Automatic compilation of parallelism in visual object recognition

Abstract

The paper presents the results of investigations into techniques for automatically generating concurrent implementations of visual object recognition systems from high-level visual domain descriptions. Visual knowledge can be represented using 'schemata'--object-centred recognition models that intentionally represent classes of related objects. Schemata form hierarchies based on relationships such as composition and function. Schemabased representations contain implicit concurrencies that can be exploited by a recognition system called a "recognition network" The paper presents the design of a simple schema representation language, and shows how it can be automatically transformed into concurrent code. This is done by building a compiler that takes schemata as input and produces a recognition network, in OCCAM, as output. The design of the compiler involves several novel features. The existence of large transputer arrays makes it possible to execute the resulting code with high physical, as well as conceptual, parallelism.

DOI: 10.1016/0141-9331(89)90122-1

7 Figures and Tables

Cite this paper

@article{Shoom1989AutomaticCO, title={Automatic compilation of parallelism in visual object recognition}, author={Greg B. Shoom and David B. Skillicorn}, journal={Microprocessors and Microsystems - Embedded Hardware Design}, year={1989}, volume={13}, pages={165-172} }