Character recognition method and apparatus

• US 5,768,451 A
• Filed: 06/02/1997
• Issued: 06/16/1998
• Est. Priority Date: 12/22/1993
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Hisamitsu, Toru, Nitta, Yoshihiko, Shimokawabe, Hiroaki, Shima, Yoshihiro, Marukawa, Katsumi
• Primary Examiner(s): Johns, Andrew W.
• Assistant Examiner(s): Davis, Monica S.
• Application Number:
  US08/867,774
• Time in Patent Office: 379 Days
• Field of Search: 364/419.08, 364/419.1, 364/419, 382/309
• US Class Current: 382/309
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/268 Lexical context

Abstract

A character recognition method is arranged to supplement an erroneously recognized character with a linguistic knowledge. In this method, the extraction of a candidate based on the search of a word dictionary occupies a large part of the operation. To speed up the extraction of a candidate word, therefore, the method is provided to search the dictionary by using a group of candidate characters or a dictionary header for processing an inflected form of a verb. Further, the present method is provided for calculating a word matching cost for improving an efficiency of modifying an erroneous recognition. The word search is done by using a "hybrid method" arranged of "candidate-character-driven word extraction" and "dictionary-driven word extraction". Moreover, the word-dictionary is arranged to have a header word composed of an inflectional ending of a verb and an auxiliary verb or a particle added to the tail of the inflectional ending. The present method attaches much importance to the difference of the matching cost about a character with a totally high confidence ratio than the difference of the matching cost about a character with a totally low confidence ratio.

Claims

1. A character recognition method comprising the steps of:

inputting a digital image including an image of a character sequence;

recognizing the image of the character sequence as a sequence of separated character patterns;

outputting a group of candidate characters, for each of the character patterns, along with a numerical confidence ratio valuegiven by an optical character reader for each output candidate character in the groups wherein the confidence ratio value provides the certainty of each candidate character as a recognition result of each of the character patterns;

extracting a group of candidate words from a word dictionary by using keys, each candidate word starting at a character position p, each key being a candidate character sequence including characters taken from the candidate characters corresponding to each position starting at the character position p;

calculating a word matching cost, for each of the extracted candidate words, by executing a word matching of each candidate word, wherein the word matching cost c(w) for w=C1, C2, . . . , Cm is represented as a sum of character matching costs of characters composing the word, wherein the character matching cost of Ci is a positive value calculated from Ci and Si={(Ci1, Vi1) (Ci2, Vi2), . . . , (CiN, ViN)}, wherein Ci1 is the j-th candidate character for Ci and Vij is its confidence ratio;

detecting a most feasible candidate word sequence in accordance with a candidate word set, the calculated word matching cost for each candidate word and a grammatical knowledge relating to a connection cost between two words; and

,

displaying the character sequence generated by connecting the most feasible candidate word sequence;

wherein the word dictionary contains pairs of entries forming the character sequence and the grammatical knowledge; and

,

wherein each of the keys is generated from the candidate characters and combinations of the candidate characters, up to a predetermined length L, starting from the position p, words up to the length L being extracted from the dictionary by coincidence with the entries in said dictionary, and words having a length larger than the length L being extracted from the dictionary by coincidence with the first L characters of the entries in the dictionary.

2. A character recognition method according to claim 1, wherein said predetermined length L is set to 2.

3. A character recognition method comprising the steps of:

inputting a digital image including an image of a character sequence;

recognizing the image of the character sequence as a sequence of separated character patterns;

outputting a group of candidate characters, for each of the character patterns, along with a numerical confidence ratio value given by an optical character reader for each output candidate character in the group wherein the confidence ratio value provides the certainty of each candidate character as a recognition result of each of the character patterns;

extracting a group of candidate words from a word dictionary by using keys, each candidate word starting at a character position p, each key being a candidate character sequence including characters taken from the candidate characters corresponding to each position starting at the character position p;

calculating a word matching cost by executing word matching for each extracted candidate word, wherein the word matching cost c(w) for w=C1, C2, . . . , Cm is represented as a sum of character matching costs of characters composing the word, wherein the character matching cost of Ci is a positive value calculated from Ci and Si={(Ci1, Vi1), (Ci2, Vi2). . . . , (CiN, ViN)}, wherein Cij is the j-th candidate character for Ci and Vij is its confidence ratio;

detecting a most feasible candidate word sequence in accordance with a candidate word set, the calculated word matching cost for each candidate word and a grammatical knowledge relating to a connection cost between two words; and

,

displaying the character sequence generated by connecting the most feasible candidate word sequence;

wherein the dictionary has a plurality of entries, each entry having a verb stem and character sequences of concatenation of two words such as the concatenation of an inflectional ending and the head of the verb stem for an auxiliary verb, a particle, a comma, a period and other words frequently appearing after the inflectional ending.

4. A character recognition method comprising the steps of:

inputting a digital image including an image of a character sequence;

recognizing the image of the character sequence as a sequence of separated character patterns;

outputting a group of candidate characters, for each of the character patterns, along with a numerical confidence ratio value given by an optical reader for each output candidate character in the group, wherein the confidence ratio value provides the certainty of each candidate character as a recognition result of each of the character patterns,

extracting a group of candidate words from a word dictionary by using keys, each candidate word starting at a character position p, each key being a candidate character sequence including characters taken from the candidate characters corresponding to each position starting at the character position p;

calculating a word matching cost, for each of the extracted candidate words, by adding all the character matching costs of the characters composing the words wherein the calculation of the character matching cost includes;

calculating the matching cost of a character C against a set S={(C1,V1), . . . , (Ci, Vj) . . . , (CN,VN)}, (where Cj is the j-th candidate character for said C, and Vj is its confidence ratio) by using a function K(C,S) having a specific property;

when said C equals Ci, K(C,S) increases as (V1-Vi) increases and K(C,S) decreases as V1 decreases;

detecting a most feasible candidate word sequence in accordance with a candidate word set, the calculated word matching cost for each candidate word and a grammatical knowledge relating to a connection cost between two words;

displaying the character sequence generated by connecting the most feasible candidate word sequence;

recording a connecting cost corresponding to the numerical information indicating how easily the words are connected;

ranking the confidence of each of the word sequences composed of a combination of the words contained in a candidate word lattice and calculating the most feasible candidate word sequence by using a linear sum of the word matching cost and the connecting cost;

displaying consecutive character sequences generated by connecting the word sequences.

5. A character recognition method comprising the steps of:

inputting a digital image including an image of a character sequence;

recognizing the image of the character sequence as a sequence of separated character patterns;

outputting a group of candidate characters, for each of the character patterns, along with a numerical confidence ratio value given by an optical character reader for each output candidate character in the group wherein the confidence ratio value provides the certainty of each candidate character as a recognition result of each of the character pattern;

extracting a group of candidate words from a word dictionary by using keys, each candidate word starting at a character position p, each key being a candidate character sequence including characters taken from the candidate characters corresponding to each position starting at the character position p;

calculating a word matching cost, for each of the extracted candidate words, by executing a word matching of each candidate word, wherein the word matching cost c(w) of w=C1, C2, . . . , Cm is represented as a sum of character matching costs of characters composing the word, wherein the character matching cost of Ci is a positive value calculated from Ci and Si={(Ci1, Vi1), (Ci2, Vi2), . . . (CiN, ViN)}, wherein Cij is the j-th candidate character for Ci and Vij is its confidence ratio;

detecting a most feasible candidate word sequence in accordance with a candidate word set, the calculated word matching cost for each candidate word and a grammatical knowledge relating to a connection cost between two words;

displaying the character sequence generated by connecting the most feasible candidate word sequence;

wherein the dictionary contains pairs of entries having a character sequence and grammatical information; and

,

said each key at the position p for extracting a candidate word is generated from each candidate character and combination of candidate characters up to a predetermined length L starting from the position p, words up to the length L are extracted from the dictionary by coincidence with entries in said dictionary, and words having a length larger than the length L are extracted from the dictionary by coincidence with a p1-th or a p2-th or . . . a pk-th character between a p1-th or a p2-th or . . . a pk-th character of an entry in the dictionary respectively, where p1, p2, . . . , pk are predetermined positive integers.

6. A character recognition method according to claim 5, wherein said predetermined length L is set to 2.

7. A character recognition apparatus comprising:

means for inputting a digital image including an image of a character sequence;

means for recognizing the image of the character sequence as a sequence of separated character patterns;

means for outputting a group of candidate characters corresponding to each of said character patterns and a numerical value of a confidence ratio given by an optical character reader to each of said output candidate characters in the group, wherein the confidence ratio value provides the certainty of each candidate character as a recognition result of each of the character pattern;

means for extracting a group of candidate words starting from each of character position, p, from a word dictionary by using keys, in which a key is a candidate character sequence each character of which is taken from said candidate characters corresponding to each position starting from p;

means for calculating word matching cost by executing word matching of each candidate word of said all extracted candidate words, wherein the word matching cost c(w) of w=C1, C2, . . . , Cm is represented as a sum of character matching costs of characters composing the word, wherein the character matching cost of Ci is a positive value calculated from Ci and Si={(Ci1, Vi1), (Ci2, Vi2), . . . , (CiN, ViN)} wherein Cij is the j-th candidate character for Ci and Vij is its confidence ratio;

means for detecting the most feasible candidate word sequence in accordance with said candidate word set, calculated word matching cost for each candidate word and grammatical knowledge relating to a connection cost between two words;

means for displaying a character sequence generated by connecting said word sequence;

wherein said dictionary contains pairs of an entry, which is a character sequence, and grammatical information therefor; and

,

said each key at the position p for extracting a candidate word is generated from each candidate character and combination of candidate characters up to predetermined length L starting from said position p, words up to said length L are extracted from said dictionary by coincidence with entries in said dictionary, and words length of which is larger than said L are extracted from said dictionary by coincidence with their first L characters and entries in said dictionary.