Browse Prior Art Database

Linear Rejection in Character Recognition Systems

IP.com Disclosure Number: IPCOM000079425D
Original Publication Date: 1973-Jul-01
Included in the Prior Art Database: 2005-Feb-26
Document File: 3 page(s) / 38K

IBM

Chow, CK: AUTHOR

Abstract

In pattern recognition systems such as character readers, for example, an important requirement is the capability of rejecting "bad" characters which could potentially cause substitution errors. There is described herein a class of methods for bad character rejection, which are advantageously suitable for use in pattern recognition systems.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 54% of the total text.

Page 1 of 3

Linear Rejection in Character Recognition Systems

In pattern recognition systems such as character readers, for example, an important requirement is the capability of rejecting "bad" characters which could potentially cause substitution errors. There is described herein a class of methods for bad character rejection, which are advantageously suitable for use in pattern recognition systems.

In a typical known pattern recognition system, the recognition process includes the step of computing for each input pattern vector V, a set of scores T(1)(V), T(2)(V),..., T(NC)(V), one for each character class, wherein the term NC represents the number of character classes. To effect the pattern recognition, the largest of these scores is selected.

For convenience, these scores are arranged in descending order as follows: T(1)(V) >/- T(2)(V) >/- ... >/- T(NC)(V) where pi = (1, 2, ..., NC) and is a permutation of the integers, 1,2, ..., NC associated with this arrangement. In this connection, it is to be noted that it is well known that, for minimum error rate systems, the foregoing T's are the logarithms of the probabilities of the classes for the given vector V.

If there is no rejection, the pattern recognition system identifies the input pattern as the 1th character, i.e., the one having the largest score. In a typical known rejection method, there is compared the largest score T(1) with the next largest score, T(2).

If the difference resulting from this comparison is less than some threshold value alpha, the system rejects. Thus, the system rejects whenever alpha > T(1)(V) - T(2)(V). It is evident that such rejection method does not avail itself of the information provided by the other scores.

In accordance with the method described herein, such other information is utilized and, in addition, the described method is linear in T's, whereby simple implementation is enabled. The method includes the computing of T(0)(V) as a convex combination of T(2), T(3),... T(NC) plus a constant in accordance with the following equation:

(Image Omitted)

Referring now to the figure, in steps 10 and 12 thereon, there are respectively carried out a computation of the recognition scores T's and the...