Research On Images Blind Deblurring Based On Blind Deconvolution

In remote sensing imaging and medical imaging, image resolution and contrastwill be decreased because of the complexity of imaging conditions. So imagerestoration has being a research focus in computer vision and image processing.Existing image restoration algorithms, such as Wiener filtering, built on the knownpoint spread function. However，point spread function were often unknown in practicebecause imaging conditions were unknown, like the relative motion between cameraand objects and atmospheric disturbance. Therefore, the research on blinddeconvolution and blind restoration is very necessary in theory and practice.This dissertation was mainly to the blurred images whose prior information wasunknown or partially known. We detailed researched on the basic theories and keytechnologies in two spects of image blind restoration and image blind assessment. Forimage blind restoration, we discussed two spects of blur type being known andunknown, deeply researching on key problems about parameter estimation andregularization technique. For image blind assessment, we respectively researched onno-reference image assessment algorithms of noise, blur degree and blocking effectsand comprehensive evaluation algorithm that was comprised of three distortionfactors. The main innovations and research results are as following:1. For the known blur type, this dissertation only studied the single blur type. Wefirstly established the parameter model of point spread function, estimated its parameters by the prior knowledge about degraded image and imaging system andthen restored the blurred image by using the estimated point spread function model.Based on this theory, after mainly studying the characteristics of defocused blurredimage, a novel scheme, which was based on the Hough transform, was proposed forestimating the radius of point spread function of defocused blurred images. Thisscheme was highly accurate and highly stable. It’s very useful for parametersestimation in the processing of image restoration.2. For the unknown blur type, the estimation of point spread function and imagerestoration were carried on simultaneously. A super total variation image blinddeblurring method with self adaptive threshold was proposed to restore the imagesdegraded by unknown point spread function. On the one hand, based on the analysisof the total variation model, the super total variation with self adaptive threshold wasproposed and the threshold was deduced by the estimation noise of image. In the case,the edge areas of image were preserved and the smooth areas were denoised in theprocessing of image restoration. On the other hand, the exact solution of optimizationmodel was obtained by using semi-quadratic regularization and auxiliary variables,which overcome shortcomings of using too many approximations in many traditionalsolutions. The experimental results demonstrated that the restoration image has moredetails and fewer blocking effects.3. This dissertation analyzed existing no-reference image quality assessmentalgorithms. Since these algorithms were not consistent with subjective assessment, anovel no-reference image quality assessment method which was based on noise，blurdegree and blocking effects was proposed. It introduced three types of imagedistortion, including noise, blur degree and blocking effects. Firstly, the standarddeviation of image noise was estimated by modified wavelet medium estimation.Secondly, the blur degree of image was obtained by using edge pixel points. Thirdly,blocking effects was represented by using characteristics of image pixel blocks.Finally, the assessment model was established by combining foregoing threedistortion types. Combining the Differential Mean Opinion Scores (DMOS) provided in the LIVE IQA DataBase, The evaluation values of this algorithm not only agreedwith PSNR in objective assessment, but also were consistent with the DMOS insubjective assessment.4. Combining the objective and subjective image quality assessment, thisdissertation analyzed and evaluated the experimental results from the stability, theefficiency and the application of algorithm, visual effects of images, as well as imageevaluation indices, like PSNR, image entropy. The experimental images and data werepresented. Also, this dissertation compared our algorithms with the same typealgorithms in effect of image restoration and image quality assessment.