Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 51

industrial applications. This is by no means intended to be an exhaustive pre-
sentation of neural network applications, which would require several topics.
The point here is to show some typical examples, and to stress the reason why
neural networks made important, possibly decisive, contributions.

1.4.2 An Application in Pattern Recognition: The Automatic
Reading of Zip Codes

Character recognition is definitely the application area where neural networks
made their first significant contributions to engineering, proving to be reliable
alternatives to classical pattern recognition methods. In the present section,
some examples and results will be described, relying on the elements of theory
and methodology provided in the previous section.
The automatic reading of postal codes is probably one of the most widely
investigated problems in picture recognition. The automatic reading of printed
envelopes and parcels is a relatively simple problem; by contrast, the huge
variety of handwritings made the recognition of handwritten addresses a truly
challenging problem. For each item handled by the postal service, a machine
must either recognize the code, or resort to human inspection when it fails to
identify the code. As indicated above, correcting a sorting mistake made by
a machine is more costly than resorting to human inspection for reading and
typing in the correct code; therefore, the most frequently used performance
criterion for such machines is the following: given a maximum misclassification
rate (say 1%), what fraction of the mail must be read by a human operator? At
present, typical performances are 5% rejection rate for 1% misclassification.
The development of automatic zip code reading was primarily spurred by
the industrial importance of the problem, but also by the fact that, as early
as 1990, large-scale data bases were made available to the general public by
the United States Postal Service (USPS), and later by the National Institute
of Science and Technology (NIST). That policy allowed many laboratories,
both in industry and in universities, to improve the state of the art, and
to validate, in a statistically significant way, the methods and procedures
that they developed; it had a general positive impact on the development of
powerful classification methods.
Figure 1.32 displays some examples from the USPS database, which fea-
tures 9,000 digits (which is not a very large number, considering the variety
of handwriting styles). The di culty of the problem is immediately apparent.
Consider the postal code in the upper right corner of the picture; one reads
68544 effortlessly, but one notes

• that the digit 6 is split into two parts,

• that digits 8 and 5 are linked together,

• that digit 5 is split into two parts, the right part being linked to digit 4!

Thus, if one decides to base the recognition of the code on the recognition of its
individual digits, the problem of segmentation must be solved first: how does