tractable computation.

5.7.4. Comparison with state-of-the-art dictionary learning based

Fig. 13. Influence of the size of bases K per class on the Extended YaleB dataset.
face recognition methods
We also compare the proposed CSDL-SRC and CSDL-CRC
methods with state-of-the-art dictionary learning based face
recognition methods, including joint dictionary learning (JDL) [34],
Fisher discrimination dictionary learning (FDDL) [29], dis-
criminative KSVD (DKSVD) [32], label consistent KSVD (LCKSVD)
[10], and dictionary learning with structure incoherence (DLSI)
[18]. Table 6 shows their comparisons for face recognition on the
AR dataset. For completeness we also include the results of SRC,
CRC, NN, and SVM.
208 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210

Fig. 14. Comparison of the original faces and the learned dictionary from the Extended YaleB dataset with ℓ2 normalized features.

Table 6 the 7 images with illumination and expression changes from

Recognition rate of competing methods on the AR Session 1 were used for training, and the other 7 images with the
dataset (%).
same condition from Session 2 were used for testing. The size of
Methods the face images is 60  43. In Table 6, α is set to 0.01 for CSDL-SRC
(power) and 0.001 for CSDL-SRC(ℓ2). β is set to 0.01 for both CSDL-
NN 71.3
SVM 85.4
SRC(power) and CSDL-SRC(ℓ2). From Table 6, our proposed CSDL
CRC 90.4 algorithm achieves comparable recognition rates to the state-of-
SRC 89.8 the-art dictionary learning based face recognition methods with
DKSVD 85.4 less complexity.
LCKSVD 89.7
DLSI 89.8
JDL 91.7
FDDL 92.0
6. Conclusion
CSDL-CRC(ℓ2) 91.3
CSDL-SRC(ℓ2) 91.1
CSDL-CRC(power) 92.1 In this paper, motivated by the fact that sparse representation
CSDL-SRC(power) 92.9 and collaborative representation are two powerful tools in face
recognition tasks and that different training samples contribute
unequally to the class specific dictionary, class specific dictionary
For a fair comparison, we follow the experimental setup in learning for both sparse representation and collaborative repre-
Yang et al. [29], whose training set and testing set are publicly sentation are proposed. A dual form of dictionary learning is thus
available.1 The default number of bases is set as the number of proposed to explicitly interpret the relationship between the
training samples. The eigenface feature of dimension 300 is used. training samples and the basis vectors. Based on these, class
The cropped AR dataset consists of over 4,000 frontal images from specific dictionary learning is optimized by alternating blockwise
126 individuals. For each individual, 26 pictures were taken in two coordinate descent and Lagrange multipliers. Experimental results
separated sessions. We chose a subset consisting of 50 male sub-
show that our proposed CSDL algorithms, i.e., CSDL-SRC and
jects and 50 female subjects in the experiment. For each subject,
CSDL-CRC, have achieved superior performance for face
recognition tasks.
1
http://www4.comp.polyu.edu.hk/  cslzhang/code/FDDL_IJCV.zip
 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210
Acknowledgment
This paper is supported partly by the National Natural Science
Foundation of China (Grant nos. 61402535, 61271407), the Natural
Science Foundation for Youths of Shandong Province, China (Grant
no. ZR2014FQ001) , Qingdao Science and Technology Project (no.
14-2-4-111-jch), and the Fundamental Research Funds for the
Central Universities, China University of Petroleum (East China)
(Grant no. 14CX02169A).
References
[1] D. Tao, L. Jin, W. Liu, X. Li, Hessian regularized support vector machines for
mobile image annotation on the cloud, IEEE Trans. on Multimedia 15 (4)
(2013) 833–844.
[2] P.N. Belhumeur, J.P. Hespanha, D. Kriegman, Eigenfaces vs. fisherfaces:
recognition using class specific linear projection, IEEE Trans. Pattern Anal.
Mach. Intell. 19 (7) (1997) 711–720.
[3] D.P. Bertsekas, Nonlinear Programming, Athena Scientific, Belmont, MA, 1999.
[4] A. Castrodad, G. Sapiro, Sparse modeling of human actions from motion
imagery, Int. J. Comput. Vis. 100 (1) (2012) 1–15.
[5] C. Chang, C. Lin, Libsvm: a library for support vector machines, ACM Trans.
Intell. Syst. Technol. 2 (3) (2011) 27.
[6] S. Gao, I.W.-H. Tsang, Y. Ma, Learning category-specific dictionary and shared
dictionary for fine-grained image categorization, IEEE Trans. Image Process. 23
(2) (2014) 623–634.
[7] A. Georghiades, P. Belhumeur, D. Kriegman, From few to many: illumination
cone models for face recognition under variable lighting and pose, IEEE Trans.
Pattern Anal. Mach. Intell. 23 (6) (2001) 643–660.
[8] X. He, S. Yan, Y. Hu, P. Niyogi, H.-J. Zhang, Face recognition using laplacianfaces,
IEEE Trans. Pattern Anal. Mach. Intell. 27 (3) (2005) 328–340.
[9] J. Ho, M. Yang, J. Lim, K. Lee, D. Kriegman, Clustering appearances of objects
under varying illumination conditions, in: Proceedings of 16th International
Conference on Computer Vision and Pattern Recognition, IEEE, Madison,
Wisconsin, 2003, pp. 1–11.
[10] Z. Jiang, Z. Lin, L.S. Davis, Label consistent k-svd: learning a discriminative
dictionary for recognition, IEEE Trans. Pattern Anal. Mach. Intell. 35 (11) (2013)
2651–2664.
[11] H. Lee, A. Battle, R. Raina, A. Ng, Efficient sparse coding algorithms, in: Pro-
ceedings of Advances in Neural Information Processing Systems, The MIT
Press, Vancouver, British Columbia, 2006, pp. 801–808.
[12] B.-D. Liu, B. Shen, X. Li, Locality sensitive dictionary learning for image clas-
sification, in: Proceedings of the 22nd International Conference on Image
Processing, IEEE, Quebec, 2015, pp. 3807-3811..
[13] B.-D. Liu, B. Shen, Y.-X. Wang, Class specific dictionary learning for face
recognition, in: Proceedings of the International Conference on Security, Pat-
tern Analysis, and Cybernetics, IEEE, Wuhan, Hubei, 2014, pp. 229–234.
[14] B.-D. Liu, Y.-X. Wang, S. Bin, Y.-J. Zhang, Y.-J. Wang, Blockwise coordinate
descent schemes for sparse representation, in: Proceedings of the 39th
International Conference on Acoustics, Speech and Signal Processing, IEEE,
Florence, 2014, pp. 5304–5308.
[15] B.-D. Liu, Y.-X. Wang, Y.-J. Zhang, Y. Zheng, Discriminant sparse coding for
image classification, in: Proceedings of the 37th International Conference on
Acoustics, Speech and Signal Processing, IEEE, Kyoto, 2012, pp. 2193–2196.
[16] J. Mairal, J. Ponce, G. Sapiro, A. Zisserman, F.R. Bach, Supervised dictionary
learning, in: Proceedings of Advances in Neural Information Processing Sys-
tems, The MIT Press, Vancouver, British Columbia, 2009, pp. 1033–1040.
[17] A. Martinez, The AR Face Database, CVC Technical Report 24, 1998.
[18] I. Ramirez, P. Sprechmann, G. Sapiro, Classification and clustering via dic-
tionary learning with structured incoherence and shared features, in: Pro-
ceedings of the 23rd International Conference on Computer Vision and Pattern
Recognition, IEEE, San Francisco, California, 2010, pp. 3501–3508.
[19] O.D. Richard, E.H. Peter, G.S. David, Pattern classification. A Wiley-Interscience,
2001, pp. 373–378.
[20] F.S. Samaria, A.C. Harter, Parameterisation of a stochastic model for human
face identification, in: Proceedings of the 2nd Workshop on Applications of
Computer Vision, IEEE, Sarasota, Florida, 1994, pp. 138-142.
[21] T. Sim, S. Baker, M. Bsat, The cmu pose, illumination, and expression (pie)
database, in: Proceedings of the fifth International Conference on Automatic
Face and Gesture Recognition, IEEE, Shanghai, 2002, pp. 46–51.
[22] P. Sprechmann, G. Sapiro, Dictionary learning and sparse coding for unsu-
pervised clustering, in: Proceedings of the 35th International Conference on
Acoustics, Speech and Signal Processing, IEEE, Dallas, Texas, 2010, pp. 2042–
2045.
[23] D. Tao, X. Li, X. Wu, S.J. Maybank, Geometric mean for subspace selection, IEEE
Trans. Pattern Anal. Mach. Intell. 31 (2) (2009) 260–274.
[24] P. Tseng, Convergence of a block coordinate descent method for non-
differentiable minimization, J. Optim. Theory Appl. 109 (3) (2001) 475–494.
[25] M. Turk, A. Pentland, Eigenfaces for recognition, J. Cogn. Neurosci. 3 (1) (1991)
71–86.
209
[26] H. Wang, C. Yuan, W. Hu, C. Sun, Supervised class-specific dictionary learning
for sparse modeling in action recognition, Pattern Recognit. 45 (11) (2012)
3902–3911.
[27] J. Wright, A.Y. Yang, A. Ganesh, S. Sastry, Y. Ma, Robust face recognition via
sparse representation, IEEE Trans. Pattern Anal. Mach. Intell. 31 (2) (2009)
210–227.
[28] Y. Yan, H. Wang, D. Suter, Multi-subregion based correlation filter bank for
robust face recognition, Pattern Recognit. 47 (11) (2014) 3487–3501.
[29] M. Yang, L. Zhang, X. Feng, D. Zhang, Sparse representation based fisher dis-
crimination dictionary learning for image classification, Int. J. Comput. Vis. 109
(3) (2014) 209–232.
[30] M. Yang, L. Zhang, J. Yang D. Zhang, Metaface learning for sparse representa-
tion based face recognition, in: Proceedings of the 17th International Con-
ference on Image Processing, IEEE, Hong Kong, 2010, pp. 1601–1604.
[31] D. Zhang, M. Yang, X. Feng, Sparse representation or collaborative repre-
sentation: which helps face recognition? in: Proceedings of the 13th Inter-
national Conference on Computer Vision, IEEE, Barcelona, 2011, pp. 471–478.
[32] Q. Zhang, B. Li, Discriminative k-svd for dictionary learning in face recognition,
in: Proceedings of the 23rd International Conference on Computer Vision and
Pattern Recognition, IEEE, San Francisco, California, 2010, pp. 2691–2698.
[33] W. Zhao, R. Chellappa, P.J. Phillips, A. Rosenfeld, Face recognition: a literature
survey, ACM Comput. Surv. 35 (4) (2003) 399–458.
[34] N. Zhou, Y. Shen, J. Peng, J. Fan, Learning inter-related visual dictionary for
object recognition, in: Proceedings of the 25th International Conference on
Computer Vision and Pattern Recognition, IEEE, Province, Rhode Island, 2012,
pp. 3490–3497.
[35] W. Liu, D. Tao, Multiview hessian regularization for image annotation, IEEE
Trans. Image Process 22 (7) (2013) 2676–2687.
[36] W. Liu, D. Tao, J. Cheng, Y. Tang, Multiview hessian discriminative sparse coding
for image annotation, Comput. Vis. Image Underst 118 (1) (2014) 50–60.
[37] W. Liu, Z. Zha, Y. Wang, K. Lu, D. Tao, p-Laplacian regularized sparse coding for
human activityr recognition, IEEE Trans. Ind. Electron. (2016).
[38] J. Yu, D. Tao, M. Wang, Adaptive hypergraph learning and its application in
image classification, IEEE Trans. Image Process 21 (7) (2012) 3262–3272.
[39] Y.-X. Wang, Y.-J. Zhang, Nonnegative matrix factorization: a comprehensive
review, IEEE Trans. Know. Data Eng 25 (6) (2013) 1336–1353.
[40] Y.-X. Wang, L.-Y. Gui, Y.-J. Zhang, Neighborhood preserving non-negative
tensor factorization for image representation, in: Proceedings of the 37th IEEE
International Conference on Acoustics, Speech and Signal Processing, Kyoto,
2012, pp. 3389–3392.
[41] Y.-X. Wang, Y.-J. Zhang, Image inpainting via weighted sparse non-negative
matrix factorization, in: Proceedings of the 18th IEEE International Conference
on Image Processing, Brussels, 2011, pp. 3409–3412.
[42] Y.-X. Wang, M. Hebert, Model recommendation: Generating object detectors
from few samples, in: Proceedings of the 28th IEEE International Conference
on Computer Vision and Pattern Recognition, Boston, Massachusetts, 2015, pp.
1619–1628.
[43] Y.-X. Wang, M. Hebert, Learned by transferring from unsupervised universal
sources, in: Proceedings of the 30th AAAI Conference on Artificial Intelligence,
Phoenix, Arizona, 2016.
[44] C. Xu, D. Tao, C. Xu, Multi-view intact space learning, IEEE Trans. Pattern Anal.
Mach. Intell 37 (12) (2015) 2531–2544.
[45] C. Xu, D. Tao, C. Xu, Large-Margin Multi-view information bottleneck, IEEE
Trans. Pattern Anal. Mach. Intell 36 (8) (2014) 1559–1572.
Bao-Di Liu received the Ph.D. degree in Electronic
Engineering from Tsinghua University. Currently, he is
an assistant professor in College of Information and
Control Engineering, China University of Petroleum,
China. His research interests include computer vision
and machine learning.
Bin Shen is a Ph.D candidate in Department of Com-
puter Science, Purdue University, West Lafayette, IN
47907, USA. Before joining Purdue, he got B.S. and M.S.
degrees from EE, Tsinghua University, Beijing, in 2007
and 2009, respectively. His research interests include
image processing, machine learning and data mining.
210 B.-D. Liu et al. / Neurocomputing 204 (2016) 198–210
Yu-Xiong Wang is a Ph.D. student in the Robotics
Institute, School of Computer Science, at Carnegie
Mellon University. His research interests include
computer vision, image processing, and machine
learning.
Xue Li received the B.S. degree in Electronic Engineer-
ing from Beijing Institute of Technology (BIT), Beijing,
China, in 2011. Currently, she is a Ph.D candidate in the
Department of Electronic Engineering at Tsinghua
University, Beijing, China. Her research interests
include image classification, automatic image annota-
tion and machine learning.
Liangke Gui received his B.S. degree in Electronic
Information Engineering from Shandong University,
Jinan, China, in 2012. He is currently a M.S. student at
the Language Technology Institute, School of Computer
Science, Carnegie Mellon University, Pittsburgh, PA
15213, USA. His research interests include computer
vision, image processing and multimodal machine
learning.
Fei Yan, received the B.S. degree in Electronic and
Information Engineering, China University of Petro-
leum. Currently, he is a staff in Lijin County Party
Committee Office, China. His research interests include
computer vision and machine learning.
Yan-Jiang Wang received the M.S. degree from Beijing
University of Aeronautics and Astronautics, Beijing,
China, in 1989 and the Ph.D. degree from Beijing Jiao-
tong University, Beijing, China, in 2001. Now he is a
professor of the College of Information and Control
Engineering, China University of Petroleum, Qingdao,
China. He is also the head of the Institute of Signal and
Information Processing, China University of Petroleum.
His research interests include pattern recognition,
computer vision, and cognitive computation.