Regularizing image restoration in uncorrelated transform domain and its applications
Sheung‐On Choy, Yuk‐Hee Chan, W.C. Siu · 1999
The main thrust of our present work is the study on the following areas where various problems remain unsolved, not properly solved, or not effectively solved: (1) performance assessment of image restoration; (2) regularization of image restoration; and (3) applications of image restoration technology in the field of image coding. In our study. some major pitfalls of using the SNRI are pinpointed, and an improved measure termed “restoration score” is devised. The proposed measure is based on a metric for pixel fidelity improvement and an incorporation of the main properties of the human visual system. It is shown that this measure is more precise than the SNRI. Another novelty is that it contains clearly-defined and meaningful reference points in its measurements. These reference points are useful in providing users with a better insight into the effectiveness of the restoration method under evaluation. In our study, we propose a more effective scheme for regularizing image restoration. In this scheme, where to emphasize restoration and where to do more regularization are evaluated separately in two independent uncorrelated transform domains, and the utilization of various statistical information concerning the image and the noise are also suggested to improve the regularization. The main advantages of the proposed regularization scheme include: (1) better utilization of a priori information regarding the ideal image and the noise; (2) higher resistance to the estimation error of the regularization weights; and (3) greater capability for a wide range of image restoration applications. In the area of image noise filtering, we derive an uncorrelated transform-domain estimator, which is then shown to be a generalized version of the adaptive LMMSE estimator. In the proposed filtering method, the uncorrelated transform-domain local statistics obtained from the noisy image are exploited. In the proposed deblurring algorithm, the smoothness imposition on the solution is adjusted by means of weighting the decorrelated image components according to their local variances computed from the processed image. It is shown by detailed experimental results that the proposed regularization scheme is more effective in suppressing the various artifacts in image deblurring and provides a better deblurring result, as compared with the conventional spatially adaptive regularization scheme. The proposed image restoration method also leads to a fast and effective deblocking technique for block-transform encoded images at high compression rates. This good performance is achieved because of the following advantages the proposed technique has: (1) efficient incorporation of the solution bound into restoration; and (2) effective exploitation of local image properties and statistical knowledge about the quantizers used. It is encouraging to find that, for all the problems concerned, the various algorithms we devise based on the proposed regularization scheme perform significantly better than many state-of-the-art techniques used to deal with the corresponding problems. (Abstract shortened by UMI.)