Multiscale binarised statistical image features for symmetric unconstrained face matching

Face recognition subject to uncontrolled imaging conditions still remains a challenge. This paper proposes a number of counteracts to partly neutralize the adverse effects of the unwanted factors on performance. First, a novel multi-scale image descriptor (MBSIF) is proposed which unlike most commonly used features for face representation employs statistics of natural images to improve its representation capacity. Second, in order to minimize the sensitivity of the recognition system to misalignment, the descriptor is computed regionally on top of a dense MRF image matching model. Similarities of the component-wise descriptors between a pair of images are then measured in an LDA space taking into account the established dense correspondences. Third, an asymmetric face MRF matching process is extended into a symmetric framework where the similarity between a pair of images is measured in two directions, improving recognition accuracy. Finally, the proposed MBSIF descriptor is jointly used with MLBP and MLPQ representations to further enhance the accuracy. The proposed approach has been evaluated in real world challenging scenarios and shown to perform very favorably compared to the state-of-the-art methods.