Non-extensive entropy algorithm for multi-region segmentation: generalization and comparison - DOI 10.5752/P.2316-9451.2013v1n2p3

Paulo Sergio Rodrigues,Gilson Antonio Giraldi

doi:10.5752/10.5752/p.2316-9451.2013v1n2p3

Abstract

Since eighties, the concept of entropy has been applied in the ﬁeld of image processing and analysis. This concept is based on Shannon entropy which is an application in the Theory of Information of the traditional Boltzmann-Gibbs entropy, proposed to the classical thermodynamic. For decades, it is known that this old formalism of entropy fails to explain some physical system if they have complex behavior such as long-rang and long-memory interactions. Recently, studies in mechanical statistics have proposed a new kind of entropy, called Tsallis entropy (or q-entropy or non-extensive entropy), which has been considered with promising results on several application in order to explain such phenomena. In this paper we proposed an algorithm for image segmentation which is based on this new kind of entropy. Our approach, called Non-Extensive Segmentation Recursive Algorithm (NESRA) is an extension of other previous methodologies to binarize images only. In order to show the robustness of the NESRA performance, we compare it with well known and traditional approaches such as bootstrap, fuzzy c-means, k-means, self-organizing map and watershed image clustering methods. We show that, in several cases, the NESRA is better or overcomes these traditional approaches in distinct class of images

Highlights

Image segmentation plays an important role on the basis of most of all computational vision systems
Regarding the large volume of algorithms designed to accomplish the whole task of image segmentation, we can note two main groups: those designed for the primary task of region clustering according to local features and those composed of small procedures
We can highlight examples such as the well known k-means, ISODATA, mean-shift, fuzzy c-means, bootstrap, self organizing maps (SOM), watershed and several others based on histograms, just to name a few

Summary

Introduction

Image segmentation plays an important role on the basis of most of all computational vision systems. In order to accomplish automatically tasks such as scene recognition, for instance, a software needs initially to separate the scene into salient regions. Regarding the large volume of algorithms designed to accomplish the whole task of image segmentation, we can note two main groups: those designed for the primary task of region clustering according to local features and those composed of small procedures. These small procedures generally post-process the output of the first group’s algorithms in order to finer segmentation. We can cite the works of Makrogiannis, Economou and Fotopoulos (2005), Makrogiannis et al (2005) and Pappas (1992)

Objectives

Results

Conclusion