Region-Based Super Resolution for Video Sequences Considering Registration Error

Super-resolution (SR) for video sequences is a technique to obtain a higher resolution image by fusing multiple low-resolution (LR) frames of the same scene. In a typical super-resolution algorithm, image registration is one of the most affective steps. The difficulty of this step results in the fact that most of the existing SR algorithms can not cope with local motions because they assume global motion. In this paper, we propose a SR algorithm that takes into account inaccurate estimates of the registration parameters and the point spread function. When frames obey the assumed global motion model, these inaccurate estimates, along with the additive Gaussian noise in the low-resolution image sequence, result in different noise level for each frame. However, in case of existence of local motion and/or occlusion, regions that have local motion and/or occlusion have different noise level. To cope with this problem, we propose to adaptively weight each segment according to its reliability. The segments are generated by segmenting the reference frame using watershed segmentation. The experimental results using real video sequences show the effectiveness of the proposed algorithm compared to three state-of-the-art SR algorithms.