Viscom: an orthogonal multiprocessor for early vision and neural computing

A categorization of important image/vision algorithms that can be efficiently implemented on a research multiprocessor called Viscom is presented. Extended C-language constructs are developed and used to specify orthogonal multiprocessing algorithms. Several early vision algorithms are illustrated, including convolution, optical flow, and 2D transforms. Motion analysis is modeled by an artificial neural network, which can also be efficiently mapped onto Viscom. Advantages of using Viscom for early vision and neural computing are discussed, and the associated hardware/software development experiences are reported.<<ETX>>