Robust Face Recognition based on Saliency Maps of Sigma Sets

We propose a robust unsupervised method for face recognition wherein saliency maps of second order statistics are employed as image descriptors. In particular, we leverage upon region covariance matrices (RCM) and their enhancement based on sigma sets for constructing saliency maps of face images. Sigma sets are of low dimension, robust to rotation and illumination changes and are efficient in distance evaluation. Further, they provide a natural way to combine multiple features and hence facilitate a simple mechanism for building otherwise tedious saliency maps. Using saliency maps thus constructed as the face descriptors brings in an additional advantage of emphasizing the most discriminative regions of a face and thereby improve recognition performance. We demonstrate the effectiveness of the proposed method for face photo-sketch recognition, wherein we achieve performance comparable to state-of-the-art without having to do sketch synthesis.