A study of appearance-based feature extraction for face recognition

Abstract

Compared with the other biometrics techniques, face recognition is non-intrusive, natural, and easy to use. In a face recognition system, the feature extraction procedure aims to improve the recognition accuracy and robustness. The most popular feature extraction methods are appearance-based methods, which regard the face images as points in the image space and learn the feature extraction scheme based on the relationship between these points. The work in this thesis is in three parts. The first part proposes fast kernel Fisher discriminant analysis (FKFDA) to accelerate the nonlinear feature extraction. The second part proposes a method to extract pose-invariant feature. The third part extracts features for face verification. Fast kernel Fisher discriminant analysis (FKFDA): Kernel Fisher discriminant analysis (KFDA) extracts a nonlinear feature from a sample by calculating as many kernel functions as the training samples. Thus, its computational efficiency is inversely proportional to the size of the training sample set. In section 3, we propose FKFDA for fast feature extraction. This FKFDA consists of two procedures. First, we select a portion of training samples based on two criteria produced by approximating the kernel principal component analysis (AKPCA). Then, referring to the selected training samples as nodes, we formulate FKFDA to improve the efficiency. In FKFDA, the discriminant vectors are expressed as linear combinations of nodes in the kernel feature space, and the extraction of a feature from a sample only requires calculating as many kernel functions as the nodes. Therefore, the proposed FKFDA has a much faster feature extraction procedure compared with the naive kernel-based methods. Experimental results suggest that the proposed FKFDA can generate well classified features. Pose-invariant feature extraction: Recognizing face images across pose is one of the challenging tasks for reliable face recognition. Section 4 presents a new method to tackle this challenge based on orthogonal discriminant vector (ODV). The result of our theoretical analysis shows that an individual's probe image captured with a new pose can be represented by a linear combination of his/her gallery images. Based on this observation, in contrast to the conventional methods which model face images of different individuals on a single manifold, we propose to model face images of different individuals on different linear manifolds. The contribution of our approach includes: 1) to prove that the orthogonality to ODVs is a pose-invariant feature.; 2) to categorize each person with a set of ODVs, where his/her face images posses zero projections while other persons' images are characterized by maximum projections; 3) to define a metric to measure the distance between a face image and an ODV, and classify the face images based on this metric. Our experimental results validate the feasibility of modelling the face images of different individuals on different linear manifolds. Feature extraction for verification: In face verification, while the positive samples are the images of one person, the negative samples can be anything else. These two classes can be quite different in both size and distribution. This imbalance degrades the performance of many feature extraction methods and classifiers. Section 5 proposes a method for extracting minimum positive and maximum negative features (in terms of absolute value) for face verification. We develop two models to yield the feature extractors. Model 1 first generates a set of candidate extractors that can minimize the positive features, and then chooses the ones among these candidates that can maximize the negative features. Model 2 first generates a set of candidate extractors that can maximize the negative features, and then chooses the ones that can minimize the positive features. Compared with the traditional feature extraction methods and classifiers, the proposed models are less likely affected by the imbalance.