Parallel vision techniques on the hypercube computer

Abstract

Parallel algorithms for programming low-level vision mechanisms on the JPL-Caltech hypercube are reported. These concern principally edge and region finding. 256x256 8bit images were used. We discuss the problem of programming a hypercube computer, and the Caltech approach to load balancing. We then discuss the distribution of images over the hypercube and the I/O problem for images. In edge finding, we programmed convolution using a separable kernel computational approach. This was tested with 5x5 and 32x32 masks. In region finding, we developed two different parallel histogram techniques. The first finds a global histogram for the image by a completely parallel technique. This method, which was developed from the Fox-Furmanski scalar product method, allows each histogram bucket to be computed by a separate processor, each processor regarding the hypercube as a different tree, and all buckets being computed in parallel by a complete interleaving of all communications required. Similarly the global histogram can then be distributed over the hypercube, so that all processors have the entire global histogram, by an completely parallel technique. The second histogramming method finds a spatially local histogram within each processor and then connects locally found regions together. Work in progress includes the application of a Hopfield neural net approach to region finding.

Additional details