Abstract
We propose a method that uses multiple Wiener filters based on region decomposition to restore natural images degraded by Gaussian noise. Usually, a Wiener filter is designed to be optimal for the entire image. However, because natural images are highly nonstationary, an effective method switches to the optimum filter for each local region. In this paper, we consider a method that repeatedly decomposes into four equal parts, the four regions obtained by dividing an image into four equal parts, and design a Wiener filter for each region. As the regions become smaller, the desired Wiener filter is not necessarily obtained because of the drop in reliability of the correlation functions used to find the Wiener filter. Therefore, we first study to what extent the restoration accuracy of a known source image is improved by ideal decomposition that repeatedly decomposes only regions where the accuracy will improve. We then present a practical method that sufficiently improves the restoration accuracy even when the source image is unknown. We next merge similar decomposed regions and investigate how many classes are essential for the decomposition and these class boundaries. Thus, we verify that the proposed decomposition implicitly accounts for properties such as the contours in the image and clearly demonstrate the suitability of the proposed method. ยฉ 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 85(8): 51โ61, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1112