Image Understanding Algorithms on Fine-Grained Tree-Structured SIMD Machines

Ibrahim, Hussein A. H.

An Important goal for researchers In computer vision is the construction vision systems that Interpret Image data in real time. Such systems typically require a large amount of computation for processing raw Image data at the lowest level, and for sophisticated decision making at the highest level Recent advances In VLSI circuitry· have led to several proposals for parallel architectures for computer vision systems. In this theSIS. we demonstrate that fine-grained tree-structured SIMD machines, which have favorable characteristics for efficient VLSI Implementation, can be used for the rapid execution of a wide range of Image understanding tasks We also Identify the limitations of these architectures and propose methods to ameliorate these difficulties. The NON-VON supercomputer, currently being constructed at Columbia University, is an example of such an architecture. The major contribution of this thesis IS the development and analysis of several parallel Image understanding algorithms for the class of architectures under consideration The algorithms developed In this research have been selected to span different levels of computer vision tasks They Include Image correlation, hlstogrammlng, connected component labeling, the computation of geometric properties, set operations, the Hough transform
method for detecting object boundaries, and the correspondence problem In
moving light display applications. The algorithms Incorporate novel approaches to reduce the effects of communication bottleneck usually associated With tree architectures



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-139A-84
Published Here
August 7, 2013