Robust discriminative principal component analysis for face recognition

Abstract

Robust face recognition is a challenging goal because of the gross similarity in shape and configuration of all human faces accompanied by the large differences between face images of the same person due to variations in lighting conditions, view points, head pose, and facial expressions. These two factors remarkably degrade the distinguishability of human faces because research on face recognition has shown that differences between images of the same face due to nuisance variations are normally greater than those bet ween different faces. This problem is further exacerbated when only one image is available per person for registration and matching. An ideal face recognition system should recognize novel images of a known face and be maximally insensitive to nuisance variations while knowing very little about the image acquisition process. There are two classical methods for face recognition with variation in lighting conditions. One approach is to represent images with features that are insensitive to illumination variations. The other approach is to construct a linear subspace for images of every face under different illumination. Although both of these techniques have been successfully applied to some extent in face recognition, it is hard to extend them for recognition with various facial expressions. Three main approaches are used for expression invariant face recognition: the first approach is to morph images to be the same shape as those used for training: the second is to apply optical flow for comparing images: and the third approach is to assign different weights to local regions that are less sensitive to expression variation. Note that similar approaches cannot be applied to solve the problem of face recognition under variable lighting conditions. An even more difficult task is to recognize face images with both illumination and expression variations due to the fact that features insensitive to illumination changes are often highly sensitive to expression variation. The problem is magnified in applications when only one sample image per face (class) is available. In this paper, we develop an algorithm - Discriminative Principal Component Analysis (DPC A) that can simultaneously deal with variations in face images due to illumination and facial expressions using only a single image per person (class). This approach takes advantages from both Principal Component Analysis (PC A) and Fisher Linear Discriminant (FLD) methods. We first apply PCA to construct a subspace for image representation. We then warp the subspace by whitening and eigen-filtering according to the within-class covariance and between-class covariance of samples to improve the class separability. The proposed technique performs well under changes in lighting conditions because features extracted by PCA are similarly altered for all classes……………