Unsupervised color image segmentation using a new Neural-Morphological procedure

Unsupervised color image segmentation using a new Neural-Morphological procedure

Hassan A. Muthanna1, R. Touahni1, A. Sbihi2, S. Eddarouich3 , R. Messoussi1

1IbnTofail University, Faculty of Sciences, LASTID, Kenitra, Morocco

2 Abdelmalek As-Saadi University, ENSA, Tanger, Morocco

3 Regional Educational Center, Rabat, Morocco

hassan55m2000@yahoo.com

Abstract— In this paper, we present a new unsupervised color image segmentation procedure using an image pixels classification. This procedure is based on the artificial neural network with competitive learning and mathematical morphology for detecting the underlying probabilities density function’s (pdf) modes. This function describes the distribution of the image pixels that composed of the three color components in the RGB space. The network training algorithm allows to localize local maxima of the (pdf), considered as the markers of the modes of this function. Then, the modes are detected using a new technique based on the morphological thickening transformation. The so detected modes are then used for the classification process. The proposed procedure does not pass by any thresholding and does not require any information a priori on the number of classes nor on the structure of their distributions in the sample.

Keywords— Color image segmentation, competitive learning, mathematical morphology, mode detection, pattern classification.

I. INTRODUCTION

Color image segmentation is one of the most important pre- processing step towards image understanding, image compression and coding. It is a process that consists to partition the image into disjoint region as sets of connected pixels that are homogenous with respect one or more color characteristics [1]. Generally, there are four main approaches to the segmentation of color image: histogram-based segmentation techniques, neighbourhood-based segmentation, physical-based segmentation techniques, and multidimensional data classification methods. In these approaches, most algorithms treated are based on threshold selection or in parameters adjustment which may change the segmentation results.

For the multidimensional data classification methods [2, 3], cluster analysis techniques attempt to separate a set of multidimensional observations into groups or clusters which share some properties of similarity. The objects are generally represented by N-dimensional vectors of observed features. The statistical approach in cluster analysis postulates that the input pattern are drawn from an underlying probability density function (pdf) which describes the distribution of the data points through the data space. Regions of high local density, which might correspond to significant classes in the population, can be found from the peaks or the modes of the density function estimated from the available patterns[4]. Then, the key problem is to partition the data space with a multimodal pdf into subspaces over which the pdf is unimodal [5].

Many clustering procedures based on modes detection concepts, have been proposed. In some of them, modes are considered as local maxima

...