Clustering Validation Indices for Sintonization of Automatic Segmentation Techniques Using Natural Scene Images

Abstract

Automatic image segmentation is a fundamental task in many applications such as video surveillance, image recovery, medical image analysis, recognition, tracking and classification of objects.  Most of the works reported in the literature on digital images segmentation need a reference image to evaluate their segmentation. However, in real-life scenarios there is not always a reference image to carry out this task. In this sense, this paper presents a study of the use of clustering validation indices and maximum entropy as cost functions to quantify the segmentation quality in order to find the optimal parameters for the automatic segmentation of digital images. The clustering validation indices allow us to have a quantitative value on how well the segmentation is done without considering any reference image. We carried out the evaluation of three segmentation algorithms, K-Means, Watershed and SRM, using four image databases with images containing different numbers and sizes of objects and different levels of illumination. The obtained results reveal that the clustering validation indices improve the precision in the identification of objects within the images in the three automatic segmentation algorithms, while these results are competitive with works reported in the literature.