up previous
Up: Survey of Current Techniques Previous: Three-dimensional Bayesian analysis

Summary

Image understanding is a very diverse and active field of research. Even an exhaustive review of subtopics such as two-dimensional segmentation could fill several volumes. Hence, there are many different approaches one could use to solve any given image understanding task. Even after trading off complexity, generality, cost, and a host of other factors, there would still be many approaches worth pursuing. In this chapter, I have described several successful approaches used to solve various problems dealing with segmentation, shape representation, and image understanding. I have shown how computational operators such as thresholding, edge detection, and ridge detection are used extensively in low-level image processing tasks. Snakes and active contour models have the potential for handling more difficult image understanding tasks that require a greater degree of autonomy. Still, higher-level models such as those proposed by Binford and Levitt are required to perform complex three-dimensional object recognition. In summary, effective image understanding requires tight coupling between high-level, hierarchical, task-dependent models and low-level computational operators. The low-level operators are responsible for generating context-dependent data under the guidance of the high-level models. The high-level agents can synthesize the low-level data in a structured manner to make decisions about how to track boundaries, how to extract regions, how to aggregrate information, and to how determine when an object has been recognized correctly.



Ramani Pichumani
Mon Jul 7 10:34:23 PDT 1997