Research On Denoising Algorithms For Images Contaminated By Random-valued Impulse Noise

Digital images are frequently corrupted by noise during their acquisition and transmission so that image denoising becomes a continuing fundamental research topic. Although random-valued impulse noise is an important type of noise, few methods for removing it can perform excellently at any noise level in any image. By analyzing and contrasting several typical denoising methods we propose a novel approach based on pixel classification. The proposed method classifies pixels in noisy image into two categories:smooth region pixels and texture region pixels with impulse noise by using cascade filtering windows and then store them in two separate images. We delete outlier pixels further in smooth region image and remove impulse noises while preserving details in texture region image. All pixels which are kept in this step are considered as original pixels and are stored in a new image. At last, we restore the image by details-preserving regularization with these original pixels. Extensive simulation results demonstrate that our method is robust and excellent for different noise levels and images comparing with other typical denoising methods.In this paper, we have two major contributions. One is that we conclude similarity constraint, scale constraint and structure constraint these three properties of cascade filtering window and illustrate its principle deeply, especially for removing noisy clusters. The other is that we propose a novel method for removing random-valued impulse noise based on pixel classification. We differentiate pixels in smooth and texture region by using different measurements so that we improve noise detection accuracy. The proposed method outperforms other state-of-the-art methods in peak signal-to-noise ratio (PSNR) and restoration quality.