Character recognition method

• US 4,903,313 A
• Filed: 07/02/1987
• Issued: 02/20/1990
• Est. Priority Date: 07/03/1986
• Status: Expired due to Term

• Current Assignee: Ricoh Company Limited
• Original Assignee: Ricoh Company Limited
• Inventors: Tachikawa, Michiyoshi
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): Mancuso, Joseph
• Application Number:
  US07/069,303
• Time in Patent Office: 964 Days
• Field of Search: 382/27, 382/22, 382/16, 382/9, 382/21, 382/10, 382/19, 382/18
• US Class Current: 382/171
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/182 by coding the contour of the pattern

Abstract

A character recognition method includes a step of assigning one of a plurality of predetermined directionality codes to each of the pixels defining the contour of a character pattern. Then, the character pattern, to which the directionality codes have been assigned, is subdivided into a plurality of mesh regions. In this case, in accordance with the present invention, the character pattern is not directly subdivided into mesh regions, and, instead, the character pattern is first subdivided into a first plurality of sub-regions and then the first plurality of sub-regions are converted into a second plurality, which is smaller in number than the first plurality, of mesh regions by combining at least some of said first plurality of sub-regions. Such a two-step subdivision process of a character pattern is advantageous in absorbing fluctuations caused by modifications of the character pattern, such as hand-written characters.

Claims

1. A character pattern recognition method comprising the steps of:

opticallly reading an original image to produce an image signal;

producing an individual input character pattern from said image signal;

extracting a predetermined feature vector from said input character pattern, said extracting step including subdividing said input character pattern into a first plurality of sub-regions in accordance with a first subdivision scheme such that predetermined feature quantities are distributed among the sub-regions substantially equally, and converting said first plurality of sub-regions into a second plurality of mesh regions in accordance with a second subdivision scheme by combining at least some of said first plurality of sub-regions into one or more of said mesh regions; and

comparing said feature vector of said input character pattern with each of the feature vectors of a plurality of known character patterns stored in a library, thereby selecting one of said plurality of known character patterns, which is closer to said input character pattern, thereby determining identification of said input character pattern.

2. The method of claim 1, wherein said extracting step includes assigning one of a plurality of predetermined directionality codes to each of pixels defining the contour of said input character pattern and subdividing said character pattern into said first plurality of sub-regions such that said assigned directionality codes are distributed substantially equally over said first plurality of sub-regions.

3. The method of claim 2, wherein each of said plurality of directionality codes corresponds to a particular arrangement of a predetermined number of pixels which includes a center pixel of interest, a first pair of top and bottom pixels located at top and bottom of said center pixel of interest, respectively, and a second pair of right and left pixels located to the right and to the left of said center pixel of interest, respectively.

4. The method of claim 2, wherein said extracting step further includes classifying said assigned directionality codes into one of a predetermined number of layers by scanning said character pattern having said directionality codes assigned to the contour pixels thereof before subdividing said character pattern into said first plurality of sub-regions.

5. The method of claim 4, wherein said scanning of said character pattern is carried out first in a first straight direction, then in a second straight direction normal to said first straight direction, and then reversing directions along said first and second directions, whereby the directionality codes detected during each scanning are classified into said predetermined number of layers in the order of detection one for each layer.

6. The method of claim 10, wherein at least two adjacent ones of said second plurality of mesh regions have a common overlapping region which is defined by a selected one or more of said first plurality of sub-regions.