Scale-based integrated microscopic computer vision techniques for micromanipulation and microassembly

Micromanipulation and microassembly are important emerging techniques developed mainly over the past decade. Microscopic computer vision is of fundamental importance to micromanipulation and microassembly in providing low-level noncontact feedback of microscale geometry, spatial relations, and motion and high-level task understanding. Because of the unique properties of microscope optics, microscopic computer vision differs significantly from macroscale computer vision. Based on these properties, this thesis develops a scale-based theoretical model for microscopic computer vision and several related techniques that are crucial to micromanipulation and microassembly. The static properties of microscopic optics are analyzed. A theoretical model based on the scale-space theory is developed to characterize the dynamic behaviors of microscopic images under depth changes. This model makes it possible to directly apply the mathematical concepts and techniques of the scale-space theory, which was originally developed for front-end vision and multi-scale image analysis, to microscopic computer vision. Based on this model, a class of new wavelet-based focus measures is constructed, which provides significantly better depth resolution and selectivity than previously reported spatial domain operators. It also provides a necessary tool for the subsequent development of microscopic computer vision techniques. Based on the theoretical model and wavelet focus measures, techniques for the auto-focusing, segmentation, and tracking of microscopic images are developed. Firstly, the concept and technique of selective focusing are developed to answer the basic question of how to focus on three-dimensional objects. Secondly, based on perceptual organization and graph partitioning, an unsupervised microscopic image segmentation technique is developed. This provides a unified segmentation approach for handling partially defocused images and does not rely on the extraction and connection of local image features such as edges. Thirdly, visual tracking techniques for microscope images that remain robust under significant depth changes are developed. These techniques are all closely related because of the unique properties of microscope optics. Microscale force/vision integration is proposed under the general framework of multisensor fusion as an essential approach to overcome the basic limitations of microscopic computer vision. The application of this integration in micromanipulation contact transition control is presented.