Robust Face Super-Resolution via Locality-Constrained Low-Rank Representation