Browse Prior Art Database

Multilevel Adaptive Decision Techniques

IP.com Disclosure Number: IPCOM000095115D
Original Publication Date: 1965-Sep-01
Included in the Prior Art Database: 2005-Mar-07
Document File: 2 page(s) / 14K

Publishing Venue

IBM

Related People

Chien, RT: AUTHOR [+3]

Abstract

A number of methods are provided for increasing the accuracy of character recognition systems by introducing multilevel, adaptive decision arrangements. The central concept is that a rejection criteria is set up to isolate the confusion groups of the analysis data as a function of the given measurements. Enough allowance is made so that the substitution-error-rate after rejection is in an acceptable range. These confusion groups are analyzed and a second level decision arrangement is constructed to break these confusion groups. When in the recognition mode, both stages are applied, either forced decision or decision with a reject level can be used at the second level. First Level Decision and Rejection.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 54% of the total text.

Page 1 of 2

Multilevel Adaptive Decision Techniques

A number of methods are provided for increasing the accuracy of character recognition systems by introducing multilevel, adaptive decision arrangements. The central concept is that a rejection criteria is set up to isolate the confusion groups of the analysis data as a function of the given measurements. Enough allowance is made so that the substitution-error-rate after rejection is in an acceptable range. These confusion groups are analyzed and a second level decision arrangement is constructed to break these confusion groups. When in the recognition mode, both stages are applied, either forced decision or decision with a reject level can be used at the second level. First Level Decision and Rejection.

In recognition systems with binary measurements, each sample of a particular character can be represented as a binary vector. The sample vectors of a particular character can then be viewed as a binary matrix with I rows and J columns. 1 is the number of samples and J the number of measurements.

From each measurement-sample matrix, there is constricted a three-level reference vector by setting up a threshold T. This setting up is in such a way that the components r. of the reference vector depend on both the average value

(Image Omitted)

and the threshold T. Specifically, Hence, for each character Ck of the K possible characters, such a reference vector Rk can be constructed.

When an unknown sample is presented for recognition, a distance is computed between the unknown binary vector and each reference vector R. The distance measure is a modified Hamming distance on the digit-by-digit basis. For each digit, the only pair that contributes a 1 unit to the total distance is when a 1 bit is matched with a 0 bit and vice versa. The distance between the unknown sample and the k/th/ reference vector Rk is denoted by d(k). The sequencing of the character set is further rearranged such that when k runs through k(1), k(2), k(3).... k(K) the result is:

(Image Omitted)

In the expression k(1) is the first candidate, k(2) is the second candidate, etc.

If forced one-level minimum decision is used, k(1) is always picked as the selected one. In two-level decision arrangements to be discussed later, the following rejection criteria are used. A threshold S is set and usually has the values 2, 3, or 4. Also, the first transitions alpha and beta are such that d(alpha +
1) - d(alpha) >/- 2 and d(beta+1)-d(beta) >/- 1....