Chinese Journal of Computers   Full Text
Title	Hybrid Linear Model Based Image Denoising
Authors	CAO Yang LUO Yu-Pin YANG Shi-Yuan
Address	(Department of Automation, Tsinghua University, Beijing 100084)
Year	2009
Issue	No.11(2260—2264)
Abstract &   Background	Abstract Most of previous image denoising method assume image signal is piecewise smooth. They suppress oscillating patterns in the image to denoise. This kind of method remove high-frequency signal from texture image. This paper proposes a hybrid linear model based image denoising method in order to preserve texture signal. The new method doesn’t use piecewise smoothness assumption. Supposing image signal is self-similar while noise is not. Statistical learning algorithm is used to cluster the image regions and extract the principle component. The principle component is used as the image signal to form the denoised image, so as to preserve most detail signal of texture image. Keywords image denoising; self-similarity; hybrid linear model Background Most of previous image denoising methods assume image signal is piecewise smooth. This kind of method remove high-frequency signal from texture image. Some recent researches utilize self-similarity of image. These methods suppose that regions of one image are similar. They try to search for similar regions and from which extract underlying image patterns. Using the patterns, denoised image can be reconstructed. In general, smooth region and edge region can be seen as two kinds of patterns. Smooth based method can only deal with these kinds of image regions. Self-similarity based method can deal with all kinds of image regions. In this paper, a new self-similarity based method is proposed. New method supposes all image regions have only a few image patterns and every pattern form a subspace in the data space. Since noise is not limited in subspace, it can be canceled by project the noisy data onto the subspaces.