A generalised framework for mapping low-level visual features to high-level semantic features

Lim , Phei-Chin (2008) A generalised framework for mapping low-level visual features to high-level semantic features. Masters thesis, Centre for Academic Information Services.

[img]
Preview
PDF
lim+phei+chin.pdf

Download (2MB) | Preview

Abstract

Current state-of-the -art techniques in addressing the semantic gaps are either dependent on machine learning approaches or knowledge-based approaches, which is seen to be lacking if either approaches is utilised individually. Hence, a systematic three-phase framework for addressing the semantic gap is presented in this study. Visual features are extracted from images using image processing algorithms, while textual features(domain concepts) are extracted through knowledge engineering processes. This research emphasises on the semantic mapping phase where mapping between visual features and domain concepts are performed in a two processes. Densities for a visual are visualized in a graphical plot to guide the human operator in mapping the visual feature to a related domain concept. Subsequently, clusters are obtained from k-Means clustering and then mapped with domain concepts' properties. A set of standard semantic concepts is then obtained where classification rules are automatically derived by employing the standard semantic concepts. A prototype system is built to demonstrate the effectiveness of the framework and the knowledge acquisition environment for the categorization of butterfly wing images. By providing an environment for bridging the semantic gap, there exists the ability in evaluating the usefulness of each visual feature and acquiring of rules for characterizing butterflies automatically. These rules ate further evaluated using three separate test sets. The average classification accuracy for three test sets is 81.8% while the average error rate was 9.1%. The remaining 9.1% is attributed to unmatched rules derived from test images, which provided the capability in further extending the rules previously obtained. Dynamic customisation for this particular domain was achieved by incorporating terminology from the application domain as textual features.

Item Type: Thesis (Masters)
Additional Information: Universiti Malaysia Sarawak, UNIMAS
Uncontrolled Keywords: Universiti Malaysia Sarawak, UNIMAS, IPTA, education, kuching, samarahan, sarawak, malaysia, universiti, university
Subjects: A General Works > AC Collections. Series. Collected works
A General Works > AC Collections. Series. Collected works

T Technology > T Technology (General)
Divisions: Academic Faculties, Institutes and Centres > Faculty of Computer Science and Information Technology
Depositing User: Karen Kornalius
Date Deposited: 10 Apr 2014 03:49
Last Modified: 23 Mar 2015 08:21
URI: http://ir.unimas.my/id/eprint/1691

Actions (For repository members only: login required)

View Item View Item