Locality-Constrained Iterative Matrix Regression for Robust Face Hallucination

摘要: The performance of traditional face recognition approaches is sharply reduced when encountered with a low-resolution (LR) probe image. basic idea super-resolution (SR) to desire high-resolution (HR) image from an observed LR one the help set training examples. In this paper, we propose locality-constrained iterative matrix regression (LCIMR) model for hallucination task and use alternating direction method multipliers solve it. LCIMR attempts directly compute representation coefficients maintain essential structural information. A locality constraint also enforced preserve sparsity simultaneously. Moreover, iteratively updates similarities reconstruction weights based on result (the hallucinated HR patch) previous iteration, giving rise improved performance. Experimental results benchmark FEI database show superiority proposed over some state-of-the-art algorithms.