Uso da divergência generalizada na clusterização aplicada à segmentação de imagens

Abstract

Image segmentation is a fundamental process in applications involving image analysis, as it is responsible for defining the regions of interest for the problem. A widely used technique for image segmentation is clustering. Clustering is based on a measure of similarity between objects to group them into regions. The use of dissimilarity measures in the metric space is quite common in pattern recognition systems, mainly in clustering algorithms. However, some studies have observed limitations with conventional distance functions in the search for similarity. In view of this, research focused on the use of non-metric divergences in the search for similarity between objects in clustering algorithms has attracted the attention of researchers. This work presents an alternative path for image segmentation using a generalized divergence as a measure of dissimilarity in Bregman hard clustering algorithm. This divergence is obtained by generalizing the relative entropies of Shannon and Tsallis, and defined in relation to a deformed exponential j. The work is developed using three deformed exponential functions, called j1, j2 and j3, and comparing their results with the classic K- textit means clustering technique and for some images with the Otsu’s method. The tests were performed for 5 groups of images, in which 2 groups are grayscale images divided into binary segmentation and segmentation in more than two regions, and 3 groups of color images, for tests in the RGB, HSV color space and with enhancement of the RGB levels. In order to analyze the segmentation results, in addition to the metrics accuracy, precision and sensitivity, the Sørensen – Dice similarity coefficient (CSD), and a qualitative evaluation obtained through the average of grades empirically attributed were used. The results showed a good performance of the proposed method with accuracy above 0.9 in all images using the j1 deformed exponential, CSD> 0.7 for 23 of the 25 images tested using the j1 and j2 deformed exponentials and average grade above 7 for 24 of the 25 images for the j1 and j2 deformed exponentials. In view of the results obtained, it is possible to conclude that the proposed method was not only shown to be applicable but also, depending on the application, it may present a better performance than methods already established in the literature.