Self-Organized Gabor Features for Pose Invariant Face Recognition

Pose-invariant face recognition using single frontal training image is considered one of the most difficult challenges in face recognition. To address this problem, we introduce a novel feature extraction method based on learning the manifold of local features. Changes in local features due to pose variations induce a nonlinear manifold in the feature space. Self-organizing map is employed to learn the manifold induced by Gabor filter response of a generic training face database captured at various pose directions. Furthermore, this manifold can be used to represent new face image as a set of points in the feature space. A modular Hausdorff distance measure, which can effectively measure the similarity between two point sets without any correspondence, is also proposed to identify unlabeled subjects. Experimental results on CMU-PIE face database show the effectiveness of the novel method against pose variations.