Mapping image processing algorithms onto abstract multiprocessors

The need for parallel computing technology is rapidly growing in several image processing applications, such as industrial quality control, bio-medical imaging, traffic control automation. Most of the image processing algorithms are inherently computationally intensive and may require vast computing power if strict time-constraints are posed. The spreading of parallel image processing techniques and systems has been driven not only by the afore-mentioned need, but the inherent parallel nature of many image processing algorithms has also eased this evolution. The execution characteristics of a certain parallel algorithm on a given architecture heavily depends on the `mutual conformance' of the mentioned algorithm and the architecture pair. Two algorithms with similar sequential performance may behave very differently in a parallel environment. In sequential algorithms the complexity is expressed in terms of operations and storage. In parallel environments these terms are not adequate for characterizing the computing efficiency - fewer operations does not directly mean shorter execution time since there is a definite overhead involved due to availability of resources and communication between processors