Discriminative learning with application to interactive facial image retrieval

The amount of digital images is growing drastically and advanced tools for searching in large image collections are therefore becoming urgently needed. Content-based image retrieval is advantageous for such a task in terms of automatic feature extraction and indexing without human labor and subjectivity in image annotations. The semantic gap between high-level semantics and low-level visual features can be reduced by the relevance feedback technique. However, most existing interactive content-based image retrieval (ICBIR) systems require a substantial amount of human evaluation labor, which leads to the evaluation fatigue problem that heavily restricts the application of ICBIR. In this thesis a solution based on discriminative learning is presented. It extends an existing ICBIR system, PicSOM, towards practical applications. The enhanced ICBIR system allows users to input partial relevance which includes not only relevance extent but also relevance reason. A multi-phase retrieval with partial relevance can adapt to the user's searching intention in a from-coarse-to-fine manner. The retrieval performance can be improved by employing supervised learning as a preprocessing step before unsupervised content-based indexing. In this work, Parzen Discriminant Analysis (PDA) is proposed to extract discriminative components from images. PDA regularizes the Informative Discriminant Analysis (IDA) objective with a greatly accelerated optimization algorithm. Moreover, discriminative Self-Organizing Maps trained with resulting features can easily handle fuzzy categorizations. The proposed techniques have been applied to interactive facial image retrieval. Both a query example and a benchmark simulation study are presented, which indicate that the first image depicting the target subject can be retrieved in a small number of rounds