Discriminative Learning Models And Algorithms For Image Deblurring

With the rapid development of smart phones and other portable mobile imaging devices,the image has played an important role for people to obtain information and communicate with others.The image is produced by the imaging sensors that accumulate incoming lights for an amount of time.During exposure,if the camera sensor moves or objects move,a blurred image will be obtained.Motion blur can produce disappointing blurred images with inevitable information loss.Thus,it is challenging to restore clear images from blurred images.Image deblurring is a typical ill-posed problem.It is important to define effective image priors that can model the statistical property of clear images.In addition,the blurred images usually contain significant noise or outliers,due to the complex imaging environment.However,most existing methods assume that the blurred images only contain a small amount of noise or no noise,and cannot handle outliers in real captured images.A few outlier handling methods mainly rely on strong prior knowledge or complex outlier detection,and is computationally inefficient.Therefore,it is urgently needed to efficiently address image deblurring in complex scenes.This thesis will focus on the property of the blur process and image deblurring for complex scenes.Then this thesis proposes effective discriminative learning models and algorithms for image deblurring.Specifically,the main contents of this thesis are as follows.(1)Blur kernel estimation via salient edges and low rank prior for blind image deblur-ring.We propose a blind image deblurring method based on salient edge selection and low rank image prior.By analyzing that the blur process changes the similarity of non-local neighboring image patches,we adopt the low rank prior to restore clear images.To further improve the accuracy of the kernel estimation,we use the salient edge selection method to extract the main structures to facilitate the blur kernel estimation.Experimental results demonstrate that the proposed method can effectively handle the blind image deblurring.(2)Learning discriminative data fitting functions for blind image deblurring.We analyze that image intensity,image gradient,and higher order image information have different effects on blur kernel estimation and image restoration.Then we learn discriminative data fitting functions to respectively estimate blur kernels and restore clear images for blind image deblurring.Extensive experiments on challenging blurred images demonstrate the efectiveness of the proposed method.(3)Blind image deblurring with outlier handling.By analyzing the effects of outliers on the goodness-of-fit in function approximation,we propose a simple but effective image deblurring method based on a robust data fidelity term.The proposed method does not require complex operations,e.g.,outlier detection,but can minimize the negative effect of outliers in the blur kernel estimation process.Experimental results illustrate that the proposed method can effectively address blind image deblurring with outlier handling.(4)Learning data terms for non-blind deblurring.We propose a simple and effective dis-criminative framework to learn data terms that can adaptively handle blurred images in the presence of severe noise and outliers.Instead of learning the distribution of the data fitting errors,we directly learn the associated shrinkage function for the data term using a cas-caded architecture,which is more flexible and effective.The proposed method performs favorably against the state-of-the-art algorithms on blurred images with outliers.(5)Spatially variant linear representation model for image deblurring.We propose a spatially variant linear representation model(SVLRM),where the target image is linearly represented by the guidance image.To estimate the linear representation coefficients,we develop an effective algorithm based on a deep convolutional neural network(CNN).The proposed deep CNN(constrained by the SVLRM)is able to estimate the spatially variant linear representation coefficients which are able to model the structural information of both the guidance and input images.We show that the proposed algorithm can be effectively applied to a variety of applications,including depth/RGB image restoration,flash/no-flash image deblurring,natural image denoising,etc.Extensive experimental results demonstrate that the proposed algorithm performs favorably against state-of-the-art methods that have been specially designed for each task.