Browse Prior Art Database

Multilevel Character Recognition System

IP.com Disclosure Number: IPCOM000074611D
Original Publication Date: 1971-May-01
Included in the Prior Art Database: 2005-Feb-23
Document File: 4 page(s) / 66K

Publishing Venue

IBM

Related People

Cutaia, A: AUTHOR

Abstract

In one type of optical character recognition (OCR) system, an entire electronic image of an input character is correlated with a set of corresponding reference patterns representing a number of ideal characters in an alphabet to be recognized. Such a "total-mask" system is capable of high recognition rates on poor-quality printing, but it is very slow, especially when the character set is large. In another type of OCR system, only selected portions of the input character are compared with references representing corresponding portions of the ideal characters. This "partial-mask" system, which commonly employs data from widely spaced scans, can be simple and rapid; on the other hand, its recognition rate is usually limited to 95-98 % or less, except for high-quality machine printing.

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 36% of the total text.

Page 1 of 4

Multilevel Character Recognition System

In one type of optical character recognition (OCR) system, an entire electronic image of an input character is correlated with a set of corresponding reference patterns representing a number of ideal characters in an alphabet to be recognized. Such a "total-mask" system is capable of high recognition rates on poor-quality printing, but it is very slow, especially when the character set is large. In another type of OCR system, only selected portions of the input character are compared with references representing corresponding portions of the ideal characters. This "partial-mask" system, which commonly employs data from widely spaced scans, can be simple and rapid; on the other hand, its recognition rate is usually limited to 95-98 % or less, except for high-quality machine printing. Moreover, conventional systems of the latter type use the same pattern information for classifying all patterns of the alphabet, so that the selection of the specific data to be correlated must be a compromise over the entire alphabet.

System 10 is a multilevel OCR machine which overcomes the foregoing deficiencies by clustering a group of decision candidates in the order of their recognition probabilities, based upon a partial-mask measurement set. These masks may differ for each character in the alphabet to be recognized. An identification is made immediately from this measurement set if certain first-level criteria are fulfilled. If no positive identification can be made, or if a conflict exists, the decision candidates are placed in an ordered list, and a second-level measurement set is applied to the input pattern. The latter set may be a total- mask set or a second partial-mask set. The individual measurements may be the same or different for each reference character. An identification of the input character is then made at the second level if certain further recognition criteria are met. System 10 differs in two major respects from prior systems in which "gross measurements" classify the pattern into one of several categories each containing similar characters (C, G, O, Q, etc.), and "fine measurements" discriminate among the individual members of each category. First, the present system allows the second-decision level to be bypassed for the great majority of input characters. Secondly, the identity of the candidates in the gross category is not fixed, but rather varies for each individual input character.

Flowchart 50 illustrates the operation of system 10. Video data for an input character is collected (51), and certain coarse features are selected (52) from a first-feature set. These features are then correlated (53) with corresponding features of the full set of reference characters. The identities of the references which match the input character most closely are stored (54) as first-level candidates in the order of their similarity. If the first-level recognition criteria are met (55, exit 56), the...