Statistical Machine Learning for Bridging the Semantic Gap in Image Retrieval

of thesis entitled: Statistical Machine Learning for Bridging the Semantic Gap in Image Retrieval Submitted by HOI, Chu Hong (Steven) With the explosive growth of multimedia data, more and more research attentions have been devoted to visual information retrieval. Image retrieval, particularly content-based image retrieval (CBIR), has been actively studied in multimedia information retrieval community in the past decade. One of the most challenging difficulties in CBIR is the semantic gap between low-level visual features and high-level semantic concepts. This thesis investigates statistical machine learning techniques for attacking the semantic gap problem in image retrieval. In this thesis, a unified learning framework, integrating supervised learning, unsupervised learning, semi-supervised learning, active learning and distance metric learning, is proposed to bridge the semantic gap in image retrieval from several novel perspectives. The first novelty in the proposed framework is the semi-supervised active learning (SSAL) scheme for online learning with users’ relevance feedback in image retrieval. Different from traditional active learning approaches, which are usually supervised, our semi-supervised solution can be much more effective in finding the informative unlabeled data for narrowing down the semantic gap in image retrieval. The second originality is the log-based relevance feedback (LRF) in that users’ log data are engaged in the online learning tasks for image retrieval. Different from traditional relevance feedback, our LRF scheme exploits the users’ log data as a critical resource for bridging