Data mining tools often only use a single type of information. The method proposed in this thesis allows the user to insert relational information into existing data mining tools that are designed... Show moreData mining tools often only use a single type of information. The method proposed in this thesis allows the user to insert relational information into existing data mining tools that are designed for content-based information. It does so by regarding the contents of the neighborhood of an element. In this way, the content variability of elements is reduced by using the homophily in the network. Show less
In this dissertation, I investigate new approaches relevant to content-based image retrieval techniques. First, the MOD paradigm is proposed, a method for detecting salient points in images. These... Show moreIn this dissertation, I investigate new approaches relevant to content-based image retrieval techniques. First, the MOD paradigm is proposed, a method for detecting salient points in images. These salient points are specifically designed to enhance image retrieval accuracy by maximizing distinctiveness. Second, the multi-dimensional maximum likelihood similarity measure is presented, which removes a critical limitation in prior research in this area and provides an improved method of comparing image features. Third, a texture classification method based on low dimensional constructed texture features is introduced which have very low computational complexity and would be suitable for real time video understanding or interactive search of very large image databases. The new approaches are tested on well respected international test sets containing representative imagery. Show less
