Entire Gabor kernel locality preserving Fisher discriminant analysis: Subspace approach for expression recognition

The important objective of this work is to utilization of entire Gabor features by enhancing the phase part of the Gabor and maximizing the Fishers ratio in nonlinear domain space by preserving the local information. Entire Gabor kernel locality preserving Fisher discriminant analysis (EGKLPFDA) approach is proposed. Both Gabor magnitude and spatially enhanced phase congruency parts are separately used for feature extraction. These two vector feature space is projected into KLPFDA subspace method by preserving the kernel discriminant locality structure of data. Projected subspace is normalized by Z-score normalization. Both normalized scores are fused by maximum fusion rule. Final score obtained from train and test image sets are used to distance matching using Euclidean distance algorithm and support vector machine (SVM) classifier is implemented to classify the expressions. Performance analysis is carried out by comparing earlier approaches. Experimental results on JAFFE, Yale, and FD database demonstrate the effectiveness of the proposed approach.

I. Introduction

Facial expression exhibits the feelings or opinions based on various situations and emotions. Expression recognition is an important task to measure the human intension and how human activity is react with environment. Appearance based approaches finds major role for expression recognition in various critical conditions like illumination variations and partial occlusions. In pattern recognition facial expression recognition is an interesting task [1]–[4]. Different approaches for recognizing human facial expressions using appearance and geometrical features are proposed by several researchers. In this research work appearance based holistic approach is focused using suitable entire Gabor filter. To synthesize a complete image face under appearance based approach, both shape and texture are important features for expression identification and classification. Wang Z., Ruan Q. [5] introduced and demonstrated local Fisher discrminant analysis based on orthogonal analysis of subspace for facial expression recognition. M. Loog, R. P. W. Duin [6] shown that how weighted pair wise Fisher criteria can be implemet on multiclass linear dimension reduction. E. K. Tang, P. N. Suganthan [7] focused on dimensional reduction in linear region using relevance weighted linear discriminant analysis (RWLDA). Singularity matrix problem in LDA is resolved by P. Belhumeur et al [8] by proposing a Fisherface method in 1997, which makes a principal component analysis based projection and a change of matrix size so that the matrix becomes nonsingular. LDA is a supervised subspace method, using class labels it discriminates the different expression classes and tries to find the subspace. It is a old and well known dimensional reduction method among discriminant family of subspace manifolds.