© Springer International Publishing Switzerland 2015
J. Yang et al. (Eds.): CCBR 2015, LNCS 9428, pp. 536–543, 2015.
DOI: 10.1007/978-3-319-25417-3_63
Parallel Nonlinear Discriminant Feature Extraction
for Face and Handwritten Digit Recognition
Qian Liu
1
, Fei Wu
1(
)
, Xiaoyuan Jing
1,2(
)
, Xiwei Dong
1
, Kun Xu
1
,
Xuejing Shi
1
, and Xiaoyu Xi
1
1
School of Automation, Nanjing University of Posts and Telecommunications, Nanjing, China
{wufei_8888,jingxy_2000}@126.com
2
State Key Laboratory of Software Engineering, School of Computer,
Wuhan University, Wuhan, China
Abstract. For recognition tasks with large amounts of data, the nonlinear dis-
criminant feature extraction technique often suffers from large computational
burden. Although some nonlinear accelerating methods have been presented,
how to greatly reduce computing time and simultaneously keep favorable rec-
ognition result is still challenging. In this paper, we introduce parallel compu-
ting into nonlinear subspace learning and build a parallel nonlinear discriminant
feature extraction framework. We firstly design a random non-overlapping
equal data division strategy to divide the whole training sample set into several
subsets and assign each computational node a subset. Then we separately learn
nonlinear discriminant subspaces from these subsets without mutual communi-
cations, and finally select the most appropriate subspace for classification. Un-
der this framework, we propose a novel nonlinear subspace learning approach,
i.e., parallel nonlinear discriminant analysis(PNDA). Experimental results on
three public face and handwritten digit image databases demonstrate the effi-
ciency and effectiveness of the proposed approach.
Keywords: Parallel nonlinear discriminant feature extraction framework ·
PNDA · Face and handwritten digit recognition
1 Introduction
Supervised subspace learning is an effective feature extraction technique for face and
handwritten digit recognition application, since it utilizes class information to extract
discriminative features. Linear discriminant analysis (LDA)[1] is a representative
supervised subspace learning method, which calculates the projective subspace by
maximizing the between-class scatter and simultaneously minimizing the within-class
scatter. To improve the performance of LDA, many methods have been addressed,
such as discriminative orthogonal neighborhood-preserving projection[2] and L
1
-norm
maximization based LDA[3].
Due to the non-linear nature of most real-world image data, many nonlinear
discriminant subspace learning methods have been developed, such as kernel