In this introduction to a special issue of Computer, the authors characterize the importance of parallel computing to image processing in computer vision, one of the “grand challenges” in the US government’s initiative on high-performance computing and communication. Although much progress has been made in parallel processing in recent years, the “immense computational challenge presented by vision is still to be met.” The reasons include our lack of clear understanding of both how to integrate algorithms from diverse areas, as required by typical vision systems, and the vision process itself.
The current approach to classifying the processing needs for vision computation into low, intermediate, and high levels is profiled. Efforts in architectural design and development have resorted to embedding components from each of the three levels into one integrated system. Comparing the results with what has been accomplished in general-purpose parallel processing for other scientific disciplines, the authors conclude that architectures for vision systems are in their infancy.
The paper ends with an outline of issues that must be addressed in the future: creating suitable architectures, system integration to provide transparency to the user, powerful and flexible programming models, user-friendly software tools, and real-time vision applications. The paper is a brief and readable portrait of the current state of parallel processing support for computer vision. It gives a useful overview for anyone interested in this subject.