User interface techniques for browsing and searching image databases

The number of digital images is rapidly growing. We need efficient browsing and searching tools to manage the large image databases. Content-based image retrieval was proposed to address this problem; it requires automatic extraction of visual features from images. Unfortunately computers can only extract low-level visual features without supervised help by humans. It is generally accepted that effective browsing requires semantic information. The information relates to "what is in the picture." For example, low-level visual feature extracted from a picture of a bear can be its color or texture. The semantic information is the fact that the picture contains a bear. Human relevance feedback approaches were proposed to bridge the gap between high-level semantic concepts and low-level visual features. Since in current technology, only the human can understand the semantic concept of images. Thus, user interaction is essential for image retrieval systems.; In our work we develop several user interface techniques for browsing and searching image databases. A review of existing image retrieval systems that use relevance feedback techniques indicates that they focus on image retrieval performance, but not on the user interface needed for relevance feedback. We propose an user interface for relevance feedback and semantic clustering. Our proposed user interface makes it easy to browse retrieved images, and easy to create semantic clustering. Previously proposed relevance feedback technique use information obtained from the most current session, but do not use information from past sessions. We show that information from past sessions can significantly improve the retrieval performance. Two methods that accumulate information from relevance feedback are proposed. The system accumulates information from multiple relevance feedback sessions, and uses the information to improve image retrieval performance and image visualization. The approach requires a mathematical tool that enables effective separation of classes by discriminant subspaces. We investigate several proposed algorithms to compute the discriminant subspaces, and determine that the Null Space Method is the most effective. The current algorithm for this method is time consuming. We develop a method that can compute the discriminant subspaces very fast. Finally we propose a practical technique for interactive browsing of visual content.