The main purpose of this paper is to introduce, through the case study of a new class of general approach of vision processing, the use of formal techniques to model complex concurrent systems using abstract specification. Moreover, we want to show that it is possible to derive implementations on parallel machines by specification transformation. The implementation is derived from the abstract specification by introducing concrete aspects of the realization, for instance the algorithms used to distribute the global knowledge of time, producing a concrete specification. The CO-OPN (Concurrent Object Oriented Petri Nets) language [4,5] is used to model both specifications, the abstract one and the concrete one. This example proves the wide spectrum of application that can be modelled using CO-OPN. CO-OPN is supported by a complete environement called SANDS (Structured Algebraic Net Development System) including a graphical editor, a compiler, a temporal logic prover and a simulator [5]. The concrete specification is used for producing the code for the parallel machine, in our case a transputer machine. The idea of developing a vision system in a modular fashion has already been raised by many authors [1]. Anatomical, physiological and psychological studies show that despite separation and specialization of visual processing, common underlying principles and mechanisms seem to apply. Moreover if modules are casted from the same mold, reusability can significantly save development time. The goal of this work is not to propose a universal vision system, but rather to investigate promising concepts.