Unconstrained Face Identification with Multi-scale Block-based Correlation

Many approaches to unconstrained face identification exploit small patches which are unaffected by distortions outside of their locality. However, small patches have limited discriminative ability, making accurate patch matching difficult. We propose a novel blockbased approach to exploit the greater discriminative information in larger areas, while maintaining robustness to local variations. A testing block contains several neighbouring testing patches. We identify all the matching training patches in a block jointly, using normalized cross correlation (NCC), as a means of reducing the uncertainty of each matching patch with the addition of the neighbouring patch information. We further propose a multi-scale extension in which we carry out block-based matching at several block sizes, where a larger block contains more neighbouring testing patches, to combine complementary information across scales for further robustness. For evaluation, we use two unconstrained datasets, cropped Labelled Faces in the Wild (LFWCrop) and Unconstrained Facial Images (UFI). Our new approach is able to significantly improve identification accuracy over existing patch-based methods, in the presence of uncontrolled pose, expression and lighting variations.