TEM Journal. Volume 8, Issue 2, Pages 372-382, ISSN 2217-8309, DOI: 10.18421/TEM82-08, May 2019. Visual Domain Ontology using OWL Lite for Semantic Image Processing Ahmad Adel Abu-Shareha, Ali Alshahrani ITC Department, Arab Open university, Riyadh, Saudi Arabia Abstract – In this paper, a visual domain ontology 1. Introduction (VDO) is constructed using OWL-Lite Language. The VDO passes through two execution phases, namely, Several image-processing applications, such as construction and inferring phases. In the construction image classification, object recognition, and image phase, OWL classes are initialized, with reference to disambiguation, are based on a supervised or semi- annotated scenes, and connected by hierarchical, supervised learning process, which involves training spatial, and content-based relationships and predicting phases. The training phase used (presence/absence of some objects depends on other objects). In the inferring phase, the VDO is used to training data to construct a model that usually infer knowledge about an unknown scene. This paper handles low-level features and/or image labels. aims to use a standard language, namely, OWL, to Predicting employs an input with unknown or represent non-standard visual knowledge; facilitate ambiguous labels and uses the constructed model to straightforward ontology enrichment; and define the predict output labels. These simple applications rules for inferring based on the constructed ontology. handle minimum visual information provided by the The OWL standardizes the constructed knowledge and constructed model. The new trends in computer facilitates advanced inferring because it is built on top vision and image processing move toward semantic of the first-order logic and description logic. The VDO technique for image classification, object then allows an efficient representation and reasoning of recognition, segmentation, and disambiguation. complex visual knowledge. In addition to Implementing the complex recognition and representation, the VDO enables easy extension, sharing, and reuse of the represented visual knowledge. predicting tasks of semantic techniques requires an extensive and expressive knowledge source to Keywords – Semantic Image Processing, Image replace the temporary models constructed on-the-fly Knowledge Representation, Visual Domain Ontology, in the training phase. Ontology is one of the well- OWL. known sources of knowledge. It is a formal knowledge representation that handles a set of concepts within a domain, their relationships, and properties [1]. Constructing an ontology to handle visual DOI: 10.18421/TEM82-08 knowledge is not a trivial task. Visual knowledge is https://dx.doi.org/10.18421/TEM82-08 complex and heterogeneous. For example, features Corresponding author: Ahmad Adel Abu-Shareha, may be represented as a vector of numerical values ITC Department, Arab Open university, Riyadh, with uncertainties. Spatial information may be Saudi Arabia represented as direction in 2D space, and content Email: [email protected] information may be represented as tuple of two or more entries. Ontology for the visual domain should Received: 04 March 2019. represent such heterogeneity and enable high-level Accepted: 25 April 2019. prediction and recognition. Moreover, ontology Published: 27 May 2019. should be general-purpose for various applications © 2019 Ahmad Adel Abu-Shareha, Ali and for facilitating the links between image Alshahrani; published by UIKTEN. This work is licensed applications toward a semantic image processing under the Creative Commons Attribution- pipeline[1]. NonCommercial-NoDerivs 3.0 License. Several existing approaches have been proposed to construct ontologies for semantic image-processing The article is published with Open Access applications, such as using ontologies for semantic at www.temjournal.com feature extractions [2],[3],[4], semantic image segmentations [5],[6], and semantic object recognition [7],[8]. These approaches are pioneers in 372 TEM Journal – Volume 8 / Number 2 / 2019 TEM Journal. Volume 8, Issue 2, Pages 372-382, ISSN 2217-8309, DOI: 10.18421/TEM82-08, May 2019. using ontologies, and their constructed and utilized via hierarchical and non-hierarchical relationships to ontologies are task oriented but suffer from the create the ontology structure, which allows the following drawbacks: storing low-level features and ontology to be represented broadly by general- their object identities only; having a weak structure; purpose data structures of graphs and trees. The and limited ability to fit the task at hand because of stated relationships between the underlying ontology the utilization of simple encoding languages and concepts are guaranteed through the relationships techniques. Moreover, these techniques cannot (represented by edges) among these concepts. construct a consistent complex visual knowledge Moreover, these edges infer relationships in the so- source. called reasoning. The ontology specification can be Scholars must construct visual knowledge represented using general-purpose programming ontology by using standard knowledge representation languages or tools; this representation is denoted as languages, such as RDF and Web Ontology general representation of ontology. However, this Language (OWL). This paper introduces visual representation cannot construct complex knowledge domain ontology (VDO), a formal representation of because the ontology grows; this phenomenon leads visual domain knowledge, by using OWL Lite. The to high chance for inconsistency and connectivity well-defined syntax of OWL represented in RDF loss because of the lack of rules to control the triples and RDF graph, its formal semantics, and its construction. Figure 1. shows a simple visual sufficient expressive power can efficiently represent ontology that represents visual information by using complex knowledge in a restricted visual domain [9]. a general-purpose graph. “Outdoor” is a concept This paper is organized as follows. Section 2 connected through a hierarchy relationship to provides a brief description of the ontology. Section “Everything” in high levels and “Animal” in low 3 highlights some of the previous works in the levels. Non-hierarchy relationship “on-top” connects ontologization of image knowledge. Section 4 concepts to one another. Properties such as “has- describes the structure and components of the Texture” and “has-Color” associate property values proposed VDO, its implementation using OWL Lite, with a specific concept. and the reasoning mechanisms that must be incorporated with the ontology. Section 5 presents Everything the implication of the proposed work. Section 6 concludes the study. Outdoor Indoor 2. Ontology Ontology is a conceptual knowledge representation that can be interpreted by both human and computer. Ontology of a given domain should cover the abstract Grass Sand information for that domain, that is, domain on-Top has-Texture components. Ontology of visual knowledge should has-Color Animal represent visual information, which can be Green Smooth summarized as follows: low-level features that are directly extracted from the images and images’ uncertainty uncertainty objects; the association of the low-level features to 0.3 0.1 object identity (feature-to-object) under uncertainty values; content information that represents the Figure 1. Graph Representation of Visual Knowledge presence/absence of several objects together in the Ontology scene; and objects’ spatial relationships and context information represented by non-visual domain- Advances in ontology representation have been knowledge [10]. For ontology components, associated with the semantic web and the emergence ontology involves concepts and hierarchical of standard ontology languages, such as RDF and relationships, which are used to represent and OWL. OWL prevents inconsistency and looseness categorize the domain components. In an ontology, and is considered an inference technique because it is each category is represented by a concept and built on top of the first-order and description logic. different categories are connected to each other Ontology concepts are defined using such languages, through hierarchical relationships. In addition, besides existing hierarchical relationships and ontology may involve additional components, such properties, as ontology classes; actual objects, which as properties and other relationships, which enrich belong to a specific ontology class, are represented as the ontology and enable efficient knowledge individuals. The OWL syntax is represented in RDF representation. Concepts are connected to each other triples. The triple structure is {subject – relationship TEM Journal – Volume 8 / Number 2 / 2019. 373 TEM Journal. Volume 8, Issue 2, Pages 372-382, ISSN 2217-8309, DOI: 10.18421/TEM82-08, May 2019. – object}. Figure 2. shows an example triple. {Apple – hasColor – Red}, {Grass – hasColor – However, visual knowledge requires a tuple of more Green}, the uncertainty triples associated with the than three elements. For example, a five-tuple (Apple, Grass) are {Apple – value – 0.6} and {Grass representation is required, structured as {Object – – value – 0.7}. These five triples will also lead to hasFeature – Feature – hasUncertainty – disassociation. The same convention applies if the Uncertainty}, to encode low-level features with uncertainty triples are associated with the associated uncertainty. A triple can encode the first relationships (e.g. “hasColor”). three elements by