Statistical learning and data mining for visual information retrieval

In recent years, the world has seen an exponential increase in the amount of digital multimedia content, in both everyday life and scientific exploration processes. This trend necessitates the development of new techniques for efficient and effective management of multimedia data. In this endeavor, statistical learning and data mining play a pivotal role. In this thesis, we report advancements in addressing three major challenges in visual information retrieval: query concept learning, image similarity measuring, and image semantics mapping. 1. To learn a query concept quickly and accurately in high-dimensional feature spaces with a limited number of training instances, we designed the maximized expected generalization algorithm (MEGA). MEGA applies the active learning method to accelerate the visual information search. MEGA selects samples which can maximize the information gain in each iteration of the relevance feedback process. MEGA can reduce the number of samples needed for learning a user's subjective query concept. Furthermore, MEGA has the capability to learn from the negative examples. Thus MEGA can initiate a visual information search without requiring good examples beforehand. 2. To accurately measure the perceptual similarity between images; we carried out extensive data mining research on a 1.5-million image database. Based on our findings, bolstered by the similarity theory of perception in the cognitive science, we derived the dynamic partial function (DPF) for characterizing the similarity between images. DPF pairwisely selects a subset of features to measure the similarity between images. By dynamically considering a subset of relevant features, DPF can cluster perceptually similar images more compactly in the feature space, and at the same time keep the dissimilar images away. 3. To better map the low-level image features to high-level semantics; we proposed the Confidence-based Dynamic Ensemble method. The dynamic ensemble method is a generic method to improve classification accuracy for multi-class classification. This method first assesses the confidence of each classification. For low-confidence classifications; the method dynamically resamples a set of relevant classes as training data for better classification. (Abstract shortened by UMI.)