Efficient Misalignment-Robust Representation for Real-Time Face Recognition

Sparse representation techniques for robust face recognition have been widely studied in the past several years. Recently face recognition with simultaneous misalignment, occlusion and other variations has achieved interesting results via robust alignment by sparse representation (RASR). In RASR, the best alignment of a testing sample is sought subject by subject in the database. However, such an exhaustive search strategy can make the time complexity of RASR prohibitive in large-scale face databases. In this paper, we propose a novel scheme, namely misalignment robust representation (MRR), by representing the misaligned testing sample in the transformed face space spanned by all subjects. The MRR seeks the best alignment via a two-step optimization with a coarse-to-fine search strategy, which needs only two deformation-recovery operations. Extensive experiments on representative face databases show that MRR has almost the same accuracy as RASR in various face recognition and verification tasks but it runs tens to hundreds of times faster than RASR. The running time of MRR is less than 1 second in the large-scale Multi-PIE face database, demonstrating its great potential for real-time face recognition.