Proceedings of 2005 International Symposium on Intelligent Signal Processing and Communication Systems EXPLICIT LOCAL SEGMENTATION BASED IMPULSIVE NOISE REDUCTION FOR COLOR IMAGES Mieng Q. PhuI, Peter E. Tischer and Hon R. Wu2 'Clayton School of Information Technology, Monash University, Clayton Campus, VIC, Australia. Mieng.Quoc.Phu @ infotech.monash.edu. au Peter.Tischer @ csse.monash.edu. au 2Software and Network Engineering, RMIT University, City Campus, VIC, Australia. henry.wu @ rmit.edu. au ABSTRACT A family of local segmentation vector filters for color image noise suppression and detail preservation is proposed. Most state-of-the-art filters alleviate impulse noise well but tend to destroy thin lines, edges and fine image details. The proposed filters facilitate local segmentation to preserve image structures and noise suppression. First the K-VMF is developed and used for local segmentation, and then a selection of vector filters is used to reconstruct the current pixel. In addition, once pixels have been marked as being noisy, their values are not used in processing subsequent pixels. The proposed filters also demonstrated acceptable results for both objective and subjective assessments. 1. INTRODUCTION In the specific area of color image restoration, filters are either categorized as component-wise or multivariate [1]. Component-wise methods deal with each channel separately, whereas multivariate filters process color pixels as vectors. Since color channels are strongly correlated, vector filters tend to perform better and produce fewer artifacts such as color bleeding and distortions. When dealing with impulsive noise, where the original pixels are completely replaced by random noise, some people advocate that the most efficient filtering approach is based on the vector order-statistic theory [2]. One of the most popular vector filters is the Vector Median filter (VMF) [3]. Other filters include Vector Direction filter (VDF/GVDF) [4], Directional Distance filter (DDF) [5], Hybrid Directional filter (HDF/AHDF) [6] and Adaptive Nearest Neighbour filter (ANNF) [7]. These classic filters remove impulse noise adequately but they tend to introduce new artifacts to image structures such as blurring, smearing and shifting. This happens because they do not classify pixels as been clean, noise, blotch of noises, or high image detail. The Multiple Window Configuration (MWC) [8] solves this drawback by the use of detection and switching. However, this filter depends on the reference image and fine tuning on specific images to achieve optimal results for that image. Another switch based filter is Neighbour Adaptive Vector filter (NAVF) [9]. All of these filters implicitly or explicitly assumed that the current window is homogeneous. This is true on most parts of the image, but there are edges, thin lines and outliers as well. Thus, the assumption of homogeneity will lead to the removal of not only noise but also the image structure. The Peer Group filter (PGF) [10] uses Fisher's Discriminant to segment pixels explicitly into groups with similar intensity and to reconstruct using the VMF and a weighting function. 2. VECTOR MEDIAN FILTER VMF is one of the most efficient and popular filter because of its simplicity and low computational cost. It is extensively used, for example, as an impulse detector, and in hybrid and switch based filters. Let x be as a multichannel sample vector and W be the window (xl, x2,..., xc) . For pixel xi, its total distance to the other pixels in the window is given (1) n d (xi) = d(xiS,xj) j=l The distance d (xi, xj) is often the Euclidean distance. Let the dw (xi) be sorted in ascending order to produced {d1, d2, ... ., I, } . The pixel associated with dI is the most 'inlying' value and is chosen as the vector median (VM) and used as the output of the VMF. Although the VMF can remove most impulsive noise, it tends to destroy image structure and blur the image as in Figure 1. When there is no noise within an image the VMF will destroy thin lines and edges. The VMF and its 0-7803-9266-3/05/$20.00 C2005 IEEE. December 13-16, 2005 Hong Kong (1) 657 - Authorized licensed use limited to: RMIT University. Downloaded on November 27, 2008 at 19:43 from IEEE Xplore. Restrictions apply.