Bayesian evidence combination for region labeling

摘要

My thesis is that a Bayesian evidence fusion framework for high-level vision has quantifiable advantages over other popular methods such as Neural Nets. Bayesian networks allow for the development of a probabilistic evidence fusion scheme for combining evidences from multiple information sources. The combined information can then be used to solve a variety of problems across multiple domains. In this thesis, I present a Bayesian evidence fusion framework for solving region labeling problems. Evidence regarding the label for a region can come from various sources. These sources can be in different modalities of measurement (for example, images, audio streams, and geometric coordinates). In order to perform effective evidence fusion, we need to accept this multi-modal evidence in a probabilistic framework. Bayesian networks provide a suitable framework for such evidence combination. I present details of knowledge engineering techniques designed to create the Bayesian network structures for evidence fusion, as well as the training methodologies. The theoretical framework is then applied to solve various region labeling problems. Automatic main subject detection provides a measure of saliency for regions or objects in an image. A large number of feature detectors, some based on the physics of the camera, and others based on computer vision and image understanding techniques, can provide evidence for main subject region in a consumer photographic image. The Bayesian evidence fusion framework allows us to construct a network that quickly combines this evidence to produce a belief map that identifies main subject regions in the image. Similarly, the problem of map building for autonomous mobile robot navigation is one of region labeling. Here, we want to assign each region of the robot's map with the labels occupied or empty. I have developed a simulator that uses Bayesian networks for evidence fusion to generate the map. This map is then used for autonomous robot navigation in the simulated environment. The thesis is validated by a benchmarking study that compares the Bayesian evidence fusion framework with popular information fusion and classification methodologies such as neural networks and statistical predictors.