Page 444 - Sensors and Control Systems in Manufacturing

398
Cha p te r
Ei g h t

FIGURE 8.14 Trained image.

be unrecognizable because they are partially or completely occluded.
And, finally, if the objects are defective (Fig. 8.14), the features are
even less predictable and hence harder to find.
Since global features are not computable from a partial view of an
object, recognition systems for these more complex tasks are forced to
work with either local features, such as small holes and corners, or
extended features like a large segment of an object’s boundary. Both
types of feature, when found, provide constraints on the position and
the orientations of their objects. Extended features are in general com-
putationally more expensive to find, but they provide more informa-
tion because they tend to be less ambiguous and more precisely
located.
Given a description of an object in terms of its features, the time
required to match this description with a set of observed features
appears to increase exponentially with the number of features. The
multiplicity of features precludes the straightforward application of
any simple matching technique. Large numbers of features have been
identified by locating a few extended features instead of many local
ones. Even though it costs more to locate extended features, the
reduction in the combinatorial explosion is often worth it. The other
approach is to start by locating just one feature and use it to restrict
the search area for nearby features. Concentrating on one feature
may be risky, but the reduction in the total number of features to be
considered is often worth it. Another approach is to sidestep the