A K-means Based Generic Segmentation System

This paper presents a creative general purpose segmentation system, potentially capable of object extraction from RGB images. The segmentation takes place by initially performing K-means clustering and then recombination. K-means algorithm uses RGB color values, diagonal busyness factor (sum of color differences among central and diagonal pixels) and epsilon spatiality factor (sum of Euclidian distances of pixels belonging to a particular cluster from their cluster center) as its clustering parameters in order to produce optionally compact or loose clusters representative of inherent color and texture. In addition, three different distribution methods are introduced to initialize central points and therefore improve the clustering accuracy. The methods are evenly spaced values, random values and evenly spaced samples respectively. In evenly spaced values, clusters (central points) are evenly distributed along the range of RGB colors available with in the image so that each cluster may partially represent a sub range of the total range of colors available with in the image. In random values, the distribution of clusters (central points) is not even. To define a central point an RGB color is randomly selected from the range of RGB colors available with in the image. In evenly spaced samples the distribution of clusters (central points) is based on X and Y (width and height) coordinates rather than color. To obtain an even distribution, total number of image pixels are calculated and then divided by the number of clusters (central points). Recombination is performed by scanning the neighborhood of each pixel in eight connected directions and determining the class (cluster) to which majority of neighbors belong. The class of central pixel (the pixel whose neighborhood is scanned) is then changed to the class (cluster) that majority of neighbors belong.