Method of reading characters and method of reading postal addresses

• US 6,246,794 B1
• Filed: 12/11/1996
• Issued: 06/12/2001
• Est. Priority Date: 12/13/1995
• Status: Expired due to Term

• Current Assignee: Hitachi-Omron Terminal Solutions Corporation (Hitachi, Ltd.)
• Original Assignee: Hitachi, Ltd.
• Inventors: Watanabe, Shigeru, Shima, Yoshihiro, Sako, Hiroshi, Kagehiro, Tatsuhiko, Koga, Masashi, Ogata, Hisao, Fujisawa, Hiromichi, Teramoto, Masato
• Primary Examiner(s): Tran, Phuoc
• Application Number:
  US08/763,515
• Time in Patent Office: 1,644 Days
• Field of Search: 382/101, 382/102, 382/177, 382/178, 382/185, 382/186, 382/229, 382/230, 382/231, 382/286
• US Class Current: 382/185
• CPC Class Codes: G06V 10/20 Image preprocessing
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/153 using recognition of characters or words
• CPC Class Codes: G06V 30/262 using context analysis, e.g. lexical, syntactic or semantic context

Abstract

A character reading method has enhanced character segmentation accuracy and character string recognition accuracy for reading correctly hand-written addresses on postal matters. The method extracts provisional character patterns from image information of the address character string (step 206), creates a table 219 of tentative character patterns and implements the character classification for the tentative character patterns (step 207), extracts, specifically for characters of the street number portion of the address character string, periphery information (vertical and horizontal lengths, vertical/horizontal length ratio, pattern spacings, etc.) of tentative character patterns (step 212), and segments the character string into characters accurately based on the information (step 215).

Claims

1. A method of reading characters by converting image information of a written surface into an electrical signal and reading characters of a character string included in the image information, said method comprising:

a first step of locating a character string description region in the electrical signal of the image information, and segmenting image information of a character string in the character string region into multiple tentative character patterns;

a second step of implementing the character classification for the tentative character patterns by making reference to a character classification dictionary thereby to obtain multiple recognition candidate characters for each tentative character pattern;

a third step of obtaining border information for the tentative character patterns;

a fourth step of obtaining the credibility of the border information of the tentative character patterns obtained in said third step by making reference to a segmentation dictionary which contains the border information by using the recognition-candidate characters obtained in said second step as the key, and applying weights to the tentative character patterns;

a fifth step of determining the character segmentation in accordance with the weights of tentative character patterns; and

a sixth step of implementing the word-wise matching by using the character classification dictionary based on a set of classified character species produced from the tentative character patterns determined in the fifth step, and identifying the characters of the character string.

2. A character reading method according to claim 1, wherein said periphery information of the tentative character patterns includes at least one of the vertical length, horizontal length, vertical/horizontal length ratio, pattern spacing and number of connected components of the character pattern, and wherein said segmentation dictionary contains information of likelihood distribution for at least one of the vertical length, horizontal length, vertical/horizontal length ratio, pattern spacing and number of connected components of each character, and wherein said reference key of the segmentation dictionary in said fourth step is the character species of candidate characters resulting from character classification.

3. A character reading method according to claim 1, wherein said fourth step includes a step of creating and revising the segmentation dictionary, said creation/revision step including a step of displaying on the screen of a display device the input image of a character string to be recognized, the tentative character patterns, and candidate characters resulting from character classification for the tentative character patterns, a step of revising the segmentation dictionary by using the display on the screen of the display device, and a step of displaying the state of character string recognition process graphically on the screen.

4. A method of reading a postal address comprising:

a first step of converting image information, which includes character string information having a town name portion and a street number portion, into an electrical signal;

a second step of locating a character string description region in the electrical signal of the image information, and extracting combinations of connected image components, which form characters in the character string description region, as tentative character patterns;

a third step of implementing the character classification for each of the tentative character patterns by making reference to the character classification dictionary thereby to obtain recognition candidate characters and the similarity of tentative character patterns and the recognition-candidate characters;

a fourth step of forming a lattice consisting of the recognition-candidate characters, implementing the matching for the lattice with a town name dictionary thereby to identify character strings of the town name portion in the tentative character patterns, and detecting the head position of the street number portion;

a fifth step of extracting, based on the information of the head position obtained in said fourth step, periphery information of tentative character patterns which correspond to recognition-candidate characters of tentative character patterns in the street number portion, and applying weights to the tentative character patterns for evaluating the credibility of the periphery information of the tentative character patterns by making reference to the segmentation dictionary, which contains likelihood of the periphery information, by using the recognition-candidate character as the key;

a sixth step of segmenting the street number portion into characters based on the weights; and

a seventh step of implementing the word-wise matching with a street number dictionary for a set of character classification results produced in said sixth step thereby to identify the character string of street number.

5. A postal address reading method according to claim 4, wherein said second step includes a step of extracting characters from the top and end of the extracted character string, and discriminating the vertical writing form or horizontal writing form based on the classification of the extracted character patterns.

6. A postal address reading method according to claim 4, wherein said pattern periphery information includes at least one of the vertical length, horizontal length, vertical/horizontal length ratio, pattern spacing and number of connected components of character pattern, and wherein said segmentation dictionary contains information of likelihood distribution for at least one of the vertical length, horizontal length, vertical/horizontal length ratio, pattern spacing and number of connected components of each character, and wherein said reference key of recognition-candidate character in said fifth step is characters in the segmentation dictionary.

7. A postal address reading method according to claim 4, wherein said fifth step includes a step of creating and revising the segmentation dictionary, said creation/revision step including a step of displaying on the screen of a display device the input image of an address character string to be recognized, said character patterns and candidate characters resulting from character classification for the character patterns, a step of revising the segmentation dictionary by using the display on the screen of the display device, and a step of displaying the state of address character string recognition process graphically on the screen.

8. A method of reading characters with a postal address reading apparatus having an image input means for converting image information of a written surface into an electrical signal and means of reading out of the image a character string written on the surface, said method comprising:

a first step of extracting the signal of the character string from the electrical signal of the image;

a second step of extracting a tentative character pattern which is deemed to form a character from the signal of the character string, or, in case a tentative character pattern cannot be determined uniquely, extracting a plurality of tentative character patterns;

a third step of implementing the character classification for the extracted tentative character pattern;

a fourth step of calculating the external form penalty based on the assessment of the periphery information depending on the possible types of error of character segmentation; and

a fifth step of confining candidates of tentative character patterns in accordance with the character classification result of said step 3 and the external form penalty calculated in said fourth step, and implementing the matching for the character pattern candidates with character strings stored in advance in a dictionary which contains character strings that can possibly be written on written surfaces, thereby recognizing the character string written on the written surface.

9. A character reading method according to claim 8, wherein said character string written on the written surface is a character string which signifies a postal address, and said dictionary which contains character strings that can possibly be written on written surfaces comprises a dictionary in which postal addresses are described.

10. A character reading method according to claim 8, wherein said fourth step uses discriminant functions which distinguish a tentative character pattern segmented correctly and a tentative pattern segmented erroneously.

11. A character reading method according to claim 10, wherein one of said discriminant functions treats as one of feature values the spacings of each tentative character pattern with the neighboring patterns.

12. A character reading method according to claim 10, wherein one of said discriminant functions treats the gap in each pattern.

13. A character reading method according to claim 10, wherein said discriminant functions of said fourth step are created by the method which comprises:

a step of extracting a tentative character pattern which is deemed to form a character string from the image of the character string, or, in case a tentative character pattern cannot be determined uniquely, extracting a plurality of tentative character patterns;

a step of entering information on as to whether or not the tentative character pattern is segmented correctly, with types of segmentation error being sorted manually in the case of incorrect segmentation;

a step of storing tentative character patterns in a memory by sorting the tentative character patterns depending on the result of said incorrect segmentation judgment step; and

a step of implementing the learning of a classifying device by using the tentative character patterns stored in the memory by said pattern storing step.