Character recognition apparatus

• US 6,738,519 B1
• Filed: 06/09/2000
• Issued: 05/18/2004
• Est. Priority Date: 06/11/1999
• Status: Expired due to Fees

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Nishiwaki, Daisuke
• Primary Examiner(s): Dastouri, Mehrdad
• Application Number:
  US09/590,261
• Time in Patent Office: 1,439 Days
• Field of Search: 382/173, 382/174, 382/175, 382/176, 382/177, 382/178, 382/179, 382/185, 382/186, 382/187, 382/195, 382/199, 382/194, 382/202, 382/203, 382/209, 382/217, 382/218, 382/159, 382/161, 382/228, 358/453, 358/462
• US Class Current: 382/228
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/2272 with lexical matching
• CPC Class Codes: G06V 30/268 Lexical context

Abstract

There are provided: an n-fold-character recognizing part for collectively recognizing an unmatched portion without segmenting character candidate patterns character by character for an image of a read-wise skipped portion, i.e., the unmatched portion upon word verification; and an n-fold-character recognizing dictionary referred to by the n-fold-character recognizing part upon recognition; to thereby conduct re-recognition independent of instability of character segmentation even when the portion read-wise skipped by the word verification includes two or more characters.

Claims

1. A character recognition apparatus comprising:

image storing means for inputting and storing a character string image;

character segmentation means for producing character candidate patterns character by character for the character string image inputted into said image storing means, and for detecting a character-contacting portion in the character string image to thereby estimate the number of characters in the character-contacting portion;

single-character recognition means for deciding character codes for the character candidate patterns character by character produced by said character segmentation means, and for outputting certainty levels of the character codes;

a single-character recognizing dictionary to be used by said single-character recognition means for decision

n-fold-character recognition means for deciding character codes corresponding to n pieces of characters, when the character-contacting portion detected by said character segmentation means is estimated to include the n pieces of characters;

an n-fold-character recognizing dictionary that stores features for verification from character string patterns for learning to be used by said n-fold-character recognition means for decision;

word verification means for outputting verified words in a descending order of a matching number of characters for the registered candidate words, in accordance with single-character recognition results output by said single-character recognition means and n-fold-character recognition results output by said n-fold-character recognition means;

a word dictionary to be used by said word verification means upon verification; and

controlling means for controlling said image storing means, said character segmentation means, said single-character recognition means and said n-fold-character recognition means, wherein said controlling means comprises;

adopting/stopping means for adopting recognition results outputted from said single-character recognition means by examining certainty levels for all of the recognition results character by character for the character string image and for stopping the recognition for the inputted character string image if all the certainty levels are larger than a predetermined value, and if outputted from said word verification means completely corresponds to the recognition result outputted from said single-character recognition means;

sending means for sending partial character string images corresponding to the unmatched portions in the single-character recognition results upon word verification toward said n-fold-character recognition means, together with an estimated number of characters and an estimated character codes of the unmatched portions, in a descending order of a matching number of characters for the unverified words when no verification results outputted from said word verification means completely corresponds to the recognition result outputted from said single-character recognition means and

means for conducting, wherein said n-fold-character recognition means re-recognizes the partial character string images sent from said sending means and outputs the recognition results and certainty levels therefor, and when each of certainty levels for pertinent recognition results is larger than a predetermined value, each of the pertinent words being verified is kept as a final candidate, and an evaluation value for each of the whole character strings being verified is calculated that uses the certainty level outputted by the single-character recognition means for the portion for which said single-character recognition means has been activated and uses the certainty level outputted by said n-fold-character recognition means for the portion for which said n-fold-character recognition means has been activated and a word, which has the largest evaluation value among the words kept as the final candidates among the verified words, is adopted as a recognition result for the inputted character string.

2. A character recognition apparatus of claim 1, wherein said character segmentation means includes means for detecting a character boundary between contiguous characters upon producing the character candidate patterns character by character from the character-contacting portion, making use of a transition of run lengths of contiguous runs upon tracing black pixels in a vertical direction relative to a character string direction for black pixels of the character-contacting portion.

5. A character recognition apparatus of claim 1, wherein said n-fold-character recognition means comprises means for assuming a rectangle surrounding a partial character string image corresponding to the character-contacting portion upon conducting recognition for the whole of the character-contacting portion without dividing the character-contacting portion into character candidate patterns character by character, for setting an area having a height corresponding to the height of the rectangle, and a width equal to or smaller than a single character width, for conducting recognition making use of a transition of a feature within said area, the transition being obtained for each movement of said area from one end of the rectangle toward the other end thereof and for outputting a certainty level of the thus obtained recognition result, while storing a recognizing dictionary therefor, in said n-fold-character recognizing dictionary.

4. A character recognition apparatus of claim 4, wherein said n-fold-character recognition means comprises means for conducting recognition for the whole of the character-contacting portion without dividing the character-contacting portion into character candidate patterns character by character and for outputting a certainty level of the recognition result thereof, while storing a recognizing dictionary for the whole recognition, in said n-fold-character recognizing dictionary.

3. A character recognition apparatus of claim 4, wherein said character segmentation means includes means for detecting a character boundary between contiguous characters upon producing the character candidate patterns character by character from the character-contacting portion, making use of a maximum point and a minimum point in a vertical direction relative to a character string direction for a character stroke of the character-contacting portion.

6. A character recognition apparatus comprising:

an image storing part that inputs and stores a character string image;

a character segmentation part that produces character candidate patterns character by character for the character string image inputted into said image storing part, and detects a character-contacting portion in the character string image to thereby estimate the number of characters in the character-contacting portion;

a single-character recognition part that decides character codes for the character candidate patterns character by character produced by said character segmentation part, and that outputs certainty levels of the character codes;

a single-character recognizing dictionary to be used by said single-character recognition part for decision;

n-fold-character recognition part that decides character codes corresponding to n pieces of characters, when the character-contacting portion detected by said character segmentation part is estimated to include the n pieces of characters;

an n-fold-character recognizing dictionary that stores features for verification from character string patterns for learning to be used by said n-fold-character recognition part for decision;

a word verification part that outputs verified words in a descending order of a matching number of characters for the registered candidate words, in accordance with single-character recognition results output by said single-character recognition part and n-fold-character recognition results output by said n-fold-character recognition part;

a word dictionary to be used by said word verification part upon verification; and

a controlling part that controls said image storing part, said character segmentation part, said single-character recognition part and said n-fold-character recognition part, wherein said controlling part comprises includes;

an adopting/stopping part that adopts recognition results outputted from said single-character recognition part by examining certainty levels for the recognition results character by character for the character string image and stops the recognition for the inputted character string image when all the certainty levels are larger than a predetermined value, and when a verification result outputted from said word verification part corresponds to the recognition result outputted from said single-character recognition part;

a sending part that sends partial character string images corresponding to the unmatched portions in the single-character recognition results upon word verification toward said n-fold-character recognition part, together with an estimated number of characters and an estimated character codes of the unmatched portions, in a descending order of a matching number of characters for the unverified words when no verification results outputted from said word verification part completely correspond to the recognition result outputted from said single-character recognition part; and

a conducting part, wherein said n-fold-character recognition part re-recognizes the partial character string images sent from said sending part and outputs the recognition results and certainty levels therefor and when each of certainty levels for pertinent recognition results is larger than a predetermined value, each of the pertinent words being verified is kept as a final candidate, and an evaluation value for each of the whole character strings being verified is calculated that uses the certainty level outputted by the single-character recognition part for the portion for which said single-character recognition part has been activated and uses the certainty level outputted by said n-fold-character recognition part for the portion for which said n-fold-character recognition part has been activated and a word which has the largest evaluation value among the words kept as the final candidates among the verified words is adopted as a recognition result for the inputted character string.

7. The character recognition apparatus of claim 6, wherein said character segmentation part includes a part that detects a character boundary between contiguous characters upon producing the character candidate patterns character by character from the character-contacting portion, making use of a transition of run lengths of contiguous runs upon tracing black pixels in a vertical direction relative to a character string direction for black pixels of the character-contacting portion.

8. A character recognition apparatus of claim 6, wherein said character segmentation part includes a part that detects a character boundary between contiguous characters upon producing the character candidate patterns character by character from the character-contacting portion, making use of a maximum point and a minimum point in a vertical direction relative to a character string direction for a character stroke of the character-contacting portion.

9. A character recognition apparatus of claim 6, wherein said n-fold-character recognition part includes a part that conducts recognition for the whole of the character-contacting portion without dividing the character-contacting portion into character candidate patterns character by character and outputs a certainty level of the recognition result thereof, while storing a recognizing dictionary for the whole recognition, in said n-fold-character recognizing dictionary.

10. A character recognition apparatus of claim 6, wherein said n-fold-character recognition part comprises a part that assumes a rectangle surrounding a partial character string image that corresponds to the character-contacting portion upon conducting recognition for the whole of the character-contacting portion without dividing the character-contacting portion into character candidate patterns character by character, sets an area having a height corresponding to the height of the rectangle, and a width equal to or smaller than a single character width, conducts recognition, making use of a transition of a feature within said area, the transition being obtained for each movement of said area from one end of the rectangle toward the other end thereof and outputs a certainty level of the thus obtained recognition result, while storing a recognizing dictionary therefor, in said n-fold-character recognizing dictionary.

11. A machine readable recording medium recorded with a character string recognition program to be executed on a computer, said character string recognition program comprising the steps of:

inputting and storing a character string image;

producing character candidate patterns character by character for the inputted character string image, and detecting a character-contacting portion in the character string image to thereby estimate the number of characters in the character-contacting portion;

deciding character codes for the produced character candidate patterns character by character, and outputting certainty levels of the character codes, wherein features for verification from character string patterns for learning are applied in said deciding step;

performing word verification and outputting verified words in a descending order of a matching number of characters for the registered candidate words, in accordance with the single-character recognition results and the n-fold-character recognition results;

outputting the recognition results as a recognition result for the inputted character string when each of the certainty levels of the single-character recognition results and the n-fold-character recognition results is larger than a predetermined value, and when a word verification result corresponds to the outputted recognition result for the produced character candidate patterns; and

when each of the certainty levels are smaller or equal to a predetermined value and involved in the character-contacting portion, then

deciding character codes corresponding to n pieces of characters for the character-contacting portion, and outputting a certainty level therefor;

sending partial character string images corresponding to the unmatched portions in the single-character recognition results upon said word verification, together with the estimated number of characters and the estimated character codes of the unmatched portions, in a descending order of a matching number of characters for the unverified words when no word verification results completely correspond to the outputted recognition result for the character candidate patterns; and

re-recognizing the partial character string images and outputting the recognition results and certainty levels therefor and when each of certainty levels for pertinent recognition results is larger than a predetermined value, each of the pertinent words being verified is kept as a final candidate, and an evaluation value for each of the whole character strings being verified is calculated that uses the outputted certainty level; and

adopting as a recognition result for the inputted character string one of said words which has a largest evaluation value among the verified words kept as the final candidates.