Efficient misalignment-robust face recognition via locality-constrained representation

Current prevailing approaches for misaligned face recognition achieve satisfactory accuracy. However, the efficiency and scalability have not yet been well addressed, which limits their applications in practical systems. To address this problem, we propose a highly efficient algorithm for misaligned face recognition, namely misalignment-robust locality-constrained representation (MRLR). Specifically, MRLR aligns the query face by appropriately harnessing the locality constraint in representation. Since MRLR avoids the exhaustive subject-by-subject search in datasets and complex operation on large matrix, the efficiency is significantly boosted. Moreover, we take the advantage of the block structure in dictionary to accelerate the derived analytical solution, making the algorithm more scalable to the large-scale datasets. Experimental results on public datasets show that MRLR substantially improves the efficiency and scalability with even better accuracy.