The choice of the proper shape recognition method is always a compromise between recognition power and computational complexity. Speed requirements of real-time applications often limit the number of possible techniques. In previous chapter the recognition powers of three shape coding techniques were demonstrated. The pairwise geometric histogram and the set of five simple shape descriptors performed well, but the chain code histogram had some limitations. The CCH cannot preserve information on the exact shape of a contour, because it only shows the probabilities (or frequencies) for the different directions present in a contour. Thus, there may be many objects with the same CCH. However, the chain code histogram is fast to calculate and it needs only a small amount of memory which can be crucial in many real-time applications. The CCH is evidently good enough for many applications, especially for those having relatively distinct classes.
The pairwise geometric histogram is computationally heavy and it requires more memory than the CCH. If objects are non-polygonal, the polygonal approximation must be made which requires more time. The PGH is suitable for problems that contain (nearly) polygonal objects. In addition, these objects can also be partially occluded.
The set of five simple shape descriptors fall in between the CCH and the PGH in both time and memory requirements. They seem to provide consistent recognition which agrees well with that of a human observer. Some of these descriptors are quite correlated [ 15 ], so only a subset of them may be sufficient in most applications.
The self-organizing map (SOM) is an ideal tool in exploring the natural structure of high-dimensional data. It makes a topology-preserving mapping from data space to 2D space. Practically all statistically significant information is preserved in the mapping. The SOM can discover features that are not explicitly present in single measurements. For example, the contours in Figure 5 (c), which were ordered by the SOM according to five descriptors, seem to have roughly three dimensions, elongation, smoothness, and bending, none of which were obtainable using only one of the original descriptors.
Jukka Iivarinen