An Experiment in Model-Based Boundary Detection

This paper describes a simple model-based approach to boundary detection for blob shaped objects. A geometric model is used to focus attention on relevant portions of the image during boundary generation. The boundary generated is that which is a globally optimal match to the shape and edge description parameters of the model. The method is illustrated using images of stained transverse sections of skeletal muscle and the adaptability of the model is demonstrated using images of cells of the corneal endothelium which have poor grey-level contrast. The task of image interpretation involves the combination of both low-level and high-level processes. Low-level processes, such as edge detection, are primarily numerical and deal directly with data in image arrays. High-level processes are more symbolic in nature and deal with abstracted image features. Models of expected image structure have been proposed as a means of providing a framework within which high-level and low-level processing can be integrated. However, a truly general approach has yet to be reported and many issues remain to be investigated. One of the problems which arise from the use of models is the computational complexity of the matching process11. Even relatively simple model-matching tasks may be of sufficient complexity to strain the performance of any presently conceivable machine10. One way of coping with this problem is to use cues to focus attention on likely interpretations, thus avoiding computational effort being wasted on the more unlikely ones. Cues can be thought of as evidence of structure which may be used in conjunction with models to generate hypotheses about the presence of other structures in the image. In the earliest stages of image interpretatio