A versatile parallel computer architecture for machine vision

In order to achieve application dependent "real-time" performance, it is necessary with the technology available today, to heavily use parallelism for image processing and some higher level treatments. Specifically, the hardware has to be matched to the type of processing to be run on it. This leads to the design and implementation of heterogeneous pipelined structures, due to the various processing levels usually implied by an application. Then, the synchronized use of such systems calls for a carefully designed programming environment, facilitating the implementation of the described application on the system. Two distinct approaches have been used for the design of image processing systems. The first is based on a modular functional point of view, whereas the second one aims to implement a reconfigurable, flexible and fully programmable structure. Both approaches led to the realization of systems able to carry out most of the low level image processings in video real-time. Developme nt of methods and design of new architectures call thus for careful studies and possibly experimentation. This should be facilitated by the set-up of development centers grouping hardware and software resources. A possible configuration of such an integrated development system for computer vision applications will be indicated, together with some representative applications.