A multilevel Bayesian network approach to image sensor fusion

Abstract

"Automatic main subject detection" refers to the problem of determining salient or interesting regions in an image. We propose the use of a Bayesian network-based approach to solving this problem in the unconstrained domain of consumer photographic images. Various image sensors, derived from the classical computer vision literature as well as other sources, can provide evidence about main subject regions in images. A traditional sensor fusion scheme, such as a Kalman filter, fuzzy logic or simple Bayesian estimation, does not provide sufficient expressive power to capture the uncertainties and dependencies exhibited by such a system. We present a multi-level Bayesian network that accurately models the system and allows for sensor integration in an evidential framework. The multi-level Bayesian network performs better than a simple single-level Bayesian network at accurately combining various image sensor data to construct a belief map identifying the main subject regions in the image. A subsequent study also shows that the multi-level Bayesian network performs better than a linear classification scheme, as well as one based on neural networks.