A two-stage approach to blind spatially-varying motion deblurring

Many blind motion deblur methods model the motion blur as a spatially invariant convolution process. However, motion blur caused by the camera movement in 3D space during shutter time often leads to spatially varying blurring effect over the image. In this paper, we proposed an efficient two-stage approach to remove spatially-varying motion blurring from a single photo. There are three main components in our approach: (i) a minimization method of estimating region-wise blur kernels by using both image information and correlations among neighboring kernels, (ii) an interpolation scheme of constructing pixel-wise blur matrix from region-wise blur kernels, and (iii) a non-blind deblurring method robust to kernel errors. The experiments showed that the proposed method outperformed the existing software based approaches on tested real images.