Probabilistic hierarchical image models

This thesis studies the probabilistic modeling of images. Two topics are covered. One is the theoretical justification for efficient image representation; the other is building a composition machine to interpret real images.; Chapter 1 overviews the major methodologies developed in the field of computer vision and chapter 2 briefly introduces the methodology we adopt.; Chapter 3 focuses on object/scene representation. It explores theoretical supporting evidence for hierarchical representation against holistic representation through a statistical testing argument. In conclusion, a composition system, with a built-in hierarchical structure, has provably better performance.; Chapter 4 focuses on implementation of a scene interpretation system (with application to license detection), which is capable of parsing a scene or accurately detecting an uncertain number of targets with cluttered background. Under the Bayesian framework, all the components (i.e., prior model; data model and posterior model) and their implementations are discussed in detail. Demonstration of the system spells out the important aspects of our methodology; it also provides empirical supporting evidence, which is consistent with the theoretical evidence covered in chapter 3.; Chapter 5 makes a conclusion and touches on some future directions.