Abstract
Relevance feedback has become a key component in CBIR system. Although most current relevance feedback approaches are based on dichotomous relevance measurement, this coarse measurement is a distortion of the reality. We study relevance feedback with multi-level relevance measurement to better identify the user needs and preferences. To validate the use of multi-level relevance measurement and our relevance feedback algorithm, we developed a CBIR prototype system - WillHunter. There are two novelties in our system, one is our SVM-based fast learning algorithm; another is the easy-to-use graphical user interface, especially the relevance-measuring instrument. Not only experiments are conducted to assess the algorithm, but also usability study is carried out to evaluate the user interface.