Image segmentation by clustering

Abstract

This paper describes a procedure for segmenting imagery using digital methods and is based on a mathematical-pattern recognition model. The technique does not require training prototypes but operates in an "unsupervised" mode. The features most useful for the given image to be segmented are retained by the algorithm without human interaction, by rejecting those attributes which do not contribute to homogeneous clustering in N-dimensional vector space. The basic procedure is a K-means clustering algorithm which converges to a local minimum in the average squared intercluster distance for a specified number of clusters. The algorithm iterates on the number of clusters, evaluating the clustering based on a parameter of clustering quality. The parameter proposed is a product of between and within cluster scatter measures, which achieves a maximum value that is postulated to represent an intrinsic number of clusters in the data. At this value, feature rejection is implemented via a Bhattacharyya measure to make the image segments more homogeneous (thereby removing "noisy" features); and reclustering is performed. The resulting parameter of clustering fidelity is maximized with segmented imagery resulting in psychovisually pleasing and culturally logical image segments.