The principal theme of this research is the development and application of deformable models for computer vision. In contrast to traditional, purely geometric object models, deformable models simulate elastic material which change shape in response to externally applied forces. The main advantage of this approach is that it produces coherent object models where low level vision techniques often suffer from pixel artifacts and may not yield closed object boundaries. The work begins by extending existing surface-based models in two fundamental ways. First, models are developed which can dynamically change to reflect the underlying topology of the object being modeled. This makes it possible to model object deformations with less complete and prior knowledge. Second, the deformable models are generalized so objects in higher dimensions can be modeled. This will enable scientists to study object behavior in time-sequence images and gain an understanding of the mechanisms involved in object deformation. In order to accurately model complex objects in nature, computational methods are needed that are capable of handling models with hundreds of thousands of elements. The approach combines two optimization techniques, simulated annealing and multiresolution analysis, to reduce both the time and space complexity of fitting deformable object models. To evaluate effectiveness of the deformable models, the object modeling techniques are applied to several challenging biomedical computer vision problems: image segmentation; shape description; and object recognition.
