Learning Probability Models for Surface Appearance and Shape by Minimax Entropy Principle

点击翻译按钮获取中文摘要

Abstract IIS-9877127Zhu, Song-ChunOhio State University$115,872 - 12 mos.Learning Probability Models for Surface Appearance and Shape by Minimax Entropy PrincipleThis is the first year funding of a three year continuing award. The research in this project is aimed at three important problems in visual learning. 1). Studying a mathematical theory for visual learning. 2). Learning realistic probability models for textures, object shapes, clutter, and texton flows, which are essential constructing components of natural scenes. 3). Developing fast algorithms for model inference and model verification. Central to the research is the minimax entropy learning theory, proposed by the PI, the co-PI, and Mumford in 1997. The minimax entropy learning theory is capable of exploring the intrinsic structures (or regularities) underlying the observed imagery and patterns. The outcomes of this learning theory are statistical models of the visual patterns for textures, object shapes, clutter, and texton flows. The learned models can be used in the Bayesian framework as likelihood functions and prior probabilities, which are essential for the success of Bayesian image analysis paradigm in applications such as image segmentation, perceptual organization, and object recognition. The PI and Co-PI have achieved good preliminary results, and together, they form an effective interdisciplinary team to conduct the research.