Relevance-Feedback Image Retrieval Based on Multiple-Instance Learning

This paper describes the development of a novel content-based image retrieval system using Multiple-Instance Learning (MIL). MIL is designed for learning on bags, each composed of a number of instances (i.e., feature vectors). For a given bag, one or more instances may be responsible for the observed classification of the bag, but their identities are unknown. What we can observe is only the label of the bag and our aim is to predict the label of any given new bag. The image retrieval problem can be mapped to an MIL problem. Our contribution is a new way to improve the effectiveness of MIL in image retrieval. We have implemented a Web-based relevance-feedback image search system to illustrate the proposed idea, which shows that the search accuracy is encouraging.