Character recognition machine utilizing language processing

• US 5,987,170 A
• Filed: 11/06/1997
• Issued: 11/16/1999
• Est. Priority Date: 09/28/1992
• Status: Expired due to Term

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Ito, Satoru, Niwa, Hisao, Yamamoto, Hiroshi, Kojima, Yoshihiro, Maekawa, Hidetsugu, Maruno, Susumu, Kouda, Toshiyuki, Kayashima, Kazuhiro, Shimeki, Yasuharu
• Primary Examiner(s): Kelley, Christopher S.
• Application Number:
  US08/965,534
• Time in Patent Office: 740 Days
• Field of Search: 382/170, 382/229, 382/227, 382/190, 382/231, 704/4, 704/10
• US Class Current: 382/170
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/274 Syntactic or semantic context, e.g. balancing

Abstract

There is disclosed a character recognition machine adapted to recognize Japanese characters such as kanjis and kanas. The machine comprises a character string storage portion, a character extraction portion, a character recognition portion, and a language processing portion. A character string to be recognized is stored as an image in the storage portion. The character extraction portion comprises a network consisting a plurality of interconnected operators each of which has numerous inputs and outputs. An evaluation function which assumes its minimum value when a character extraction produces the best results is calculated by the operators simultaneously so as to minimize the value of the function. The character recognition portion calculates degrees of similarity of a character pattern to various character categories, the character pattern being applied from the character extraction portion. The language processing portion receives these degrees of similarity and selects a character category which seems to provide the most correct combination of characters in terms of vocabulary and grammar.

Claims

1. A character extraction machine comprising:

a character string image input portion for receiving an image consisting of a character string;

a memory for holding the received character string;

a projection profile histogram-calculating portion for calculating a concentration value of histogram which is obtained by counting pixels in the character string;

an initial parameter-setting portion for receiving an output from said projection profile histogram-calculating portion and determining values of processing parameters from heights of the character string, lengths of spaces of projection profile histogram of the character string, and other factors which are obtained from the projection profile histogram, the parameters being used to estimate character positions;

a histogram distribution smoothing network portion formed by a network for receiving an output from said projection profile histogram-calculating portion and an output from said initial parameter-setting portion, said network being composed of interconnected operators each having a multiplicity of inputs and one output, said operators corresponding to positions of pixels in a direction of the character string, and said histogram distribution smoothing network portion for acting to minimize a function which assumes its minimum value when values of the projection profile histogram of character portions have substantially the same value, to thereby smooth the projection profile histogram, thus reducing effects of noises present in the received character string;

an extraction position-estimating network portion formed by a second network for receiving an output from said projection profile histogram-calculating portion and an output from said initial parameter-setting portion, said second network being composed of interconnected operators each having a multiplicity of inputs and one output, said operators corresponding to positions of pixels in a direction of the character string, and said extraction position-estimating network portion for acting to minimize a function which assumes its minimum value when an extracted position of the whole character string has been optimally extracted, to thereby estimate an extraction position of the character;

an extraction position-determining portion for determining character extraction positions according to values outputted from said extraction position-estimating network portion;

a character pattern output portion for reading a character pattern of each character from said memory according to a signal outputted from said extraction position-determining portion;

a character recognition portion for recognizing what character is represented by each character pattern produced from said character pattern output portion;

a language processing portion for receiving an output from said character recognition portion and calculating an evaluation value of each phrase of the character string applied heretofore, said evaluation value indicating a degree of correctness in terms of vocabulary and grammar;

a final character position-determining portion for receiving an output from said language processing portion and determining whether characters in the phrase have been correctly extracted in terms of vocabulary and grammar and which, if the characters have not been correctly extracted, producing positions of a first character and a last character because probability of erroneous estimation of character positions in said phrase is high and which, if the characters have been correctly extracted, producing these character positions as final character extraction positions;

a re-extraction position-determining portion for receiving an output from said final character position-determining portion and determining a position at which a character position should be estimated again;

a character squareness degree-modifying portion for estimating a squareness degree from a character string which has been judged to have correct character positions by said final character position-determining portion; and

a parameter-modifying portion for receiving an output from said re-extraction position-determining portion and an output from said character squareness degree-modifying portion and modifying said processing parameters used to estimate the character positions.

2. A character extraction machine in accordance with claim 1, further comprising:

an extraction candidate point-setting portion for detecting positions having a possibility of being character extraction positions, from an output from said projection profile histogram-calculating portion; and

an initial value-setting portion for determining initial output values from said operators in said extraction position-estimating network portion on a basis of an output from said extraction candidate point-setting portion and an output from said projection profile histogram-calculating portion.

3. A pattern recognition machine comprising:

a plurality of feature-extracting portions, each of which is for finding each feature vector from an applied pattern;

a plurality of single feature-recognizing portions, each of which is for finding degrees of similarity of the applied pattern from each feature vector, the degrees of similarity indicating degrees to which the applied pattern belong to various categories;

a category-discerning portion for discerning the applied pattern using the degrees of similarity obtained by said plurality of single feature-recognizing portions;

each of said single feature-recognizing portions including;

a group dictionary storing a plurality of group references feature vectors representing category groups each consisting of a set of patterns having similar feature vectors;

a fuzzy rough classification portion for calculating group assignment degrees from the group reference feature vectors and from feature vectors of the applied pattern, the group assignment degree indicating degrees to which the applied pattern belong to the category groups, respectively;

a plurality of subclassification portions for finding inside group similarity degrees from the feature vectors of the applied pattern, the inside group similarity degrees indicating degrees to which the applied pattern belongs to categories contained in the category groups;

a group-selecting portion for selecting a plurality of category groups from the group assignment degrees;

a subclassification portion input signal-selecting portion for selecting the subclassification portion which can receive the feature vectors of the applied pattern according to information about selection of groups obtained from said group-selecting portion; and

a single feature similarity degree-calculating portion for finding degrees of similarity of the applied pattern to the categories from a group assignment degree of the category group selected by said group-selecting portion and from the inside group similarity degrees obtained by said selected subclassification portions;

said single feature similarity degree-calculating portion including;

a plurality of multipliers for multiplying the group assignment degree of the category group selected by said group-selecting portion by all inside group similarity degrees, respectively, obtained from said subclassification portions receiving the feature vectors of the applied pattern by said subclassification portion input signal-selecting portion; and

a category similarity degree-calculating portion for selecting plural output values having larger values from output values from the multipliers for each category and calculating a sum of these output values;

said category-discerning portion including;

a plurality of similarity degree-normalizing portions for normalizing the degrees of similarity of the categories obtained from said single feature-recognizing portions corresponding to the feature vectors, respectively, with a maximum one of these degrees of similarity to thereby transform the degrees of similarity into normalized degrees of similarity; and

a category decision portion for finding integrated degrees of similarity of the categories using the normalized degrees of similarity obtained by said similarity degree-normalizing portions and discerning the applied pattern.

10. A character recognition machine in accordance with claim 3, wherein said fuzzy rough classification portion comprises:

a plurality of distance-calculating portions for calculating distances of the feature vectors of said applied character patterns from all group reference feature vectors stored in said group dictionary;

a plurality of dividers for calculating inverses of outputs from said distance-calculating portions;

an adder for summing up the outputs from said dividers;

a plurality of multipliers for multiplying an output from said adder by the outputs from said distance-calculating portions; and

a plurality of dividers for calculating inverses of outputs from said multipliers.

11. A character recognition machine in accordance with claim 3, wherein each of said subclassification portions comprises:

a character category dictionary in which character category reference feature vectors indicating typical values of character categories of said character patterns are stored;

a plurality of distance-calculating portions for calculating distances of said feature vectors from all character category reference feature vectors stored in said character category dictionary;

a plurality of dividers for calculating inverses of outputs from said distance-calculating portions;

an adder for summing up the outputs from said dividers;

a plurality of multipliers for multiplying an output from said adder by the outputs from said distance-calculating portions; and

a plurality of dividers for calculating inverses of outputs from said multipliers.

12. A character recognition machine in accordance with claim 3, wherein

each of said subclassification portions has a multilayered structure and comprises a plurality of processing portions each having a multiplicity of inputs and one output, the processing portions being interconnected in a network in such a way that there is no coupling inside each layer of said multilayered structure and that signals are propagated only to upper layers, and wherein each of said processing portions comprises a weighting coefficient storage portion for holding a plurality of weighting coefficients, input portions for receiving input signals, multiplier means for weighting the input signals applied to the input portions with the weighting coefficients held in said weighting coefficient storage portion, an adder means for summing up the input signals weighted by said multiplier means, and a threshold value processing means for limiting an output from said adder means within a given range.

4. A character recognition machine comprising:

a character recognition portion including;

a plurality of feature-extracting portions, each of which is for finding each feature vector from a character pattern produced by a character extraction portion which divides a text image into plural regions for each character;

a plurality of feature-recognizing portions, each of which is for finding degrees of similarity from each feature vector, the degrees of similarity indicating degrees at which the character pattern belongs to various character categories, respectively;

a candidate character-selecting portion for finding a plurality of second candidate character categories and their degrees of similarity to the character pattern from the degrees of similarity obtained from said single feature-recognizing portions; and

a post-recognition processing portion for selecting a final character category from the second character candidate categories using information about sizes and positions of the character pattern;

each of said single feature-recognizing portions including;

a group dictionary storing a plurality of group reference feature vectors representing character category groups each consisting of a set of character pattern having similar feature vectors;

a fuzzy rough classification portion for calculating group assignment degrees from the group reference feature vectors and from feature vectors of the applied character pattern, the group assignment degree indicating degrees to which the applied character pattern belongs to the character category groups, respectively;

a plurality of subclassification portions for finding inside group similarity degrees from the feature vectors of the applied character pattern, the inside group similarity degrees indicating degrees to which the applied character pattern belongs to character categories contained in the character category groups;

a group-selecting portion for selecting a plurality of character category groups from the group assignment degrees;

a subclassification portion input signal-selecting portion for selecting the subclassification portion which can receive the feature vectors of the applied character pattern according to information about selection of groups obtained from said group-selecting portion; and

a single feature similarity degree-calculating portion for finding degrees of similarity of the applied character pattern to the character categories from a group assignment degree of the character category group selected by said group-selecting portion and from the inside group similarity degrees obtained by said selected subclassification portions;

said single feature similarity degree-calculating portion including;

a plurality of multipliers for multiplying the group assignment degree of the character category group selected by said group-selecting portion by all inside group similarity degrees, respectively, obtained from said subclassification portions receiving the feature vectors of the applied character pattern by said subclassification portion input signal-selecting portion; and

a category similarity degree-calculating portion for selecting plural output values having larger values from output values having larger values from output values from the multipliers for each character category and calculating a sum of these output values;

said candidate character-selecting portion including;

a plurality of similarity degree-normalizing portions for normalizing the degrees of similarity of various character categories obtained from said single feature-recognizing portions corresponding to the feature vectors, respectively, with a maximum one of these degrees of similarity to thereby transform the degrees of similarity into normalized degrees of similarity;

a plurality of first candidate character-selecting portions for selecting plural first candidate character categories from all character categories using information about sizes and positions in the character pattern and the normalized degrees of similarity; and

a second character-selecting portion for summing up all normalized degrees of similarity of the selected first candidate character categories for each same category, finding integrated degrees of similarity of all the first candidate character categories, selecting plural first candidate character categories having larger values as second candidate character categories, and delivering the second candidate character categories and their integrated degrees of similarity.

5. A character recognition machine in accordance with claim 4, wherein

said post-recognition processing portion has:

a character information storage portion storing information about reference sizes and reference positions in character patterns for each character category;

a post-processed character set storage portion storing a set of post-processed character categories; and

a post-processed character set decision portion which compares second candidate character categories obtained from said candidate character-selecting portion with all sets of post-processed character categories stored in said post-processed character set storage portion and which, if said second candidate character categories form said set of post-processed character categories, reads information about the reference sizes and reference positions in the character categories corresponding to said set of post-processed character categories from said character information storage portion, compares information about sizes and positions in the applied character pattern with the read information, and determines any character category of the set of post-processed character categories as a new second candidate character category, and

wherein said first candidate character-selecting portion compares information about the reference sizes and reference positions in the character categories stored in said character information storage portion with the information about the sizes and positions in said applied character pattern to select said plural first candidate character categories.

6. The character recognition machine in accordance with claim 4, wherein

said first candidate character-selecting portions select character categories having normalized degrees of similarity greater than a certain threshold value, from said similarity degree-normalizing portions, and then compare information about the reference sizes and reference positions of the character categories stored in said character information storage portion with information about the sizes and positions in said applied character pattern, to thereby select said plural first candidate character categories.

7. A character recognition machine in accordance with claim 4, further comprising;

a one-character reject control portion for informing said character extraction portion of positions in a character string of said character pattern to said character extraction portion, if a sum of numbers of first candidate character categories obtained from said first candidate character-selecting portion is zero, with respect to some character pattern produced from said character extraction portion.

8. A character recognition machine in accordance with claim 4, further comprising:

a teacher signal-generating portion which, if a first candidate character of the second candidate categories obtained from said post-recognition processing portion differs from a correct character category of the applied character pattern, produces the correct character category;

an incorrect recognition frequency storage portion storing an incorrect recognition frequency for each set of post-processed character category, said incorrect recognition frequency indicating a frequency at which a combination of said correct character category and said first candidate character category agrees with the set of post-processed character categories;

an incorrect recognition frequency-updating portion which compares the combination of said correct character category and said first candidate character category with the set of post-processed character categories stored in said post-processed character set storage portion and which, if they agree, updates an incorrect recognition frequency of corresponding one of sets of post-processed character categories in said post-processed character set storage portion and which, if the incorrect recognition frequency exceeds a certain permitted number, produces the corresponding set of post-processed character categories; and

a character information-modifying portion for modifying information about reference sizes and reference positions of sets of post-processed character categories obtained from said incorrect recognition frequency-updating portion, using information about sizes and positions in said character pattern.

9. A character recognition machine in accordance claim 4, wherein said candidate character-selecting portion comprises:

a plurality of similarity degree-normalizing portions for normalizing degrees of similarity of character categories obtained from said single feature-recognizing portions with a maximum value of these degrees of similarity to transform them into normalized degrees of similarity;

a plurality of first candidate character-selecting portions for selecting a plurality of first candidate character categories from all character categories, using information about the sizes and positions in said character pattern and said normalized degrees of similarity;

a similarity degree-nonlinearly transforming portion for nonlinearly transforming the normalized degrees of similarity of said first candidate categories and dividing the nonlinearly transformed, normalized degrees of similarity by a sum of these degrees of similarity to find nonlinearly transformed degrees of similarity of said first candidate categories; and

a second candidate character-selecting portion for summing up nonlinearly transformed degrees of similarity of selected first candidate character categories for each same category so as to find integrated degrees of similarity of all the first candidate character categories, selecting plural first candidate character categories having larger integrated degrees of similarity as second candidate character categories, and producing these second candidate character categories and their integrated degrees of similarity.

13. A character recognition machine having a language processing portion capable of processing language, said language processing portion comprising:

a character string image storage means for storing an image input;

a character recognition means for outputting a set of candidate characters and an evaluation;

a word search means for receiving the set of candidate characters and determining a set of candidate words, using a word dictionary, based on a degree of similarity between the candidate characters and the characters in various character positions in the character string image;

a phrase search means for finding candidate phrases from the set of candidate words, using a grammatical dictionary;

a phrase evaluating value-calculating means for calculating correctness of each phrase in terms of vocabulary and grammar;

a phrase-selecting means for selecting one phrase, based on the evaluating values of phrases, and producing a modified character string;

a word incorrectness degree-calculating means for finding an incorrect recognition degree of each word in the modified character string, based on data about correctness percentage of each correct word and the evaluation made by said character recognition means;

a word incorrect amendment degree-calculating means for finding an incorrect amendment degree of each word in the modified character string; and

a reject processing means for determining characters to be rejected from the word incorrect amendment degree.

14. A character recognition machine having a language processing portion capable of processing language, said language processing portion comprising:

a character string image storage means for storing an image input;

a word search means for receiving a set of candidate characters and determining a set of candidate words, using a word dictionary, based on a degree of similarity between the candidate characters and the characters in various character positions in the character string image;

a phrase search means for finding candidate phrases from the set of candidate words, using a grammatical dictionary;

a phrase evaluating value-calculating means for calculating correctness of each phrase in terms of vocabulary and grammar;

a phrase-selecting means for selecting one phrase, based on evaluating values of phrases, and producing a modified character string;

a word incorrectness degree-calculating means for finding an incorrect recognition degree of each word in the modified character string, based on data about correctness percentage of each amended word and an evaluation made by a character recognition means;

a word incorrect amendment degree-calculating means for finding a word incorrect amendment degree of each word in the modified character string;

a reject processing means for determining characters to be rejected from the word incorrect amendment degree; and

a rejected character-replacing means for replacing the rejected characters in the modified character string by a character category having a maximum degree of similarity and outputted from said character recognition means.

15. A character recognition machine comprising:

a character string image storage portion, a character extraction portion, a character recognition portion, and a language processing portion; and

said character string image storage portion for storing a character string to be recognized as a character string image;

said character extraction portion connected with said character string image storage portion and comprising a plurality of operators arranged in a direction of said character string, each of said operators having a multiplicity of inputs and one outputs, the operators being interconnected so as to form a network, and

said character extraction portion acting to determine boundary positions of characters by the use of values of initialized processing parameters, to obtain a character pattern of each character and information about sizes and positions in this character pattern, to deliver these character patterns and information to said character recognition portion, to again process a text image in character positions specified by said character recognition portion or by said language processing portion with modified values of said processing parameters, to find new boundary positions of characters in said character positions, to obtain a new character pattern and new information about sizes and positions in this character pattern, and to again deliver these character pattern and information to said character recognition portion;

a character recognition portion for finding degrees of similarity of a character pattern applied from said character extraction portion, said degrees of similarity indicating degrees to which said character pattern belongs to various character categories, said character recognition portion acting to inform said character extraction portion of position of said character pattern in said character string if said degrees of similarity are not matched to information about sizes and positions in the character pattern applied from said character extraction portion, and if said degrees of similarity are matched to information about sizes and positions in the character pattern applied from said character extraction portion, said character recognition portion acting to deliver said degrees of similarity to said language processing portion, or to deliver degrees of similarity to the character category group restricted by said language processing portion for a character pattern in character positions specified by said language processing portion, and

to learn a character pattern as a teacher pattern if said language processing portion specifies a correct character category, said character pattern being in character positions of said specified correct character category, and

a language processing portion for receiving said degrees of similarity of character categories in various character positions found by said character recognition portion, selecting as a forecasted correct character category a combination of character categories which is correctest in terms of vocabulary and grammar, producing certain character categories and character positions corresponding to a certain character to said character recognition portion if categories of said certain forecasted correct character are restricted to said certain character category, producing said forecasted correct character and character positions corresponding to said forecested correct character to said character recognition portion if said forecasted correct character category differs from a character category applied from said character recognition portion and having a maximum degree of similarity, and producing character positions corresponding to a part to said character extraction portion if said part has no combination of character categories which seems to be correct in terms of vocabulary and grammar.

16. A character recognition machine in accordance with claim 15, where

said character extraction portion is a character extraction machine including:

a character string image input portion for receiving an image consisting of a character string;

a memory for holding the received character string;

a projection profile histogram-calculating portion for calculating a concentration value of histogram which is obtained by counting pixels in the character string;

an initial parameter-setting portion for receiving an output from said projection profile histogram-calculating portion and determining values of processing parameters from heights of the character string, lengths of spaces of projection profile histogram of the character string, and other factors which are obtained from the projection profile histogram, the parameters being used to estimate character positions;

a histogram distribution smoothing network portion formed by a network for receiving an output from said projection profile histogram-calculating portion and an output from said initial parameter-setting portion, said network being composed of interconnected operators each having a multiplicity of inputs and one output, said operators corresponding to positions of pixels in a direction of the character string,

and said histogram distribution smoothing network portion for acting to minimize a function which assumes its minimum value when values of the projection profile histogram of character portions have substantially the same value, to thereby smooth the projection profile histogram, thus reducing effects of noises present in the received character string;

an extraction position-estimating network portion formed by a second network for receiving an output from said projection profile histogram-calculating portion and an output from said initial parameter-setting portion, said second network being composed of interconnected operators each having a multiplicity of inputs and one output, said operators corresponding to positions of pixels in a direction of the character string, and said extraction position-estimating network portion for acting to minimize a function which assumes its minimum value when an extracted position of the whole character string has been optimally extracted, to thereby estimate an extraction position of the character;

an extraction position-determining portion for determining character extraction positions according to values outputted from said extraction position-estimating network portion;

a character pattern output portion for reading a character pattern of each character from said memory according to a signal outputted from said extraction position-determining portion;

a character recognition portion for recognizing what character is represented by each character pattern produced from said character pattern output portion;

a language processing portion for receiving an output from said character recognition portion and calculating an evaluation value of each phrase of the character string applied heretofore, said evaluation value indicating a degree of correctness in terms of vocabulary and grammar;

a final character position-determining portion for receiving an output from said language processing portion and determining whether characters in the phrase have been correctly extracted in terms of vocabulary and grammar and which, if the characters have not been correctly extracted, producing positions of a first character and a last character because probability of erroneous estimation of character positions in said phrase is high and which, if the characters have been correctly extracted, producing these character positions as final character extraction positions;

a re-extraction position-determining portion for receiving an output from said final character position-determining portion and determining a position at which a character position should be estimated again;

a character squareness degree-modifying portion for estimating a squareness degree from a character string which has been judged to have correct character positions by said final character position-determining portion; and

a parameter-modifying portion for receiving an output from said re-extraction position-determining portion and an output from said character squareness degree-modifying portion and modifying said processing parameters used to estimate the character positions.

17. A character recognition machine in accordance with claim 15, wherein

said character recognition portion is a pattern recognition machine including:

a plurality of feature-extracting portions, each of which is for finding each feature vector from an applied pattern;

a plurality of single feature-recognizing portions, each of which is for finding degrees of similarity of the applied pattern from each feature vector, the degrees of similarity indicating degrees to which the applied pattern belong to various categories;

a category-discerning portion for discerning the applied pattern using the degrees of similarity obtained by said plurality of single feature-recognizing portions;

each of said single feature-recognizing portions including;

a group dictionary storing a plurality of group references feature vectors representing category groups each consisting of a set of patterns having similar feature vectors;

a fuzzy rough classification portion for calculating group assignment degrees from the group reference feature vectors and from feature vectors of the applied pattern, the group assignment degree indicating degrees to which the applied pattern belong to the category groups, respectively;

a plurality of subclassification portions for finding inside group similarity degrees from the feature vectors of the applied pattern, the inside group similarity degrees indicating degrees to which the applied pattern belongs to categories contained in the category groups;

a group-selecting portion for selecting a plurality of category groups from the group assignment degrees;

a subclassification portion input signal-selecting portion for selecting the subclassification portion which can receive the feature vectors of the applied pattern according to information about selection of groups obtained from said group-selecting portion; and

a single feature similarity degree-calculating portion for finding degrees of similarity of the applied pattern to the categories from a group assignment degree of the category group selected by said group-selecting portion and from the inside group similarity degrees obtained by said selected subclassification portions;

said single feature similarity degree-calculating portion including;

a plurality of multipliers for multiplying the group assignment degree of the category group selected by said group-selecting portion by all inside group similarity degrees, respectively, obtained from said subclassification portions receiving the feature vectors of the applied pattern by said subclassification portion input signal-selecting portion; and

a category similarity degree-calculating portion for selecting plural output values having larger values from output values from the multipliers for each category and calculating a sum of these output values;

said category-discerning portion including;

a plurality of similarity degree-normalizing portions for normalizing the degrees of similarity of the categories obtained from said single feature-recognizing portions corresponding to the feature vectors, respectively, with a maximum one of these degrees of similarity to thereby transform the degrees of similarity into normalized degrees of similarity; and

a category decision portion for finding integrated degrees of similarity of the categories using the normalized degrees of similarity obtained by said similarity degree-normalizing portions and discerning the applied pattern.

18. A character recognition machine in accordance with claim 15, wherein

said character recognition portion is a character recognition machine including:

a plurality of feature-extracting portions, each of which is for finding each feature vector from a character pattern produced by a character extraction portion which divides a text image into plural regions for each character;

a plurality of feature-recognizing portions, each of which is for finding degrees of similarity from each feature vector, the degrees of similarity indicating degrees at which the character pattern belongs to various character categories, respectively;

a candidate character-selecting portion for finding a plurality of second candidate character categories and their degrees of similarity to the character pattern from the degrees of similarity obtained from said single feature-recognizing portions; and

a post-recognition processing portion for selecting a final character category from the second character candidate categories using information about sizes and positions of the character pattern;

each of said single feature-recognizing portions including;

a group dictionary storing a plurality of group reference feature vectors representing character category groups each consisting of a set of character pattern having similar feature vectors;

a fuzzy rough classification portion for calculating group assignment degrees from the group reference feature vectors and from feature vectors of the applied character pattern, the group assignment degree indicating degrees to which the applied character pattern belongs to the character category groups, respectively;

a plurality of subclassification portions for finding inside group similarity degrees from the feature vectors of the applied character pattern, the inside group similarity degrees indicating degrees to which the applied character pattern belongs to character categories contained in the character category groups;

a group-selecting portion for selecting a plurality of character category groups from the group assignment degrees;

a subclassification portion input signal-selecting portion for selecting the subclassification portion which can receive the feature vectors of the applied character pattern according to information about selection of groups obtained from said group-selecting portion; and

a single feature similarity degree-calculating portion for finding degrees of similarity of the applied character pattern to the character categories from a group assignment degree of the character category group selected by said group-selecting portion and from the inside group similarity degrees obtained by said selected subclassification portions;

said single feature similarity degree-calculating portion including;

a plurality of multipliers for multiplying the group assignment degree of the character category group selected by said group-selecting portion by all inside group similarity degrees, respectively, obtained from said subclassification portions receiving the feature vectors of the applied character pattern by said subclassification portion input signal-selecting portion; and

a category similarity degree-calculating portion for selecting plural output values having larger values from output values having larger values from output values from the multipliers for each character category and calculating a sum of these output values;

said candidate character-selecting portion including;

a plurality of similarity degree-normalizing portions for normalizing the degrees of similarity of various character categories obtained from said single feature-recognizing portions corresponding to the feature vectors, respectively, with a maximum one of these degrees of similarity to thereby transform the degrees of similarity into normalized degrees of similarity;

a plurality of first candidate character-selecting portions for selecting plural first candidate character categories from all character categories using information about sizes and positions in the character pattern and the normalized degrees of similarity; and

a second character-selecting portion for summing up all normalized degrees of similarity of the selected first candidate character categories for each same category, finding integrated degrees of similarity of all the first candidate character categories, selecting plural first candidate character categories having larger values as second candidate character categories, and delivering the second candidate character categories and their integrated degrees of similarity;

said post-recognition processing portion selects plural final candidate character categories and finds their degrees of similarity from the plural second candidate character categories by the use of information about sizes and positions in the character pattern and delivers these final candidate character categories and their degrees of similarity as values evaluating characters in the final candidates character categories to said language processing portion.

19. A character recognition machine in accordance with claim 15, wherein

said language processing portion is a language processing portion including:

a character string image storage means for storing an image input;

a character recognition means for outputting a set of candidate characters and an evaluation;

a word search means for receiving the set of candidate characters and determining a set of candidate words, using a word dictionary, based on a degree of similarity between the candidate characters and the characters in various character positions in the character string image;

a phrase search means for finding candidate phrases from the set of candidate words, using a grammatical dictionary;

a phrase evaluating value-calculating means for calculating correctness of each phrase in terms of vocabulary and grammar;

a phrase-selecting means for selecting one phrase, based on the evaluating values of phrases, and producing a modified character string;

a word incorrectness degree-calculating means for finding an incorrect recognition degree of each word in the modified character string, based on data about correctness percentage of each correct word and the evaluation made by said character recognition means;

a word incorrect amendment degree-calculating means for finding an incorrect amendment degree of each word in the modified character string; and

a reject processing means for determining characters to be rejected from the word incorrect amendment degree.

20. A character recognition machine having a language processing portion capable of processing language, said language processing portion comprising:

a character string image storage means for storing an image input;

a word search means for receiving a set of candidate characters and determining a set of candidate words, using a word dictionary, based on a degree of similarity between the candidate characters and the characters in various character positions in the character string image;

a phrase search means for finding candidate phrases from the set of candidate words, using a grammatical dictionary;

a phrase evaluating value-calculating means for calculating correctness of each phrase in terms of vocabulary and grammar;

a phrase-selecting means for selecting one phrase, based on the evaluating values of phrases, and producing a modified character string;

a word incorrectness degree-calculating means for finding an incorrect recognition degree of each word in the modified character string, based on data about correctness percentage of each correct word and an evaluation made by a character recognition portion;

a word incorrect amendment degree-calculating means for finding a word incorrect amendment degree of each word in the modified character string; and

a reject processing means for determining characters to be rejected from said word incorrect amendment degree and for outputting the rejected characters as misrecognized characters.

21. A character recognition machine having a language processing portion capable of processing language, said language processing portion comprising:

a character string image storage means for storing an image input;

a word search means for receiving a set of candidate characters and determining a set of candidate words, using a word dictionary, based on a degree of similarity between the candidate characters and the characters in various character positions in the character string image;

a phrase search means for finding candidate phrases from the set of candidate words, using a grammatical dictionary;

a phrase evaluating value-calculating means for calculating correctness of each phrase in terms of vocabulary and grammar;

a phrase-selecting means for selecting one phrase, based on evaluating values of phrases, and producing a modified character string;

a word incorrectness degree-calculating means for finding an incorrect recognition degree of each word in the modified character string, based on data about correctness percentage of each amended word and an evaluation made by a character recognition portion;

a word incorrect amendment degree-calculating means for finding a word incorrect amendment degree of each word in the modified character string;

a reject processing means for determining characters to be rejected from the word incorrect amendment degree;

a rejected character-replacing means for replacing the rejected characters in the modified character string by a character category having a maximum degree of similarity and outputted from said character recognition means; and

a re-extraction position-indicating means for producing a re-extraction position of the rejected characters to a character extraction means which divides a text image into plural regions for each character.