Automated mood boards - Ontology-based semantic image retrieval
Syed Abdullah, Engku
The main goal of this research is to support concept designers’ search for inspirational and meaningful images in developing mood boards. Finding the right images has\ud become a well-known challenge as the amount of images stored and shared on the Internet and elsewhere keeps increasing steadily and rapidly. The development of\ud image retrieval technologies, which collect, store and pre-process image information to return relevant images instantly in response to users’ needs, have achieved great\ud progress in the last decade.\ud However, the keyword-based content description and query processing techniques for Image Retrieval (IR) currently used have their limitations. Most of these techniques\ud are adapted from the Information Retrieval research, and therefore provide limited capabilities to grasp and exploit conceptualisations due to their inability to handle\ud ambiguity, synonymy, and semantic constraints. Conceptual search (i.e. searching by meaning rather than literal strings) aims to solve the limitations of the keyword-based\ud models.\ud Starting from this point, this thesis investigates the existing IR models, which are oriented to the exploitation of domain knowledge in support of semantic search\ud capabilities, with a focus on the use of lexical ontologies to improve the semantic perspective. It introduces a technique for extracting semantic DNA (SDNA) from\ud textual image annotations and constructing semantic image signatures. The semantic signatures are called semantic chromosomes; they contain semantic information\ud related to the images.\ud Central to the method of constructing semantic signatures is the concept disambiguation technique developed, which identifies the most relevant SDNA by measuring the semantic importance of each word/phrase in the image annotation. In\ud addition, a conceptual model of an ontology-based system for generating visual mood boards is proposed. The proposed model, which is adapted from the Vector Space Model, exploits the use of semantic chromosomes in semantic indexing and assessing the semantic similarity of images within a collection.