Effective use of SIMD parallelism in low- and intermediate-level vision

The authors consider two examples from image understanding-focus-of-attention vision and contour image analysis-and present new parallel-processing methods that effectively support these types of computations. The research is a blend of theory and practice. On the one hand, the aim is to develop algorithms whose properties are well understood and can be formally related to key aspects of machine models. On the other hand, it is desired that the algorithms be easy to implement and practical in terms of their actual processing times on existing parallel machines. The experimental research was conducted on a 16384-processor Connection Machine CM2, and results of algorithm implementations on that machine are presented.<<ETX>>