Dynamic scene deblurring

Abstract

Blurring artifacts are the most common flaws in photographs. To remove these artifacts, many deblurring methods which restore sharp images from blurry ones have been studied considerably in the field of computational photography. However, state-of-the-art deblurring methods are based on a strong assumption that the captured scenes are static, and thus a great many things still remain to be done. In particular, these conventional methods fail to deblur blurry images captured in dynamic environments which have spatially varying blurs caused by various sources such as camera shake including out-of-plane motion, moving objects, depth variation, and so on. Therefore, the deblurring problem becomes more difficult and deeply challenging for dynamic scenes. Therefore, in this dissertation, addressing the deblurring problem of general dynamic scenes is a goal, and new solutions are introduced, that remove spatially varying blurs in dynamic scenes unlike conventional methods built on the assumption that the captured scenes are static. Three kinds of dynamic scene deblurring methods are proposed to achieve this goal, and they are based on: (1) segmentation, (2) sharp exemplar, (3) kernel-parametrization. The proposed approaches are introduced from segment-wise to pixel-wise approaches, and pixel-wise varying general blurs are handled in the end. First, the segmentation-based deblurring method estimates the latent image, multiple different kernels, and associated segments jointly. With the aid of the joint approach, segmentation-based method could achieve accurate blur kernel within a segment, remove segment-wise varying blurs, and reduce artifacts at the motion boundaries which are common in conventional approaches. Next, an \textit{exemplar}-based deblurring method is proposed, which utilizes a sharp exemplar to estimate highly accurate blur kernel and overcomes the limitations of the segmentation-based method that cannot handle small or texture-less segments. Lastly, the deblurring method using kernel-parametrization approximates the locally varying kernel as linear using motion flows. Thus the proposed method based on kernel-parametrization is generally applicable to remove pixel-wise varying blurs, and estimates the latent image and motion flow at the same time. With the proposed methods, significantly improved deblurring qualities are achieved, and intensive experimental evaluations demonstrate the superiority of the proposed methods in dynamic scene deblurring, in which state-of-the-art methods fail to deblur.