Statistical learning for multimedia information retrieval and their applications
Qi Tian, Yijuan Lu · 2008
Recent years have witnessed a fast development of Multimedia Information Retrieval (MIR). The content extraction, indexing, and retrieval of large scale multimedia data sets continue to be one of the most challenging and fast-growing research areas. Digital libraries, education and training, media commerce, home media, bioinformatics, biometrics, and medical multimedia databases have created a worldwide need for new paradigms and techniques, such as statistical learning methods on how to browse, search, and summarize multimedia collections. However, statistical learning in Multimedia Information Retrieval still faces some challenges such as semantic gap, small sample size and high dimensionality. In this dissertation, novel methods are proposed and generalized from high to low manner to alleviate these problems and, consequently, facilitate concept selection, data modeling and classification in Multimedia Information Retrieval and other applications. At the high-level, a novel framework to develop a lexicon of high-level concepts with small semantic gaps (LCSS) is constructed. By quantitatively studying and formulating the semantic gap problem, these visually and semantically consistent concepts are automatically selected and show their promising application potential for concept detection, automatic annotation, and multimedia information retrieval. At the middle-level, Fast Adaptive Discriminant Analysis (FADA), Two-Dimensional Adaptive Discriminant Analysis (2DADA) and integrated boosting (i.Boosting) are explored to enhance the Adaptive Discriminant Analysis (ADA). At the low-level, Principal Feature Analysis (PFA) is developed for feature dimension reduction. Those algorithms are successfully applied to several real-world applications on image annotation, face recognition, and microarray analysis. In addition, in the specific application of microarray analysis, to fill a gap between comprehension and interpretation of microarray expression data, we introduce Relevance Feedback to microarray analysis and propose a generalized Kernel Discriminant-EM algorithm (KDEM), which allows for a better classification in a nonlinear feature subspace. We combine Relevance Feedback and KDEM together to construct an efficient and effective semi-supervised learning framework. We also introduce Hybrid Discriminant Analysis by combining PCA and LDA to microarray analysis. Our experiments on various applications show that these methods can efficiently alleviate the semantic gap, small sample size and high dimensionality problems.