On Increasing the Objectiveness of Segmentation Results

Abstract

Unsupervised image segmentation is a complicated task due to the lack of appropriate measures for judging the quality of segmentation results. At present, most segmentation results are evaluated visually and qualitatively. In a previous paper [1], we presented a system that performs image segmentation without using neither ground-truth information nor human judgement. A genetic algorithm is used to efficiently search the set of parameters that maximises a segmentation quality criterion based on the integration of information from region boundaries and edge pixels. The performance of that system is degraded by the need of binarizing the edge maps, resulting in an information loss that may even hinder the convergence of the GA. In this paper we describe the solutions adopted to overcome this problem and to increase the objectiveness of segmentation results. Basically, human visual criteria are firstly embodied in an automatic image-contrast enhancement-process. The connectivity-code intermediate representation [2] and saliency-enhancing operations derived from the psychology field are then used to increase the significance of edge-pixels that have a small intensity but a high perceptual importance. Both computer simulated and real images have been used to test the performance of our adaptive segmentation system. The increased objectiveness of the segmentation results is confirmed both perceptually and by higher fitness values than those obtained by the previous system.