A learning-based POCS algorithm for face image super-resolution reconstruction

Super-resolution (SR) reconstruction, particularly on face images, can be widely used in forensic analysis and video surveillance. In this paper, we investigate the statistical characteristics of face images and incorporate them into SR reconstruction in terms of deterministic sets. Based on the set theoretic formulation, the projection onto convex sets (POCS) algorithm is applied to find the solution to face image reconstruction. Compared with the traditional POCS based SR methods, the proposed approach imposes additional constraints to the solution. The experimental results on frontal face images show that the proposed approach gains a better performance both on noise suppression and reconstruction quality.