Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Committee Chair

Frigui, Hichem

Author's Keywords

Content-based image retrieval; Semantic gap; Image annotation; Multimedia


Images, Photographic--Databases; Image processing--Digital techniques


To manage large image databases, Content-Based Image Retrieval (CBIR) emerged as a new research subject. CBIR involves the development of automated methods to use visual features in searching and retrieving. Unfortunately, the performance of most CBIR systems is inherently constrained by the low-level visual features because they cannot adequately express the user's high-level concepts. This is known as the semantic gap problem. This dissertation introduces a new approach to CBIR that attempts to bridge the semantic gap. Our approach includes four components. The first one learns a multi-modal thesaurus that associates low-level visual profiles with high-level keywords. This is accomplished through image segmentation, feature extraction, and clustering of image regions. The second component uses the thesaurus to annotate images in an unsupervised way. This is accomplished through fuzzy membership functions to label new regions based on their proximity to the profiles in the thesaurus. The third component consists of an efficient and effective method for fusing the retrieval results from the multi-modal features. Our method is based on learning and adapting fuzzy membership functions to the distribution of the features' distances and assigning a degree of worthiness to each feature. The fourth component provides the user with the option to perform hybrid querying and query expansion. This allows the enrichment of a visual query with textual data extracted from the automatically labeled images in the database. The four components are integrated into a complete CBIR system that can run in three different and complementary modes. The first mode allows the user to query using an example image. The second mode allows the user to specify positive and/or negative sample regions that should or should not be included in the retrieved images. The third mode uses a Graphical Text Interface to allow the user to browse the database interactively using a combination of low-level features and high-level concepts. The proposed system and ail of its components and modes are implemented and validated using a large data collection for accuracy, performance, and improvement over traditional CBIR techniques.