Semantic object recognition by merging decision tree with object ontology Wafa Damak Higher Institute of Computer and Multimedia Computer Imaging Electronics System (CEM Lab) Sfax, Tunisia damakwafa@gmail.com Issam Rebai Department of Computer Science Telecom Bretagne Brest, France issam.rebai@telecom-bretagne.eu Imene Khanfir Kallel Computer Imaging Electronics System (CEM Lab) Sfax Engineering School Sfax, Tunisia imen.khanfir@betatex.com.tn Abstract—In this work, we propose an object recognition strategy in a domestic environment. Our contribution is to use low-level features extracted from images with high-level concepts generated from an ontology of domestic objects to get richer decision. It consists in developing a semantic classification by providing for a white cane user the class of the obstacle and the scene in which it is located. The classification is performed with a decision tree that provides a better recognition rate than SVM. The combination of color and texture features resolves the ambiguities of shape features for some objects that have similar shape. Keywords—Object recognition; classification; learning; SVM; decision tree; ontology. I. INTRODUCTION Color, texture and shape features, called low-level features are directly extracted from numeric image representation and, without human interpretation, they have no link with semantic present in the image. So, these features are independent from usage context and semantic abstraction. These latest are not included in image, but are found elsewhere. Then, we must find a way providing image semantics. Several mechanisms of knowledge domain representation exist in the literature. Recent works use ontology to clearly describe knowledge about one research field. It defines a set of concepts, their characteristics and their relationship between each other. An ontology is a specification of an abstract and simplified world view that is represented for a specific purpose [13]. It serves to analyze the knowledge in a field through modeling it in order to identify and to provide its semantics. In this paper we are interesting in developing and implementing a recognition strategy of objects detected in an indoor environment (airport, shopping mall, etc.). In order to simplify the scene, we are restricted to the case of an habitation. This work is intended for a visual impaired people device (a smart white cane), helping them to be familiarized with their environment. Since the white cane user needs to be informed of any obstacle crossing his way: what does it like? and what is the scene in which he is located?, semantic based approach is the most appropriate for these needs. Our contribution aims to enrich our object classifier with a domestic object ontology in order to inform visually impaired of the context in which he is located (kitchen, bathroom, etc.). The remainder of this paper is organized as follows: In section II, we present the related work. Section III defines the proposed approach. We then present test and evaluation in section IV, and enclose with conclusions and perspectives. II. RELATED WORK Object recognition is one of the most active research fields in computer vision since several classification approaches have been proposed [2]. The recognition process is considerably complex because of object appearance variation in images. Among the most important causes of object appearance variation, we cite viewpoint changing, where object undergoes geometric transformations (rotation, translation, scaling), brightness change, occultation (some object parts may be not visible) and intra-class variation where multiple objects can have very different visual appearances such as different chair models illustrated in Fig. 1. When intra-class variation is more important than interclass variation, the recognition task becomes impossible. Fig. 1. Differents models of chairs. The various object recognition approaches can generally be grouped into two main categories: visual content based 1st International Conference on Advanced Technologies for Signal and Image Processing - ATSIP'2014 March 17-19, 2014, Sousse, Tunisia IFI-145 978-1-4799-4888-8/14/$31.00 ©2014 IEEE 65