Current trends in image segmentation indicate that in order to achieve connsistently good results, some form of domain knowledge must be used. Knowledge-based image-segmentation systems traditionally use models of objects and their interrelationships to guide or postprocess the low-level segmentation operation. A major shortcoming of that appproach is the size and complexity of the knowledge base needed for all but trivial domains. The knowledge is often quite difficult to embed in the system and usually is not transferred easily to other domains. Also, processing time can be large. The goal of the proposed work is to explore the use of a small set of domain constraints, namely lighting, camera, and general object surface characteristics, to produce acceptable segmentation results without the need to classify objects beforehand. As images can be constructed graphically using only surface orientation, color, and reflectivity, along with the position and type of lighting, it is hypothesized that general domain information will often be sufficient to perform adequate segmentation. The proposed system will use this knowledge to drive a low-level expert segmentation system, determining parameters, selecting strategies, and resolving ambiguous interpretations.
