An interactive system for mental face retrieval
Yuchun Fang, Donald Geman, Nozha Boujemaa · 2005
We propose a system to "retrieve" the mental image of a face from a large database using Bayesian inference and relevance feedback. Since the "target image" exists only in the mind of the user, mental image retrieval differs sharply from standard, example-based retrieval and has not been widely studied. In designing the relevance feedback engine, we adopt probabilistic models for the display and answer processes. The answer model is designed to capture properties of human cognition in choosing among displayed faces. The images in each display are selected according to heuristics inspired by maximizing the conditional mutual information between the answer and the target given the previous feedback. Simulations and real tests validate show that the relevance feedback engine operates in real-time and locates the target in a reasonable number of displays.