Scope and method of study. The rapid development of information technologies and the advent of the World-Wide Web have resulted in a tremendous increase in the amount of available multimedia information. As a result, there is a need for effective mechanisms to search large collections of multimedia data, especially images. In order to alleviate some of the problems associated with text-based approaches to image retrieval, content-based image retrieval (CBIR) was proposed. The idea is to search on images directly. A set of low-level features, which can be either global or region-based, are extracted from an image to represent its visual content. Retrieval of images is then done by image example where a query image is given as input by the user. The relevance of a database image to the query image is proportional to their feature-based similarity. Those feature representations deemed the most "similar" are returned to the user as the retrieval set. Unfortunately, human notion of similarity is usually based on high-level abstractions, such as activities, events, or emotions displayed in an image. As a result, images with high feature-based similarity may be completely different in terms of semantics. This discrepancy between low-level features and high-level concepts is known as the semantic gap. Relevance feedback (RF) is a supervised learning technique that, by gathering semantic information from user interaction, can reduce the semantic gap and improve retrieval performance. We can distinguish two different types of information provided by RF. The short-term learning obtained within a single query session is intra-query learning. The long-term learning accumulated over the course of many query sessions is inter-query learning. While intra-query learning has been widely used in the literature, less research has been focused on exploiting inter-query learning. In this dissertation, the problem of mapping the low-level physical characterization of images to high-level semantic concepts is addressed by focusing on inter-query learning in CBIR with both global and region-based image representations. While the focus is on inter-query learning, novel intea-query learning approaches and image representations are also presented.; Findings and conclusions. We demonstrated the superior performance of the proposed approaches over other methods and confirmed that image retrieval performance is constantly improved by the integration of inter-query learning.
展开▼