Novel Indexing Method of Relations Between Salient Objects

Since the last decade, image retrieval has been used in several application domains such as GIS, medicine, etc. Queries are formulated using different types of features such low-level features of images (histograms, color distribution, etc.), spatial and temporal relations between salient objects, semantic features, etc. In this paper, we propose a novel method for identifying and indexing several types of relations between salient objects. Spatial relations are used here to show how our method can provide high expressive power to relations in comparison to the traditional methods. INTRODUCTION During the last decade, a lot of work has been done in information technology in order to integrate image retrieval in the standard data processing environments. Image retrieval is involved in several domains [3, 6, 10, 13] such as GIS, Medicine, Surveillance, etc. where queries criteria are based on different types of features such as metadata [12, 14, 19], low-level features [1, 2, 16], semantic features [4, 15, 18, 23], etc. Principally, relations between salient objects are very important. In medicine, for instance, the spatial data in surgical or radiation therapy of brain tumors is decisive because the location of a tumor has profound implications on a therapeutic decision [17, 20]. Hence, it is crucial to provide a precise and powerful system to express spatial relations. In the literature, three major types of spatial relations are proposed [5]: • Metric relations: measure the distance between salient objects [7]. For instance, the metric relation “far” between two objects A and B indicates that each pair of points Ai and Bj has a distance grater than a certain value δ. • Directional relations: describe the order between two salient objects according to a direction, or the localisation of salient object inside images [8]. In the literature, fourteen directional relations are considered: • Strict: north, south, east, and west. • Mixture: north-east, north-west, south-east, and south-west. • Positional: left, right, up, down, front and behind. Directional relations are rotation variant and there is a need to have referential base. Furthermore, directional relations do not exist in certain configurations. • Topological relations: describe the intersection and the incidence between objects [9, 11]. Egenhofer [9] has identified six basic relations: Disjoint, Meet, Overlap, Cover, Contain, and Equal. Topological relations present several characteristics that are exclusive to two objects (i.e., there is one and only one topological relation between two objects). Furthermore, topological relations have absolute value because of their constant existence between objects. Another interesting characteristic of topological relations is that they are transformation, translation, scaling, and zooming invariant. In spite of all the proposed work to represent complex visual situations, several shortcomings exist in the methods of spatial relation computations. For instance, Figure 1 shows two different spatial situations of three salient objects that are described by the same spatial relations in both cases: topological relations: a1 Touch a2, a1 Touch a3, a2 Touch a3; and directional relations: a1 Above a3, a2 Above a3, a1 Left a2. The existing systems do not have the required expressive power to represent these situations. Thus, in this paper, we address this issue and propose a novel method that can easily compute several types of FIgure 1: Two different spatial situations relations between salient objects with better expressions. The rest of this paper is organized as follows. In section 2, we present our method for identifying relations. In section 3, we discussed how our method gives better results using spatial features. Finally, conclusions are given in section 4. PROPOSITION The 9-Intersection model proposed by Egenhofer [23], represents each shape “A” as a combination of three parts: interior A°, boundary ∂A and exterior A-. The topological relations between two shapes are obtained by applying an intersection matrix between these parts (Figure 2). Each intersection is characterized by an empty (∅) or non-empty