An SIMD-MIMD architecture for image processing and pattern recognition

The author reports the study of architectures for applications in which the nature of the computing evolves from fine grain parallelism (massive parallel SIMD c.q. highly pipelined computing) to coarse grain computing (MIMD computing). Although this phenomena is apparent in many scientific computing problems, it is most evident in image processing, where the computing activity evolves from global filtering over the image, to operations only involving the structures in the image-objects, object features-, to operations only involving the measured attributes of the structures and their manual relations. This can be considered as an evolution of abstraction level, whereas each abstraction level works on different data-structures and has another granularity. Not to be confused with multi-resolution processing (pyramidal processing, e.g., split and merge techniques), where the processing takes place on another resolution of similar data in a similar data structure. Quite common are algorithms that use feedback from higher abstraction levels to lower abstraction levels, to adapt and control the performance of the lower levels. The author reports the investigation on new computational models and principles of computer architectures that smoothly support efficient processing on each of the granularity levels of an application's task. It proposes a mixed SIMD/MIMD architecture with an emphasis on the interface between these two paradigms. The use of the architecture is elucidated using the distance transform, a path search algorithm and region growing as application examples.