Content-based visualization and retrieval for image libraries

This thesis investigates the content-based image retrieval (CBIR) systems and content-based visualization and layouts for image libraries.; This thesis presents a powerful enhancement of recent CBIR techniques which integrates spatial and local feature information and combines learning and vision techniques for interactive image retrieval and visualization. Our approaches address the fundamental problems of content-based image retrieval from low-level feature representation to mid-level on-line learning, and to high-level user modeling. For high-level user modeling, a user-centric system is proposed for optimized visualization and layouts for content-based image retrieval. A novel subspace feature-weighting scheme is proposed to modify 2-D layouts in a variety of content-dependent ways. For mid-level on-line learning, discriminant analysis in the expectation-maximization (EM) framework and support vector machines (SVM) are both applied to achieve an improved image classification. For low-level feature representation, (1) a novel wavelet-based salient-point detector is presented and evaluated for image retrieval system; (2) a user-defined region-of-interest (ROI) is proposed to integrate spatial and local information; (3) a novel feature dimensionality reduction scheme is proposed using principal feature analysis; and (4) scale invariance evaluation of visual features is conducted.; Monte Carlo simulations with machine-generated layouts, pilot user studies, and experimental retrieval results from large image databases have demonstrated the capability of the proposed approaches for image retrieval applications.