Accuracy improving methods for content-based image searches in large multimedia databases

Abstract

The task of content-based image retrieval (CBIR) from a large volume of image database has recently received a great deal of attention from database community. There have been many researches to improve the accuracy and the efficiency of image retrievals. Among various issues in CBIR, the development of an accurate and compact image descriptor is the most fundamental and challenging task. Various compact image description schemes have been proposed, but most existing methods simply aggregate local descriptors into a single vector without any consideration on the importance of local descriptors. In these methods, noisy or less important local descriptors are treated equally to the important local descriptors, and there-by such compact image descriptors suffer from background clutters of an image. In this dissertation, we propose two methods to improve accuracies of content-based image retrieval on a large volume of multimedia database. In the first part of this dissertation, we propose a novel compact image descriptor that can improve the search accuracy significantly through weighted aggregation of local descriptors based on their relative importance in an image. Using saliency analysis of an image, we assign low weights to local descriptors extracted from noisy or less important region (i.e., background) of the image, whereas high weights are assigned to local descriptors extracted from salient objects. By assigning low weights to noisy or less important local descriptors, the proposed compact image descriptor alleviates the adverse effects caused by noisy descriptors, and thereby is very robust to background clutters. In the second part of this dissertation, we propose a novel multimodal query processing method, called image-keyword joint query processing, which integrates the benefits of keyword-based image retrieval with content-based image retrievals. As the two different types of queries are processed individually in the existing approaches, the final result may not contain relevant images to the query at all. To solve the problem, we propose a hybrid index structure that can process both visual and textual information simultaneously.