1 Interest of the Combination of Classifiers for Volumetric Textures Classification Elmoez BEN OTHMEN, Mohamed Ali CHERNI and Mounir SAYADI SICISI Unit, ESSTT, University of Tunis, 5 Av. Taha Hussein, 1008, Tunis, Tunisia E-mails: moez_ben_othmen@yahoo.fr, mohamed.ali.cherni@gmail.com, mounir.sayadi@esstt.rnu.tn Abstract- Nowadays, classification is applied in various fields such as pattern and writing recognition, prints checking, faces identification, medical images analysis, 2D textures characterization and volumetric textures characterization. Indeed, the three- dimensional field is considered among one of the most important fields in image processing because of the great quantity of information that can be extracted. In this work, we try to improve the performances of classification for volumetric textures images by proposing a multiple classifier systems (MCS) based method combining three Euclidean classifiers: simple Euclidean classifier (ES), normal Euclidean classifier (EN) and balanced Euclidean classifier (EB). Thereafter, we compared the performance of the proposed method to the Euclidean methods (ES, EN and EB). The hybrid presented approach has proven to be more efficient in classification and mostly robust against Gaussian noise. Keywords- volumetric images textures, combination of classifiers, Multiple to Classify System: MCS I. INTRODUCTION The analysis of the textured images is an important field and many researchers worked on this axis. The field of image processing can be divided into three axis: segmentation, synthesis and classification. The literature summarizes the extraction in various types: statistics, parametric and frequential. All these methods were mainly developed and tested on two-dimensional texture. Recently, some of these methods were studied to analyze volumetric texture. In fact, the three- dimensional field is very rich in information what makes the classification very complex and highlights the concept of combination of the classifiers [5]. Indeed, several classifiers can deliver different answers for the distribution of the image and the class to which it corresponds. This is due mainly to the specific error of the classifier. This error rises from the model of decision of the classifier and the used database. The behavior of each classifier is given by providing different basic information for the textured images. The various results of the classifiers are then combined in order to improve classification. Our article is organized as follows: section II describes the volumetric texture classification using the co-occurrence matrix (GLM3D). Section III presents the freely accessed database of volumetric textures used in this paper. Also, the aspect of combining classifiers is raised in this section. Moreover, the following section, section IV, gives the significance of classifiers as well as the definitions of the various types of studied Euclidean classifiers as well as the system of combined classifiers (MCS). Section V demonstrates the superiority of the hybrid proposed method of classification against the methods containing single classifiers as well as the robustness of this method against Gaussian noise. Finally, section VI concludes this work by resuming the performed works. II. DESCRIPTION OF CO-OCCURRENCE MATRICES FOR VOLUMETRIC DATA In this section, we present the 3D matrix of co- occurrence or the space method depending on gray levels. In fact, it makes it possible to determine the frequency of appearance of a formed "distance" for voxel separated by a certain distance D in a particular direction. A co-occurrence matrix for volumetric data is an n x n matrix, where n represents the number of gray-levels within an image. For reasons of speed computing, the number of gray levels can be reduced if one chooses to bin them .Thus, the size of the co-occurrence matrix is decreased This matrix acts as an accumulator so that M [i , j] counts the number of pixel pairs having the intensities i and j. However, this matrix is defined by specifying a displacement d = (dx, dy, dz), where dx and dy are the same as described for 2D co-occurrence matrices, and dz represents the number of pixels moved along the z-axis of the three-dimensional image. We take the matrices while resulting and measuring the space dependence of the values of gray-level by computing the devices of following texture of Haralick[3]. U.S. Government work not protected by U.S. copyright