Class specific discriminant dictionary learning with kernels for
face recognition
Bao-Di Liu
1
; College of Information and Control Engineering, China University of Petroleum; Qingdao, 266580, China
Yuting Wang; Department of Informatics, Karlsruhe Institute of Technology; Karlsruhe, 76131, Germany
Liangke Gui; School of Computer Science, Carnegie Mellon University; Pittsburgh, PA 15213, USA
Yu-Xiong Wang; School of Computer Science, Carnegie Mellon University; Pittsburgh, PA 15213, USA
Bin Shen; Department of Computer Science, Purdue University; West Lafayette, IN 47907 USA
Xue Li; Department of Electronic Engineering, Tsinghua University; Beijing 100084, China
Yan-Jiang Wang; College of Information and Control Engineering, China University of Petroleum; Qingdao, 266580, China
Abstract
The past few years have witnessed the impressive perfor-
mance of sparse representation based classification (SRC) for
visual recognition. However, the SRC technique may lead to
high residual error and poor performance due that the training
samples in each class contribute equally to the dictionary in the
corresponding class. This inspired the emergence of class spe-
cific dictionary learning algorithm. In this paper, we propose
a novel approach—class specific dictionary learning combined
with linear discriminant analysis constraints in Reproducing Ker-
nel Hilbert Space (KCSDL-LDA), which modifies and extends the
conventional class specific dictionary learning (CSDL) algorithm
in several aspects. First, we propose a novel class specific dic-
tionary learning scheme that considers the weight of each sam-
ple for each class when generating the dictionary in that class.
Second, we extend the novel class specific dictionary learning
scheme to the Reproducing Kernel Hilbert Space, in which non-
linear structure can be extracted and represented to improve the
classification accuracy. Finally, we further enhance the classifi-
cation performance by combing class specific dictionary learning
with linear discriminant analysis constraints in Reproducing Ker-
nel Hilbert Spaces. Extensive experimental results on several face
recognition benchmark datasets, such as Extended YaleB dataset,
CMU PIE dataset and AR dataset, demonstrate the superior per-
formance of our proposed KCSDL-LDA.
Introduction
The past few years have witnessed the impressive perfor-
mance of dictionary learning for sparse representation in visual
computation areas, such as image annotation [1], image inpaint-
ing [2], image classification [3], face recognition [4] and image
denoising [5]. Different from traditional decomposition frame-
works like Principal Component Analysis (PCA), Non-negative
Matrix Factorization (NMF) [6] and low-rank factorization, s-
parse representation is capable of generating sparse codes under
over-complete bases to represent the data more adaptively and
flexibly.
Face recognition, one of the successful applications of s-
parse representation, is a classical yet challenging research top-
ic in computer vision and pattern recognition [7]. Effective face
recognition usually involves two important stages: 1) feature ex-
1
thu.liubaodi@gmail.com
traction, 2) classifier construction and face prediction. For the
first stage, Turk et al. performed principal component analysis
(PCA) to extract Eigenfaces [8]. He et al. proposed Laplacian-
faces [9] to preserve local information. Belhumeur et al. extracted
Fisherfaces [10] to maximize the ratio of between-class scatter to
within-class scatter. For the latter stage, Richard et al. introduced
a nearest neighbor method [11] to predict the label of a test image
using its nearest neighbors in the training samples. Tao et al. pre-
sented a nearest subspace method [12] to assign the label of a test
image by comparing its reconstruction error for each category.
Under the nearest subspace framework, Wright et al. [4] de-
scribed a sparse representation based classification (SRC) sys-
tem and achieved an impressive performance for face recognition.
Given a test sample, the sparse representation technique repre-
sents it as a sparse linear combination of the train samples. The
predicted label is determined by the residual error from each class.
Zhang et al. [13] illustrated a collaborative representation based
classification (CRC) system. Similar to SRC, CRC represents a
test sample as the linear combination of almost all the training
samples. Moreover, Zhang et al. demonstrated that it was the
collaborative representation rather than the sparse representation
that makes the nearest subspace method powerful for classifica-
tion. Overall, both SRC and CRC algorithms directly use the
training samples as the dictionary for each class. This may lead
to high residual error and poor performance due that the training
samples in each class contribute equally to the dictionary in the
corresponding class. Therefore, the emergence of class specific
dictionary learning algorithm attracts the attention of many re-
searchers. They focus on learning a dictionary enforced by some
discriminative criteria that can reduce the residual error greatly
and achieve a superior performance for classification tasks.
So far, existing discriminative dictionary learning approach-
es are mainly categorized into three types: shared dictionary
learning, class specific dictionary learning and hybrid dictionary
learning. In shared dictionary learning, the bases are learned with
all the training samples together. The discriminative information
is often embedded into the dictionary learning procedure. Mairal
et al. learned a discriminative dictionary [14] with a linear classi-
fier of coding coefficients. Liu et al. embedded the linear dis-
criminant analysis [15] into the dictionary. Zhang et al. ob-
tained a discriminative dictionary by integrated the label infor-
mation [16] into the dictionary learning. The shared dictionary
learning approaches usually lead to a small-sized dictionary and
©2016 Society for Imaging Science and Technology
DOI: 10.2352/ISSN.2470-1173.2016.11.IMAWM-456
IS&T International Symposium on Electronic Imaging 2016
Imaging and Multimedia Analytics in a Web and Mobile World 2016 IMAWM-456.1