Motion Blur Estimation:Theories,Algorithms And Applications

Image deblurring is one of the most fundamental and challenging problems in computer vision and image processing communities.The factors that lead to image blur usually include camera shake,object motion,and focus/defocus.Blind image deblurring aims to recover a sharp latent image from a blurred image.The blur caused by camera shake can be considered as the convolution of a blur kernel and a clear image,where the blur kernel describes the camera trajectories of camera shake during the exposure time.This is an ill-posed problem as only the blurred image is available while the blur kernel and latent image are unknown.The recent years have witnessed significant advances in single image deblurring due to the increasing popularity of convenient digital imaging equipments and the rapid progress of the related intelligent fields.Existing deblurring methods usually focus on specific images(e.g.,nat-ural images,text images,saturated images,and face images,etc.)and have difficult in handling noise and outliers.In addition,these methods usually needs a huge computation time.However,the real world blurred images usually contain large noise and significant outliers.Therefore,how to efficiently recover clear images from real world blurred images is an urgent issue and also an important research topic in computer vision,computer graphics and image processing communities.This thesis will focus on the theories of blur process and proposes several efficient motion blur estimation algorithms.Specifically,the main contents of this thesis are as follows:(1)Robust blind image deblurring with outliers.We propose a robust motion blur esti-mation method to handle the blurred images with large noise and significant outliers.We analyze the effects of outliers on kernel estimation and present an efficient algorithm by exploiting reliable edges and removing outliers for kernel estimation.In addition,we ana-lyze the effects of outliers on image restoration and propose an efficient image restoration algorithm.Experimental results demonstrate that the proposed method is able to handle large noise and significant outliers in motion blur estimation.(2)Fast L0-regularized kernel estimation for blind image deblurring.Based on the sparse property of image gradients,we propose a fast L0-regularized gradient image prior for blur kernel estimation.The proposed method is able to avoid complex edge-selection step,which greatly simplifies the blur kernel estimation.The proposed algorithm performs favorably against state-of-the-art methods in terms of accuracy and running time.(3)Deblurring face images with exemplars and beyond.Existing methods are less effective as only few edges can be extracted from blurred face images for blur kernel estimation.We address the problem of deblurring face images by exploiting facial structures.The matched structure from exemplar face images is used to guide the kernel estimation pro-cess.In addition,the proposed method can be extend to handle other specific deblurring tasks.The proposed algorithm performs favorably against state-of-the-art methods on face images and other kinds of images.(4)Text image deblurring and beyond.We propose a simple yet effective L0-regularized prior based on intensity and gradient for text image deblurring.The proposed method does not require any complex edge-selection and character detection which are critical to the state-of-the-art text deblurring algorithms.In addition,we show that the proposed method can be effectively applied to deblur low-illumination images.The proposed algorithm performs favorably against state-of-the-art methods on text images and low-illumination images.(5)Deblurring images via dark channel prior.We theoretically analyze the properties of blur process and propose an effective blind image deblurring algorithm based on the dark channel prior and prove the sparse property of the clear images.Based on the theoreti-cal analysis,we propose a effective algorithm based on the sparsity of the dark channel.However,imposing sparsity on the dark channel introduces a non-convex non-linear opti-mization problem.We introduce a linear approximation to solve this problem.Extensive experiments demonstrate that the proposed deblurring algorithm achieves state-of-the-art results on natural images and performs favorably against methods designed for specific scenarios.(6)Object motion deblurring.We propose a novel model for object motion deblurring based on a maximum a posterior formulation in which soft-segmentation is incorporated for object layer estimation.Experimental results demonstrate that the proposed algorithm performs favorably against the state-of-the-art object motion deblurring methods on chal-lenging scenarios.