Robust Video Super-resolution Using Low-rank Matrix Completion

In this paper, we propose a robust super-resolution method using low rank matrix completion for videos with local motions and local deformations. It is based on the multi-frame low rank matrix completion super-resolution (MCSR) framework proposed by Chen. Nonlocal multi-scale similar patches are extracted in registration instead of optical flow for complex motions. By rearranging patches extracted from low resolution frames, super-resolution problem is converted to matrix completion. Low resolution patches is represented as observed entries in a low-rank matrix. We adopt alternating direction method of multipliers (ADMM) to minimize nuclear norm and introduce a weighted fusion method to acquire final high resolution patches. Experimental results showed that the proposed method outperformed MCSR on videos with complex motions.