Applying Multiple-Instance Learning to Content-Based Image Retrieval

Washington University, Saint Louis MO

Investigators

Abstract

Applying Multiple-Instance Learning to Content-Based Image Retrieval With the development of the web, there has been an explosion in the volume of both textual and digital media. Effective use of that information requires efficient technology for locating relevant material. Algorithms for textual search, such as string matching for keyword searches, are well understood. However, effective techniques to retrieve images based on semantic content need to be developed. Manual keyword annotation is not feasible due to the size of the image repositories, the rich contents of the images, and the subjectivity of human perception. This project applies new segmentation methods and multiple-instance learning, a new machine learning technique, to the image search problem. The research activities promise to enable effective image search in which a user provide an example of a desired image and indicates which of a small number of candidate images are desirable. Unlike existing content-based image retrieval (CBIR) systems, this project uses multiple-instance learning to automatically determine which portion(s) of the image are important to the user. Just as textual search technology has revolutionized the way people search for textual information, efficient image retrieval based on semantic content has the potential for large productivity gains in the way people work with images, including better utilization of the increasing volume of information available as images, with accompanying economic benefits. Undergraduate and graduate students will receive training through independent research experiences, classroom presentations, and course final projects. The PI will encourage more women to study computer science through interactions with high school women through the outreach activities of Washington University's Society of Women Engineers (SWE) and through interaction with college women through mentoring activities of SWE. Dissemination of the research results will be through conference presentations and papers, journal publications. Also papers and software will made available at www.cs.wustl.edu/~sg.