A New Approach for Unsupervised Classification in Image Segmentation

Image segmentation is a fundamental problem in image analysis and understanding. Among the existing approaches proposed to solve this problem, unsupervised classification or clustering is often involved in an early step to partition the space of pixel intensities (i.e. either grey levels, colours or spectral signatures). Since it completely ignores pixel neighbourhoods, a second step is then necessary to ensure spatial analysis (e.g. with a connected component labeling) in order to identify the regions built from the segmentation process. The lack of spatial information is a major drawback of the classification-based segmentation approaches, thus many solutions (where classification is used together with other techniques) have been proposed in the literature. In this paper, we propose a new formulation of the unsupervised classification which is able to perform image segmentation without requiring the need for some additional techniques. More precisely, we introduce a kmeans-like method where data to be clustered are pixels themselves (and not anymore their intensities or colours) and where distances between points and class centres are not anymore Euclidean but topographical. Segmentation is then an iterative process, where at each iteration resulting classes can be seen as influence zones in the context of mathematical morphology. This comparison provides some efficient algorithms proposed in this field (such as hierarchical queue-based solutions), while adding the iterative property of unsupervised classification methods considered here. Finally, we illustrate the potential of our approach by some segmentation results obtained on artificial and natural images.