CUHK electronic theses & dissertations collection

Lastly, in the context of motion deblurring, we discuss a few new motion deblurring problems that are significant to blur kernel estimation and nonblind deconvolution. We found that strong edges do not always profit kernel estimation, but instead under certain circumstance degrade it. This finding leads to a new metric to measure the usefulness of image edges in motion deblurring and a gradient selection process to mitigate their possible adverse effect. It makes possible to solve for very large blur PSFs which easily fail existing blind deblurring methods. We also propose an efficient and high-quality kernel estimation method based on the spatial prior and the iterative support detection (ISD) kernel refinement, which avoids hard threshold of the kernel elements to enforce sparsity. We employ the TV-ℓ1 deconvolution model, solved with a new variable substitution scheme to robustly suppress noise.This thesis covers complete discussion of motion estimation and deblurring and presents new methods to tackle them. In the context of motion estimation, we study the problem of estimating 2D apparent motion from two or more input images, referred to as optical flow estimation. We discuss several specific fundamental problems in existing optical flow estimation frameworks, including 1) estimating flow vectors for textureless and occluded regions, which was regarded as infeasible and with large ambiguities, and 2) the incapability of the commonly employed coarse-to-fine multi-scale scheme to preserve motion structures in several challenging circumstances.To address the problem of multi-scale estimation, we extend the coarse-to-fine scheme by complementing the initialization at each scale with sparse feature matching, based on the observation that fine motion structures, especially those with significant and abrupt displacement transition, cannot always be correctly reconstructed owing to an incorrect initialization. We also introduce the adaption of the objective function and development of a new optimization procedure, which constitute a unified system for both large- and small-displacement optical flow estimation. The effectiveness of our method is borne out by extensive experiments on small-displacement benchmark dataset as well as the challenging large-displacement optical flow data.To further increase the sub-pixel accuracy, we study how resolution changes affect the flow estimates. We show that by simple upsampling, we can effectively reduce errors for sub-pixel correspondence. In addition, we identify the regularization bias problem and explore its relationship to the image resolution. We propose a general fine-to-coarse framework to compute sub-pixel color matching for different computer vision problems. Various experiments were performed on motion estimation and stereo matching data. We are able to reduce errors by up to 30%, which would otherwise be very difficult to achieve through other conventional optimization methods.We propose novel methods to solve these problems. Firstly, we introduce a segmentation based variational model to regularize flow estimates for textureless and occluded regions. Parametric and Non-parametric optical flow models are combined, using a confidence map to measure the rigidity of the moving regions. The resulted flow field is with high quality even at motion discontinuity and textureless regions and is very useful for applications such as video editing.Xu, Li.Adviser: Jiaya Jia.Source: Dissertation Abstracts International, Volume: 73-03, Section: B, page: .Thesis (Ph.D.)--Chinese University of Hong Kong, 2010.Includes bibliographical references (leaves 126-137).Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web.Electronic reproduction. [Ann Arbor, MI] : ProQuest Information and Learning, [201-] System requirements: Adobe Acrobat Reader. Available via World Wide Web.Abstract also in Chinese