A spatially variant motion blur removal technique for single image deblurring

Abstract

This paper presents an effective single image spatially variant motion blur removal technique. Motion blur during the image capture occurs due to the relative motion between the capturing device and image being captured. This blur becomes spatially variant if it varies with position in an image. Removal of such space/shift variant blur from a single image is a challenging problem. To solve this problem, the blurred image is divided into smaller subimages assuming that each subimage is uniformly blurred. In the proposed technique each subimage is transformed into frequency domain for estimating motion blur parameters. Proposed blur parameter estimation implies dual Fourier spectrum computation and Radon transformation steps to obtain estimated values of blur length and blur angle respectively. These estimated motion blur parameters are used to prepare a local parametric blur model. These local parametric blur models are deconvolved with the blurry subimages and thus restores the original image. Restoration step is performed using parametric Wiener filtering. However, due to piecewise uniform consideration, proposed technique introduces some blocking artifacts which are later removed by applying post processing steps on the restored image. To demonstrate the usefulness of this technique for natural scenes and real blurred images it is tested on Berkeley Segmentation dataset and standard test images. The proposed technique shows effective deblurring of images under spatial variance conditions.