Letters
Two-stage dimensionality reduction approach based on 2DLDA and fuzzy
rough sets technique
Hao-Xin Zhao, Hong-Jie Xing
n
, Xi-Zhao Wang
Key Laboratory of Machine Learning and Computational Intelligence, College of Mathematics and Computer Science, Hebei University, Baoding 071002, Hebei Province, China
article info
Article history:
Received 15 April 2011
Received in revised form
10 June 2011
Accepted 22 June 2011
Communicated by D. Wang
Available online 3 August 2011
Keywords:
2DLDA
Fuzzy rough sets
Feature extraction
Dimensionality reduction
Face recognition
abstract
Traditional two-dimensional linear discriminant analysis (2DLDA) can deal with discriminant informa-
tion between classes and directly extract features from image matrices. However, 2DLDA essentially
works solely in the row-direction of images. Therefore, the features extracted by 2DLDA may contain
redundant information. In this letter, a dimensionality reduction method based on 2DLDA and fuzzy
rough sets technique is proposed to deal with the foresaid problem. Experimental results on the four
benchmark face databases demonstrate that the proposed method is superior to its related methods.
& 2011 Elsevier B.V. All rights reserved.
1. Introduction
To overcome the curse of dimensionality problem during the
procedure of face recognition, many dimensionality reduction
techniques have been developed [1,2]. Among them, linear dis-
criminant analysis (LDA) is one of the mostly used methods.
However, when the number of features is very large while the
number of samples is very small, the singularity of the within-
class scatter matrix of LDA may occur. To solve this problem,
regularized discriminant analysis (RDA) [4] and the Fisherface
method [3] have been proposed.
However, LDA and its improved versions are all image-as-
vector based methods which may destroy the structure of the
images. To fully utilize the information underlying the given
images, image-as-matrix based methods, such as 2DPCA, 2DLDA,
and 2DLPP, were proposed, respectively, in [5–7]. Among them, it
is observed that 2DLDA can reduce the computational complexity
of covariance matrix and enhance recognition rate in comparison
with LDA. Towards 2DPCA, Zhang and Zhou [11] pointed out that
it works only in the row-direction of images and there still
contains redundant information in the extracted features. To
alleviate this problem, they proposed 2-Directional 2-Dimen-
sional PCA ((2D)
2
PCA). Inspired by (2D)
2
PCA, Noushath et al. [8]
pointed out that 2DLDA has the similar problem. There may exist
redundant information within the features extracted by 2DLDA.
Recently, Zhai et al. [10] utilized 2DPCA and fuzzy rough sets
technique to remove the redundant information extracted by
2DPCA and applied it to face recognition. They demonstrated that
the proposed method can produce higher recognition rates
compared to 2DPCA on three benchmark face databases.
Motivated by the 2DPCA and fuzzy rough sets based dimen-
sionality reduction technique [10], we propose a new method
based on 2DLDA and fuzzy rough sets technique (2DLDAFRS) to
further remove the redundant features extracted by 2DLDA. There
are two stages to construct the proposed 2DLDAFRS. In the first
stage, 2DLDA is utilized to extract relevant features from the
given image matrices. In the second stage, the attribute reduction
method based on fuzzy rough sets is employed to select impor-
tant features and get rid of redundant ones from the obtained
features. Different from the algorithm in literature [10], we utilize
a heuristic algorithm in the second stage to find a single reduction
rather than all the reductions according to the frequency of the
features in the obtained discernibility matrix.
Compared to its related approaches, the proposed method has
three advantages which are listed below:
– It can remove the redundant information existing in the
features extracted by 2DLDA and uses less features to get
better performance than 2DLDA.
– Since utilizing 2DLDA in the first stage, it is regarded as a
supervised method which is different but superior to the
method based on 2DPCA and fuzzy rough sets technique [10].
– It uses a heuristic algorithm to reduce attributes and has less time-
complexity than the 2DPCA and fuzzy rough sets based method.
Contents lists available at ScienceDirect
journal ho mepage: www.elsevier.com/locate/neucom
Neurocomputing
0925-2312/$ - see front matter & 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.neucom.2011.06.020
n
Corresponding author. Tel.: þ86 15933987805.
E-mail address: hjxing@hbu.edu.cn (H.-J. Xing).
Neurocomputing 74 (2011) 3722–3727