Fusion of Measures for Image Segmentation Evaluation

Abstract

Image segmentation is an important task in image processing. However, no universally accepted quality scheme exists for evaluating the performance of various segmentation algorithms or just different parameterizations of the same algorithm. In this paper, an extension of a fusion-based framework for evaluating image segmentation quality is proposed. This framework uses supervised image segmentation evaluation measures as features. These features are combined together and used to train and test a number of classifiers. Preliminary results for this framework, using seven evaluation measures, were reported with an accuracy rate of 80%. In this study, ten image segmentation evaluation measures are used, nine of which have already been proposed in literature. Moreover, one novel measure is proposed, based on the Discrete Cosine Transform (DCT), and is thus named the DCT metric. Before applying it in the fusion-based framework, the DCT metric is first compared with some state-of-the-art evaluation measures. Experimental results demonstrate that the DCT metric outperforms some existing measures. The extended fusion-based framework for image segmentation evaluation proposed in the study outperforms the original fusion-based framework, with an accuracy rate of 86% and a large Kappa value equal to 0.72. Hence, the novelty in this paper is in two aspects: firstly, the DCT metric and secondly, the extension of the fusion-based framework for evaluation of image segmentation quality.