Browse Prior Art Database

FEATURE CONTENT PRESCAN FOR WORD RECOGNITION

IP.com Disclosure Number: IPCOM000015770D
Original Publication Date: 2002-Oct-19
Included in the Prior Art Database: 2003-Jun-21
Document File: 2 page(s) / 61K

Publishing Venue

IBM

Abstract

Problem

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 2

FEATURE CONTENT PRESCAN FOR WORD RECOGNITION

Problem

    Described herein is a method that can significantly reduce the time needed to find the best image match of a word in a collection of other word images. This task is typically needed in OCR-based word recognition, when there are content constraints, such as words that must be country names.

Solution

    The patent US4355371 describes a method called Alpha Content Prescan that speeds up spell-checking by comparing character presence, before performing more elaborate string matching. Character presence is compared by representing groups of characters as bits in a bitmap, XORing the bitmaps for the checked word with those of a list of correctly spelled words, and counting the set bits (1's) in the result. The number of set bits is not more than the number of character mismatches between the checked word and the candidate, and can be used to reject the candidate altogether, or to sort candidates by descending likelihood of being the best match. Mapping characters to bits can take into account the likelihood of character substitution. For instance, if M and N are likely to be confused, it is advantageous to map them both to the same bit, so the alpha content prescan will not count such substitutions as errors, and reject the correct candidate.

    This invention addresses the problem of speeding up automatic recognition of handwritten words by a computer. In this case, there is much more likelihood of character substitution, so the idea of mapping characters to the same bit may render the prescan ineffective. For example, in some handwriting recognition engines, the characters H and N are easily confused, and so are H and A. But A and M are almost never confused. The reason for this is that H and N are confusable when the H is slanted, so the horizontal line becomes similar to the diagonal in the N, and H and A are confusable when the to of the A is not fully closed, and the H is not slanted. If each bit in the bitmap represents a character, then either all of H, N, and A will be mapped to a single bit, causing A to be confusable with N, or none will, so that the knowledge that A is confusable with H, and H with N, is lost.

    The mai...