TOWARDS FACIAL ASYMMETRY BASED FACE RECOGNITION

Abstract

Face recognition, as an active research area over the past three decades, still poses many challenges. Recognition of age-separated face images (age invariant face recognition) based on facial asymmetry is one of such challenges. Successful solutions to this recognition paradigm would allow the facial photographs to be matched against face images with temporal variations. Facial asymmetry, which refers to non-correspondence in shape, size, and arrangement of facial landmarks on both sides of the face, is an intrinsic recognition-specific facial feature used for face recognition task. The contributions of this dissertation are focused on recognition of age-separated face images using facial asymmetry. We introduce to use a feature description scheme suitable to represent facial asymmetry. The introduced feature description is adaptable to recognize age-separated face images and extract demographic information such as age group, gender, and race from a given face image. Based on the introduced feature description, this dissertation offers the following three main contributions to recognize age-separated face images. The first contribution is a matching-scores space based approach to recognize age separated face images. In the proposed framework, matching scores of holistic, local, and asymmetric facial features are combined in a matching-score space (MSS) with Support Vector Machine (SVM) as a classifier to separate genuine and imposter classes. Experimental results on three publically available benchmark facial aging databases show the efficacy of proposed approach compared to some existing state-of-the-art approaches. The second contribution is focused on the role of facial asymmetry based age group estimation in recognizing age-separated face images. We provide a hierarchical approach to perform age group estimation task. The role of various asymmetric facial regions in recognizing age-separated face images of different age groups is investigated. We integrate the knowledge learned from age group estimation into face recognition algorithm to enhance the recognition performance of age-separated face images. The viability of this approach is demonstrated on two benchmark facial aging databases. The experimental results suggest that integration of age group estimates into face recognition algorithm enhances the recognition performance of age separated face images, considerably. The third contribution is examination of the role of facial asymmetry in demographic estimation (i.e. age group, gender, and race) of a query face image in a face recognition system. The role of different asymmetric facial regions in recognizing face images with different demographic attributes is presented. We integrate the demographic estimates into a face recognition algorithm to enhance the recognition accuracy of age-separated face images. Experiments are conducted on benchmark facial aging databases to validate the performance of proposed approach. The experimental results suggest that proposed approach is more adaptable to recognize age-separated face images compared to some existing state-of-the-art methods.