Machine Vision and Applications (1995) 8:85-100 Machine Vision and Applications 9 Springer-Veflag 1995 Proteus: a reconfigurable computational network for computer vision Robert M. Haralick, Arun K. Somani, Craig Wittenbrink, Robert Johnson, Kenneth Cooper, Linda G. Shapiro, Ihsin I". Phillips, Jenq-Neng Hwang, William Cheung, Yung Hsi Yao, Chung-Ho Chen, Larry Yang, Brian Daugherty, Bob Lorbeski, Kent Loving, Tom Miller, Larye Parkins, Steve Soos Department of Electrical Engineering, ~T-10, Applied Physics Laboratory, HN-10, University of Washington, Seattle, Washington 98195, USA Abstract. The Proteus architecture is a highly parallel, multi- ple instruction, multiple data machine (MIMD) optimized for large granularity tasks such as machine vision and image pro- cessing. The system can achieve 20 gigaflops (80 gigaflops peak). It accepts data via multiple serial links at a rate of up to 640 MB/s. The system employs a hierarchical recon- figurable interconnection network with the highest level being a circuit-switched enhanced hypercube, serial interconnection network for internal data transfers. The system is designed to use 256 to 1024 RISC processors. The processors use 1-MB external read~write allocating caches for reduced multipro- cessor contention. The system detects, locates, and replaces faulty subsystems using redundant hardware to facilitate fault tolerance. The parallelism is directly controllable through an advanced software system for partitioning, scheduling, and development. System software includes a translator for the INSIGHT language, a parallel debugger, low- and high-level simulators, and a message-passing system for all control needs. Image-processing application software includes a variety of point operators, neighborhood operators, convolution, and the mathematical morphology operations of binary and gray-scale dilation, erosion, opening, and closing. Key words: Proteus - INSIGHT - Reconfigurable compu- tational network - MIMD - Enhanced hypercube - Parallel computing 1 Introduction A variety of parallel architectures can be used for image pro- cessing applications. To cut short the review of these architec- ture types, we point out some political parallels that have been informally made by some noted researchers. The array can be likened to a Fascist dictator leading a march. The pipeline can be likened to a capitalist assembly line. The pyramid can be likened to the cell block hierarchy of Communist dictatorships; the multiprocessor systems, to parliamentary committees at Correspondence to: R.M. Haralick work, and the network, to political anarchy. And the recon- figurable network just cannot get its act together. Particular examples of these architectures include processor arrays such as Illiac IV, CLIP4, CLIP7, DAP, MPP, and GAPE systolic ar- rays such as WARP, pipelines such as Cytocomputer, Genesis, and VAP and multiprocess or systems such as PASM, POLYP, ZMOB, GOP, PICAP, TOSPICS, DIR FLIP, PM4, and pyra- mid systems. Scholarly reviews of these architectures can be found in the papers discussed in [6]. The purpose of this paper is to discuss an architecture in which the act of the reconfigurable network can be put together for computer vision in a large-grain parallelism mode. We do this in the context of knowing that reconfigurable networks for image processing have not been used and that there has been little discussion of architectures that are simultaneously suitable for low-level image processing, high-level computer vision, and the computational tasks required to direct robots and material handlers. All this must be done in near real time in the factory setting. Our viewpoint differs from the usual discussions of com- puter architecture. We believe that all paradigms that map image data onto processors like the arrays and pyramids, or paradigms that map specialized tasks onto various processors (such as the multiprocessor systems), must, of necessity, cre- me specialized and inflexible systems. We believe that the watchword of computer vision is flexibility. There must be the image-in and image-out operations of low-level vision. There must be the image-in and data structures-out operations of mi- dlevel vision. There must be the data structures-in and data structures-out operations of high-level vision. The purpose of any computer vision system is to be economically useful in the factories of the society in which it functions. Therefore, there must be the capability for performing numerical calcu- lations, data formatting operations, communication reporting operations, and real-time control of external devices such as material handlers and robots. This suggests that the approach must be integrated. To approach the design, we must step back and understand that the low-level neighborhood operators dis- cussed in today's archival literature can be much more com- plex than the Roberts and Sobel variety. We must understand