Video Deblurring Using Non-uniform Motion Field Based On Texture Synthesis

Along with the popularization of video recorders,capturing videos becomes an important way to make a note about people's daily lives.However,non-uniform motion blur caused by moving objects or camera jitter is a common phenomenon in video frames captured by hand hold digital cameras,especially captureed by amateurs.State-of-the-art video deblurring methods,such as directly estimating motion blur kernels and applying deconvolution on each frame,introduce extra artifacts or do not handle diverse moving objects and involuntary camera-shake blurs.What's more,it is hard to handle non-uniform motion blur and keep both spatially and temporally coherent in deblurring results using existing video deblurring methods.In this thesis,we build a non-uniform motion model for video frames and propose a new video deblurring method in which sharp patches in a video sequence are matched to reconstruct blurry frames.We find that the homograph-based motion model can approaximate the style of motion blurs.A multi-layer homograph based video deblurring method is proposed to deal with such large related motion blur.Our method firstly divides a blurring frame into several motionbased patches using optical flow between the blurring frame and its adjacent sharp frames.Then,estimating blur kernel of each patch and blurring corresponding sharp frame.Finally,matching blurred sharp frame with blurring frame and replace these blurring regions with sharp ones.However,such homograph-based motion model cannot approximate the real style of motion blurs.To find the proper patch for a blurry one,we enrich the search space with non-uniform motion blur and use a generalized PatchMatch-based random search strategy to handle rotation,scale,and blur change in different frames.Instead of pixel-wise or regular patch-wise image representation,we use superpixel-based image representation that involves the color and motion field to gather similar pixels.Our non-uniform motion blur kernels are estimated for these irregular patches using the motion field.To further improve the effectiveness of the method,we used GPU-based acceleration using OpenGL Shanding Language(GLSL).The experimental results show that the proposed method can reconstruct a blurry frame from videos with complex moving objects and deblur background and moving objects with results that are superior to those of conventional deblurring algorithms.