Robust deblurring based on prediction of informative structure

This study presents a robust motion deblurring method in which an adaptive prediction is used to extract the informative regions for kernel estimation. The prediction not only sharpens the blurry edges, but also adaptively predicts the large scale structure for kernel estimation. It allows to only use the alternating minimisation with a computationally efficient Gaussian prior for both the image and kernel while without employing thoughtful attention such as multi-scale scheme or kernel refinement. Extensive experiments were carried out to validate the proposed method and to compare it with some previous approaches. The experiment results demonstrated that the approach achieves, if not better than, state-of-the-art results for uniformly blurred images.