Semantics Extraction in Visual Domain Based on WordNet

Semantics extraction based on local context detection, for the instance´s entities, out of a global context in what so called contextualization with reference to WordNet is presented. The entities of single visual content intend to express a single or few interrelated topics in a constrained domain. Thus, the context of the entities is detected and identified as the part of the global context that is the mostly shared among the underlying entities. The proposed method is based on Level-Based Overlapping (LBO). First, a tree is obtained by tracing the hypernymy paths of all the entities over WordNet up to the root node. Second, a balanced tree is generated by injecting virtual nodes into the original ones to facilitate level-based selection. Finally, a node selection at each level is implemented over the generated balanced tree, as such, a subset of nodes that represent the maximum overlap is selected at each level through optimization technique. The result of each level is propagated to the next level using filtering. The direct implementation of the proposed approach is utilized for automatic visual domain ontology extraction, which has been addressed previously as manual process. The proposed technique is also, implemented for image disambiguation. The results show the potential of the proposed method for the implemented semantic-based applications.