Research On Adaptive Regularized Image Restoration

The demand of image restoration is rapidly increased since the obtained images in many conditions are often degraded due to relative motion, atmospheric turbulence, lack of focus, noise existing and other factors. Image restoration is an ill-posed problem. Regularization approach for image restoration is preferred though there are many restoration algorithms. Due to the adaptive regularized approach can restore the image's edge and texture and smooth the noise at the same time, it is advantage than the global regularized approach and has been attended widely in recent years. This thesis focuses on adaptive regularized image restoration.Firstly, how to evaluate the quality of restored image is discussed,and a blind approach based on dividing image into different region for restored image quality assessment is proposed. The degraded image is firstly divided into edge, texture and flat region since the influence of noise and artifacts for various regions are different. Then, the average local variance of edge region and flat region are computed separately. At last, the method to compute the integrated evaluation is given. Compared to the classical statistic approach, the proposed approach need not the original image and taking account the different influence of various contents in image for quality assessment. Secondly, the adaptive regularized approaches for image restoration have been studied from three aspects: noise-distribution estimation, the characteristic of image restoration and prior image constraint.In the study based on noise-distribution, an approach to compute local regularization parameter and design local regularization operator adaptively is proposed according to the distribution of the noise. The approach includes three steps. Firstly, the distribution of noise is evaluated in space domain or wavelet domain. Secondly, the local regularization parameters are computed adaptively according to the distribution of estimated noise. Thirdly, local regularization operator is designed based on the judgment of edge and noise. Compared to the existing adaptive approach, the proposed approach needs not the original image or the accurate noise-variance value to compute the local regularization parameter, and work well in the presence of serious noise. In the study based on the characteristic of image restoration, the approach to construct regularized constraint items is proposed according to the basic characteristic of image restoration. Then, two new regularized constraint items are constructed, which called the artifacts removing constraint item and edge restoration constraint item. New regularized constraint items can remove artifacts and restore edge in maximum extend.In the study based on prior image constraint, the approach using minimum information discrimination to achieve the restored image constrained by prior image is proposed, and two approaches are given, which called minimum-mean-square-error validation and fiducially strategy to count regularization parameter adaptively for new constraint item. The new constrained term not only leads to the much similar restored image to the prior image in gray-value distribution, but also estimate the more accurately point spread function (PSF) in blind image restoration.Lastly, the proposed adaptive regularization approaches applying to super-resolution image reconstruction have been studied. In the situations where both PSF and motion estimation are inaccurate, the proposed approach has better reconstructed result than classical approaches.