Image situated agents provide a suitable framework to implement locally cooperative adapted strategies for image segmentation. They also render easier a priori knowledge integration that provides new constraints that are required for all image processing steps (from segmentation to interpretation). We present a conceptual framework for the software architecture of a low level vision system. The latest, which is based on image situated agents, is analyzed and described in three steps : 1. A global and structural description of the agents' organization. In this description step, we establish the links between agents. We propose the use of the irregular pyramid that imposes its structure to agents population in order to guaranty a global tractable and convergent behavior. 2. A local functional and behavioral agent's description. We propose a specific implementation of our agents software architecture. In the latest, two types of agents that represent region or edge image primitives are locally interacting inside the pyramid. Our goal is to show how our methodology allows a rich flexible and distributed implementation of aspects such as model integration, and -region/region, edge/region- cooperative strategies. 3. Finally, a global comparative and functional analysis checks that the set of local interactions produces a good image segmentation. Comparative evaluation is performed on medical and synthetic images.
