Multiscale overlapping blocks binarized statistical image features descriptor with flip-free distance for face verification in the wild

In this work, an effective face verification system based on a fusion of the multiscale overlapping blocks binarized statistical image features (BSIF) descriptor and a flip-free distance is proposed. First, we propose a BSIF with overlapping blocks descriptor and extend it to a multiscale framework. Then, after applying dimensionality reduction, the projected vectors for each scale are scored using two prevalent face verification classifiers: triangular similarity metric learning and the Joint Bayesian method. Moreover, a flip-free distance is applied to boost overall performance. Finally, the different scores for different scales are fused using a support vector machine to further improve performance. We evaluate the proposed face verification system under restricted and unrestricted protocols, for which, in both cases, we achieve very competitive results (90.05 and 93.41%) for the problem of face verification on the Labeled Faces in the Wild dataset.