Features and statistical classifiers for face image analysis

Song, Qing (2001) Features and statistical classifiers for face image analysis. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (8MB)

Abstract

This thesis presents the systematic analysis of feature spaces and classification schemes for face image processing. Linear discriminants, probabilistic classifiers, and nearest neighbour classifiers are applied to face/nonface classification in various feature spaces including original grayscale space, face-image-whitened space, anything-image-whitened space, and double-whitened space. According to the classification error rates, the probabilistic classifiers performed the best, followed by nearest neighbour classifiers, and then the linear discriminant classifier. However, the former two kinds of classifiers are more computationally demanding. No matter what kind of classifier is used, the whitened space with reduced dimensionality improves classification performance. -- A new feature extraction technique, named dominant feature extraction, is invented and applied to face/nonface classification with encouraging results. This technique extracts the features corresponding to the mean-difference and variance-difference of two classes. Other classification schemes, including the repeated Fisher's Linear Discriminant (FLD) and a moving-centre scheme, are newly proposed and tested. The Maximum Likelihood (ML) classifier based on hyperellipsoid distribution is applied for the first time to face/nonface classification. -- Face images are conventionally represented by grayscales. This work presents a new representation that includes motion vectors, obtained through optical flow analysis between an input image and a neutral template, and a deformation residue that is the difference between the deformed input image and the template. The face images compose a convex cluster in this representation space. The viability of this space is tested and demonstrated through classification experiments on face detection, expression analysis, pose estimation, and face recognition. When the FLD is applied to face/nonface classification and smiling/nonsmiling face classification, the new representation of face images outperforms the traditional grayscale representation. Face recognition experiments using the nearest neighbour classifier on the Olivetti and Oracle Research Laboratory (ORL) face database shows that the deformation residue representation is superior to all other representations. These promising results demonstrate that as a general-purpose space, the derived representation space is suitable for almost all aspects of face image processing.

Actions (login required)

View Item