Robotics Reference
In-Depth Information
different people, none of whom is the person whose writing the system
will be attempting to recognize. By storing data from many writers with
many different cursive styles, the system is able to compare any person's
writing with a range of stored samples. The more writers employed in
the data collection process, the more accurate will be the system.
Writer-dependent systems can be even more accurate than writer-
independent systems because they are tasked only with recognizing the
writing of certain people. It is the writing of those people, the “training
database”, which forms the data against which the recognition sample is
compared. The training database in a writer-dependent system is nor-
mally large enough to contain samples of all the important variations
and peculiarities of the styles of the small group of writers for which
it is intended, it is not so large that it attempts to model every possi-
ble variation and peculiarity present across the whole range of natural
handwriting styles. It is because these systems are not so ambitious that
they offer a greater accuracy of recognition—the database against which
a character or a word is being compared will be smaller than the corre-
sponding database for a writer-independent system, and hence there is
less scope for error in the recognition process. But it is possible to have
the best of both worlds, by combining a writer-independent system with
writer-dependent data in order to “learn” the writing of a particular user
and thereby to improve the performance of a system with time.
Recognizing the Edges and Shapes in Images
Much of the early work on computer vision was devoted to the prob-
lems of finding lines and edges in images, and using this information to
identify shapes. The pioneer in this field was Lawrence Roberts who, in
1965, was the first to attempt to solve the task of automatically recog-
nizing three-dimensional objects. Two years earlier, while a graduate stu-
dent at MIT, Roberts had already published an algorithm for generating
a two-dimensional perspective view of a three-dimensional object. His
algorithm reduced the visual information in a scene to distinct solid ob-
jects with straight line edges, which were then further reduced to flat
surfaces that were also defined by straight lines. The locations of the
endpoints of each line were stored in the computer together with data re-
lating to specific viewing locations, the direction of sight from a viewing
location to the object, and the position of a flat surface on which the pro-
gram was required to project the two-dimensional image. Based on this
