Adaptive Blind Image Restoration Based On Total Variation Regularization And Multi-resolution Analysis

Image deblurring and restoration problems are well known inverse problems.Blind image restoration, which recovers the image and simultaneously identifies theblur, is always a big challenge. This dissertation intensively explores blind imagerestoration methods which are mainly based on two kinds of techniques:multi-resolution analysis and total variation (TV) regularization.Wavelet and multi-resolution analysis theory is one of the most useful tools inthe image processing field. Multi-resolution analysis, especially the wavelet transformprovides unique benefits in image representation and processing. While waveletapplications in image compression have become extremely prevalent, their use inrestoration has not been exploited to the same degree. The application of wavelet inrestoration is the main work in this dissertation.Partial Differential Equation (PDE) is the most potential mathematic tool in theimage processing field, total variation regularization based on PDE has achieved greatsuccess in image restoration and preserving edges. The major focus in this research isto get more precise understanding of total variation regularization model to constructmore effective total variation minimizing restoration schemes which have adaptiveability.Firstly, we introduce some basic knowledge about image restoration and relatedmathematic theory, analyze the degradation model of image and the ill-posed problemduring restoration, and we develop an automatic algorithm using Fourier-Mellintransform to classify and identify the point spread function with cepstrum. Secondly,we introduce some classic methods of image restoration. Next, we propose a doubly local Wiener filtering denoising method using adaptive directional windows and MeanShift algorithm, the resulting denoising performance is substantially close to the bestresults reported in recent literature. Then, we propose an adaptive total variation blindrestoration algorithm, which can automatically change parameters according to thenoise level, and experimental results show that it improves the restoring performancesignificantly under the presence of high noise level. At last, a hybrid multiframe blinddeconvolution algorithm is proposed for restoring turbulence-degraded images, whichcombines multi-resolution analysis and total variation regularization, recovers imagesremarkably, and reduces the calculating workload greatly.