Edge Preserving MRF For Image Denoising

Denoising is an important task in image processing, and the results will have a direct impact on the following work. At present, many denoising methods are effective in removing the noise, but miss lots of image edges and structure information which causes fuzzy edges. Therefore, how to get rid of the noise and meanwhile, achieving better edge keeping results is the key research direction. Denoising based on MRF (Markov Random Field) transforms denoising into the process of estimating the original image through noise image and MRF models. In order to get better result of edge preserving, the key is to establish a MRF model which can accurately describe edges. Based on the detailed analysis of current MRF denoising models, we describe the image edge characteristics and set up two kinds of MRF models for edge preserving on the basis of edge detection and regional segmentation. This paper’s main work and innovative points are as follows:1. It proposes a MRF denoising algorithm with a combination of edge detection. It combines the gray mutation characteristics of edges with continuity to describe the edge and build a new MRF model. Compared with other models, this model describes gray mutation and its continuity, and it can better describe the characteristics of image edges.2. It proposes a MRF denoising algorithm with a combination of regional segmentation. The algorithm uses regional segmentation knowledge to analyze whether adjacent pixels are located in edges or not for the probability analysis, it can improve the accuracy of distinguishing edges and lower interact between adjacent pixels to achieve edge keeping.The above two kinds of MRF denoising algorithms are based on traditional models and complete the purpose of edge preserving by combining edge detection and regional segmentation to improve discriminative ability of edges. Experimental results show that the proposed method has advantage in the removal of the image noise, especially in the edge keeping which proves the validity of this method.