Occlusion-Aware Motion Deblurring for Bilayer Scenes

Abstract

In this thesis, a novel blur model that can deal with occlusion in the blurred image from a scene with depth discontinuities is proposed. Existing deblurring methods usually ignore the occlusion that occurs near the depth variations but it causes severe artifacts near the object boundary, which is a critical factor in deblurring. Based on the analysis about the blur kernel near the depth discontinuities for a two-layer image model, a new occlusion-aware blur model which can make use of the information of occluded regions is proposed. Proposed model jointly recovers the depth map, foreground mask and restored image with accurate object boundary from two blurred observations. Also, a highly accurate optimization method is provided based on MCMC. Comparative experimental results on synthetic and real blurred images demonstrate convincingly that proposed model gives satisfactory results.