The Role Learning in Perceptual Organization of Complex Images

点击翻译按钮获取中文摘要

Abstract IIS-9907141Sarkar, SudeepUniversity of South Florida$73,407 - 12 mos.The Role of Learning in Perceptual Organization of Complex ImagesThis is the first year funding of a three year continuing award. Perceptual organization is the ability to efficiently group low-levelprimitives, such as pixels or edge segments or regions, so as to formobject hypotheses with minimal domain knowledge. This process is anessential link between the noisy low-level segmentation processes andthe sophisticated, computationally expensive, object recognitionprocesses. This work will address three major scientific issuesrelated to perceptual organization in computer vision: (i) How toefficiently form large groups of extended low-level primitives byassembling small groups of features (grouplets)? The concept of graphspectra is an ideal mechanism for this purpose. (ii) How to encodethe dependency of grouping parameters on image statistics? Jointprobability distributions encoded as Bayesian networks offer anefficient abstraction. (iii) What is the role of learning inperceptual organization? The specific roles of learning explored wouldbe threefold: (a) learning the image to grouping parameter mapping,(b) learning the forms of the groups, and the associated hierarchy,from object models, and (c) learning the primitive type, such ascorners, textured-regions, or edges, that is appropriate for a domain.