JP2011048499A - Recognition result correction device, image processor, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve accuracy of either character recognition or voice recognition without preparing a document set for recognition. <P>SOLUTION: A character recognition device 10 includes: a character recognition part 31 that performs character recognition on an image read by a character image reading part 21, and acquires candidate characters and character scores showing its certainty; a morphological analysis part 32 that performs morphological analysis on a character string where the candidate characters are arranged; a retrieval key creation part 33 that creates a retrieval key by connecting the candidate characters based on the result of the morphological analysis; a word evaluation part 34 that calculates retrieval scores by retrieving a document data storage part 41 whose selection has been accepted by a DB selection acceptance part 22 by using the retrieval key, and calculates word scores based on the retrieval scores; and a total evaluation part 35 that calculates total scores from the character stores and the word scores, and displays the total scores on a recognition result display part 23. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a recognition result correction device, an image processing device, and a program.

  In the character recognition device, a candidate character string is optimized using a character transition probability table that stores word identification information corresponding to a combination of a character having a transition relationship from one character to another character and a character having the transition relationship, A word dictionary storing hierarchical information indicating a hierarchical relationship between words with the word identification information for identification is searched based on the word identification information corresponding to the optimized candidate character string, and no input corresponding to the hierarchical information Is known (see, for example, Patent Document 1).

  Japan that characters of the same character type often appear continuously when outputting one of the character codes representing one or more characters that are highly similar to written data as a recognition result Also known is a technology that uses the general tendency of words to preferentially output character codes that represent characters that have the same character type as the character code that was output as the previous recognition result. (For example, refer to Patent Document 2).

  Words in the input sentence that are not registered in the dictionary using morpheme analysis means for obtaining an arbitrary number of morpheme analysis candidates in descending order of probability, word hypothesis generation means based on word models, and similar word search means based on similar word models There is also known a technique for correctly identifying the notation and part of speech, presenting the correct word even when the correct character is not included in the candidate character, and presenting the pair of the word sequence and the part of speech sequence in the most likely order (for example, And Patent Document 3).

  Both using a structured font representation and linear processing time expansion / contraction matching as an online discriminator that recognizes the time series pattern of writing points, and a modified secondary discriminant function as an offline discriminator that recognizes non-time series character image patterns In addition, a technique is also known in which the number of feature dimensions of the off-line pattern and the specific place of the modified secondary discriminant function are reduced, and further, context processing is used as post-processing (see Non-Patent Document 1, for example).

JP-A-9-305716 JP-A-8-180137 JP-A-8-315078

Hideto Oda, Ran Toki, Mototeru Onuma, Atsuko Tokuno, Seito Toshiro, Masaki Nakagawa, “Miniaturization of an online handwritten character classifier integrated with an offline classifier”, IEICE, September 1, 2007, Electronic Information IEICE Transactions D, Vol. J90-D, no. 9, p. 2583-2594

  An object of the present invention is to improve the accuracy of one of character recognition processing and voice recognition processing without preparing a document set for recognition processing.

The invention according to claim 1 is a first method for acquiring a certainty factor in the recognition process of a specific character included in a character string obtained as a result of the recognition process of any one of the character recognition process and the voice recognition process. An acquisition unit; a generation unit that generates a search word including the specific character included in the character string; and a character immediately before or immediately after the specific character included in the character string; and uses other than the recognition process A second acquisition unit that acquires an index relating to use of the search word by searching the document set prepared for the search using the search word generated by the generation unit; and the first acquisition A recognition result correcting apparatus comprising: correcting means for correcting the certainty factor of the specific character acquired by the means based on the index acquired by the second acquiring means.
The invention according to claim 2 further includes identification information receiving means for receiving identification information for identifying a unit that requires individual accuracy with respect to the recognition processing, and the second acquisition means is received by the identification information receiving means. 2. The recognition result correcting apparatus according to claim 1, wherein the document set associated in advance with the identification information is searched.
According to a third aspect of the present invention, the second acquisition unit obtains a plurality of document sets previously associated with the identification information received by the identification information reception unit using the search word. The index is acquired based on the frequency of use of the search term in each document set and the weight of each document set associated in advance with the identification information received by the identification information receiving means. The recognition result correcting apparatus according to claim 2.
According to a fourth aspect of the present invention, a confirmation instruction accepting unit that accepts an instruction for confirming the specific character included in the character string as a result of the recognition process; and a specification when the confirmation instruction accepting unit accepts the instruction And updating means for updating the weight of the specific document set based on the frequency of use of the search word obtained by searching the document set using the search word. Item 4. The recognition result correction device according to Item 3.
The invention according to claim 5 further includes selection instruction receiving means for receiving an instruction of a user who selects the document set, and the second acquisition means is selected by the instruction received by the selection instruction receiving means. 2. The recognition result correcting apparatus according to claim 1, wherein the document set is searched.
The invention according to claim 6 further comprises morpheme analysis means for performing morpheme analysis on the character string, and the generation means includes the morpheme analysis included in the search word based on a result of morpheme analysis by the morpheme analysis means. 6. The recognition result correcting apparatus according to claim 1, wherein a character immediately before or after a specific character is specified.
According to a seventh aspect of the present invention, there is provided a reading unit that reads an image from a recording medium on which the image is recorded, and a character string obtained as a result of character recognition performed on the image read by the reading unit. A first acquisition unit configured to acquire a certainty factor in the character recognition of the specific character included; the specific character included in the character string; and a character immediately before or immediately after the specific character included in the character string. Generating a search term including: and a search for a set of documents prepared for use other than the character recognition by using the search term generated by the generation unit. A second acquisition unit that acquires an index relating to use, and the certainty factor of the specific character acquired by the first acquisition unit is corrected based on the index acquired by the second acquisition unit Image processing, and a display unit that displays the specific character included in the character string based on the certainty factor of the specific character after correction by the correction unit. Device.
The invention according to claim 8 acquires a certainty factor in the recognition process of a specific character included in the character string obtained as a result of one of the character recognition process and the voice recognition process. A function, a function for generating a search term including the specific character included in the character string, and a character immediately before or immediately after the specific character included in the character string, and for uses other than the recognition process A function for obtaining an index relating to use of the search word by searching the document set prepared in the search word, and a function for correcting the certainty factor of the specific character based on the index It is a program for realizing.

According to the first aspect of the present invention, the accuracy of any one of the character recognition process and the voice recognition process can be improved without preparing a document set for the recognition process.
According to the second aspect of the present invention, the accuracy of the recognition process can be improved so as to meet the individual accuracy requirements regarding the recognition process.
According to the invention of claim 3, the accuracy of the recognition processing can be improved by using a plurality of document sets so as to meet individual accuracy requirements regarding the recognition processing.
According to the invention of claim 4, as the recognition process is repeated, the accuracy of the recognition process is improved.
According to the invention of claim 5, the accuracy of the recognition processing can be improved by using the document set freely selected by the user.
According to the sixth aspect of the present invention, it is possible to generate a search term used for improving the accuracy of the recognition processing without taking time and effort.
According to the invention of claim 7, the accuracy of character recognition can be improved without preparing a document set for character recognition.
According to the invention of claim 8, the accuracy of either the character recognition process or the voice recognition process can be improved without preparing a document set for the recognition process.

It is the block diagram which showed the function structural example of the character recognition apparatus in the 1st Embodiment of this invention. It is the flowchart which showed the operation example of the character recognition apparatus in the 1st Embodiment of this invention. It is the figure which showed the example of the character score calculated in the 1st Embodiment and 2nd Embodiment of this invention. It is the figure which showed the example of the search score calculated in the 1st Embodiment of this invention. It is the figure which showed the example of the total score calculated in the 1st Embodiment and 2nd Embodiment of this invention. It is the block diagram which showed the function structural example of the character recognition apparatus in the 2nd Embodiment of this invention. It is the figure which showed the example of the memory content of the DB weight holding | maintenance part in the 2nd Embodiment of this invention. It is the flowchart which showed the operation example of the character recognition apparatus in the 2nd Embodiment of this invention. It is the figure which showed the example of the search score calculated in the 2nd Embodiment of this invention. It is a hardware block diagram of the computer which can implement | achieve embodiment of this invention.

Embodiments of the present invention will be described below in detail with reference to the accompanying drawings.
In the present embodiment, when character recognition or voice recognition is performed, the accuracy of the recognition is simply improved. In order to improve the accuracy of such recognition, it is generally necessary to prepare an enormous amount of data such as N-gram probabilities, but in this embodiment, the results of character recognition and speech recognition are unnatural. If this happens, search the database prepared for uses other than character recognition and voice recognition, and correct the result without taking time (hereinafter, the database used for this search is referred to as “search target DB”). "). As described above, the present embodiment can be applied to both character recognition and voice recognition. However, in the following, a case where the present embodiment is applied to character recognition will be described as an example.

[First Embodiment]
In the first embodiment, the user explicitly selects the search target DB.
FIG. 1 is a block diagram illustrating a functional configuration example of the character recognition device according to the first embodiment.
As illustrated, the character recognition device 10 includes a UI (User Interface) unit 20, a processing unit 30, and a DB (Database) unit 40.
The UI unit 20 is a part that receives information input by the user and outputs information to the user, and includes a character image reading unit 21, a DB selection receiving unit 22, and a recognition result display unit 23. .
The processing unit 30 is a part that corrects the result of character recognition based on the information received by the UI unit 20 using the DB unit 40 and displays the result on the UI unit 20. The morphological analysis unit 32, the search key creation unit 33, the word evaluation unit 34, and the comprehensive evaluation unit 35 are provided.

  The DB unit 40 includes a database used by the processing unit 30, and includes document data storage units 41a, 41b, and 41c. Although the document data storage units 41a, 41b, and 41c are shown in the figure, they may be referred to as the document data storage unit 41 when it is not necessary to distinguish them. Although only three document data storage units 41 are shown in the figure, four or more document data storage units 41 may be provided. Furthermore, although the DB unit 40 is shown as a part of the character recognition device 10 in the figure, the DB unit 40 may exist outside the character recognition device 10. In the present embodiment, a DB unit 40 is provided as an example of a document set.

First, components of the UI unit 20 will be described.
The character image reading unit 21 reads a character image recorded on a recording medium such as paper. Here, the character image may be an image of a handwritten character handwritten with a writing instrument, or may be an image of a printed character printed with a printer. The character image reading unit 21 is, for example, a scanner. The character image reading unit 21 is a CCD system in which reflected light with respect to light emitted from a light source to a document is reduced by a lens and received by a CCD (Charge Coupled Devices), or an LED light source is irradiated on a document in order. It is preferable to use a CIS system in which reflected light with respect to the received light is received by a CIS (Contact Image Sensor). In the present embodiment, a character image reading unit 21 is provided as an example of a reading unit that reads an image.

The DB selection receiving unit 22 receives selection information indicating the document data storage unit 41 selected as a search target DB by the user prior to the search of the DB unit 40. In the present embodiment, a DB selection receiving unit 22 is provided as an example of a selection instruction receiving unit that receives an instruction of a user who selects a document set.
The recognition result display unit 23 displays the recognition result obtained by the processing unit 30. At this time, the recognition result may be displayed based on the score calculated by the processing unit 30. Here, as the recognition result display unit 23, for example, an LCD (Liquid Crystal Display) may be used. In the present embodiment, the recognition result display unit 23 is provided as an example of a display unit that displays specific characters based on the certainty after correction.

Next, components of the processing unit 30 will be described.
The character recognition unit 31 performs character recognition by cutting out a portion corresponding to each character from the character image read by the character image reading unit 21. In addition, as a character recognition result, for a predetermined number of candidate characters, a character code representing the candidate characters and a character score of the candidate characters are held. Here, the character score is a value indicating the certainty as a character recognition result of the candidate character. In addition, as a character recognition method in the character recognition unit 31, a method using an offline recognition technology for recognizing characters as a graphic pattern, or an online recognition technology using the information when time-series information is available. In many cases, a method of using the method is used, and a method of making an integrated judgment based on a combination of both is often used. In the present embodiment, a character score is used as an example of the certainty factor of a specific character, and the character recognition unit 31 is provided as an example of a first acquisition unit that acquires the certainty factor.

  The morpheme analysis unit 32 decomposes a character string made up of candidate characters extracted by character recognition by the character recognition unit 31 into a plurality of words for each part of speech, and gives part of speech information to each word. In the present embodiment, a morpheme analysis unit 32 is provided as an example of a morpheme analysis unit that performs morpheme analysis on a character string.

  The search key creation unit 33 determines a division unit based on the result of the morpheme analysis by the morpheme analysis unit 32, and generates a search key by combining candidate characters for each division unit. Here, if the word obtained by the morphological analysis unit 32 includes only one character, the division unit is determined by combining the character with the previous or subsequent word. At this time, it may be combined with either the previous or subsequent word, but is preferably combined with the word having the higher character score. Alternatively, the connection between specific parts of speech may be determined with priority, or may be determined based on the length of the word. In addition, search key generation is performed by selecting a predetermined number of combinations of candidate characters having a high character score, in order to prevent the number of search keys from becoming enormous by combining candidate characters, It may be performed by combining candidate characters whose score is higher than a preset threshold value. In the present embodiment, a search key is used as an example of a search term, and a search key creation unit 33 is provided as an example of a generation unit that generates a search term.

  The word evaluation unit 34 searches the document data storage unit 41 indicated by the selection information received by the DB selection reception unit 22 using the search key generated by the search key creation unit 33, and uses a score based on the search result. A certain search score is calculated. In this case, the number of documents hit by the search may be simply used as the search score, or the total appearance frequency of the search keys in the documents hit by the search may be used as the search score. And based on this search score, the word score which is a score as a word is calculated. In the present embodiment, a word score is used as an example of an index related to use of a search word, and a word evaluation unit 34 is provided as an example of a second acquisition unit that acquires the index. In this specification, the term “word score” is used for convenience, but the combination of candidate characters to which this score is attached does not necessarily constitute a grammatical word.

  The comprehensive evaluation unit 35 integrates the character score obtained by the character recognition by the character recognition unit 31 and the word score obtained by the search by the word evaluation unit 34, and finally obtains the character recognition result of each character image. decide. In the present embodiment, the comprehensive evaluation unit 35 is provided as an example of a correction unit that corrects the certainty factor of a specific character based on an index.

Furthermore, the components of the DB unit 40 will be described.
The document data storage unit 41 is a general document database that is pre-indexed so as to be searchable, and a database selected from these is a search target DB used by the processing unit 30. For example, the document database may be owned by an individual or an organization, or may be a set of web pages used in an Internet search engine.

Next, the operation of the character recognition device 10 in the first embodiment will be described.
FIG. 2 is a flowchart showing an operation example of the character recognition device 10 at this time.
First, when a user inputs a handwritten character image or a printed character image to the character recognition device 10, the character image reading unit 21 reads the character image (step 101).
When the user performs an operation of selecting one search target DB, the DB selection receiving unit 22 receives selection information indicating the selection content (step 102). In this case, as the search target DB, as described above, a general document database prepared for another use is used instead of the dictionary database for character recognition.

When the UI unit 20 acquires the information in this way, the UI unit 20 outputs the information to the processing unit 30.
Then, first, the character recognition unit 31 performs character recognition on the character image output from the character image reading unit 21 (step 103). Thereby, the character recognizing unit 31 obtains the character code and the character score indicating the certainty (confidence, likelihood) of the predetermined number of candidate characters. The character code and the character score are stored in a memory that can be referred to from each function of the processing unit 30.
Next, the morphological analysis unit 32 decomposes the character string made up of candidate characters obtained by the character recognition unit 31 into a plurality of words by morphological analysis (step 104). At this time, parts of speech information such as nouns, verbs, particles, and auxiliary verbs are attached to each word obtained by morphological analysis.

  Next, the search key creation unit 33 determines a division unit based on the result of the morpheme analysis by the morpheme analysis unit 32 (step 105). For example, among words obtained by morphological analysis, words attached with participle or part of speech information of auxiliary verbs are connected as a division unit if the preceding and following words are nouns, adjectives, and verbs. In other cases, words obtained by morphological analysis are used as division units as they are.

After that, the search key creation unit 33 creates a search key used for searching the search target DB for each division unit.
That is, the search key creation unit 33 first focuses on one division unit (step 106).
Next, a combination of candidate characters corresponding to each character included in the division unit of interest is acquired (step 107). Normally, all of the acquired combinations of candidate characters may be used as search keys.
However, in such a search key generation method, a huge number of search keys may be generated. Specifically, assuming that the number of candidate characters is m and the number of characters in the division unit is n, m-th n search keys are generated. Therefore, in order to avoid such a situation, only combinations of candidate characters whose average character score of candidate characters is larger than a predetermined threshold are used as search keys. Accordingly, the search key creation unit 33 refers to the character score stored in step 103 and determines whether the average value of the character scores of the candidate characters is greater than a predetermined threshold (step 108).

  As a result, if the average value of the character scores is equal to or less than the threshold value, the determination in step 108 is performed for other candidate character combinations. On the other hand, if the average value of the character scores is greater than the threshold value, the candidate character combination is used as a search key (step 109), and it is determined whether there are other candidate character combinations (step 110). Then, if there are other candidate character combinations, the processes of steps 107 to 109 are repeated. If there are no other candidate character combinations, the processing of the word evaluation unit 34 is performed.

  Then, the word evaluation unit 34 receives one or more search keys created by the search key creation unit 33, and uses this search key to store the document data storage unit 41 indicated by the selection information received by the DB selection reception unit 22. (Step 111). Then, based on the search result, a search score is calculated, and based on this search score, a word score indicating the probability as a word in which candidate characters are connected is calculated (step 112). Here, the search score may be calculated by, for example, the number of documents searched with the search key (number of search hits), the sum of the appearance frequencies of the search keys in the searched document, or the like.

Next, the comprehensive evaluation unit 35 integrates the character score obtained by the character recognition unit 31 in step 103 and the word score calculated by the word evaluation unit 34 in step 112 to obtain the final certainty for each candidate character. Is calculated (step 115). Then, the total score is sent to the recognition result display unit 23, and the recognition result display unit 23 adds the final rank to the candidate characters and displays them (step 116). As this rank, for example, it is conceivable to adopt a rank such that the higher the total score calculated in step 115, the higher the rank.
Furthermore, the comprehensive evaluation unit 35 determines whether there are other division units (step 117). If there is another division unit, the control is returned to the search key creation unit 33, and the processing of steps 106 to 116 is repeated. If there is no other division unit, the operation of the first embodiment ends.

  The total score calculated by the above operation example is expressed by the following equation.

_Tij is a total score of the jth candidate character of the ith character in a certain division unit.
C _ij is the character score of the j-th candidate character of the i-th character in a certain division unit.
S _ijk is a search score obtained when the search target DB is searched with the k-th search key including the j-th candidate character of the i-th character in a certain division unit. As S _ijk , the number of hits, the appearance frequency of search keys in all hit documents, etc. may be used. Further, even when the appearance frequency is 0, a prior frequency 1 or the like may be added to all search keys so that the search score does not become 0.
The above formula indicates that the overall score T _ij can be obtained by multiplying the character score C _ij by the word score that is the maximum value when k of S _ijk is varied.

Here, the operation example shown in FIG. 2 will be described using a specific example.
FIG. 3 is a diagram showing the character score C _ij calculated by the character recognition unit 31 in step 103 and passed to the search key creation unit 33 and the comprehensive evaluation unit 35 via a memory (not shown). However, here, it is assumed that attention is paid to a division unit consisting of only two characters in step 106. Such division units are based on the result of morphological analysis in step 104, but it is up to you to decide which candidate character character string is to be subjected to morphological analysis among candidate characters for each character. . For example, it is conceivable to perform a morphological analysis on a character string composed of candidate characters having the highest character score (in the example of the figure, a character string “... Shita ...”).
In the figure, for example, the first candidate character of the first character is “shi”, the character score C ₁₁ is 93, the second candidate character of the first character is “ki”, and the character score C ₁₂ is shown to be 90. The first candidate character of the second character is “da”, the character score C ₂₁ is 95, the second candidate character of the second character is “old”, and the character score C ₂₂ is 75. It is shown that there is.

FIG. 4 is a diagram showing the search score S _ijk calculated by the word evaluation unit 34 in step 112. However, the threshold used in step 108 is 80, and the search score is shown for a combination of candidate characters whose average character score exceeds this threshold.
For example, among the search keys including the first candidate character “shi” of the first character, those having an average character score of more than 80 are “shida” and “shiji”. Therefore, in the figure, a search score S ₁₁₁ obtained by searching the search target DB using “Shida” as a search key, and a search score obtained by searching the search target DB using “Shi old” as a search key. S ₁₁₂ is shown. Maximum a is search score S ₁₁₁ in this and search score S ₁₁₁ search score S ₁₁₂ as a word score, it is passed to the comprehensive evaluation unit 35 via a memory (not shown). Similarly, among the search key including the second candidate character “K”, the search key including the third candidate character “Sat”, and the search key including the fourth candidate character “K”, the average value of the character scores is A search score when using search keys exceeding 80 is also shown. Since there is no search key including the fifth candidate character “yu” and the average character score exceeds 80, a field for storing the search score when the search key including the candidate character “yu” is used. Is not provided.

Of the search keys that include the first candidate character “da” of the second character, those with an average character score exceeding 80 are “Shida”, “Kita”, “Tsuchida”, and “ Rice field ". Therefore, in the figure, the search target DB search search score S ₂₁₁ obtained as a search key "Sita", "Kita" the search key as the search target DB search was obtained search score S ₂₁₂ , a search score S ₂₁₃ obtained by searching the search target DB using “Tsuchida” as a search key, and a search score S ₂₁₄ obtained by searching the search target DB using “Kuda” as a search key It is shown. As search score _{S 213} the word score is the largest among the search score _{S 211} and the search score _{S 212} and the search score _{S 213} and the search score _{S 214,} is passed to the comprehensive evaluation unit 35 via a memory (not shown). Similarly, the search score when the search key including the second candidate character “Old” using the search key with an average value of the character score exceeding 80 is also shown. It should be noted that a search key including the third candidate character “mouth”, a search key including the fourth candidate character “ten”, a search key including the fifth candidate character “X”, and an average character score of 80 Since there is nothing exceeding, there is not provided a column for storing a search score when a search key including candidate characters “mouth”, “ten”, and “X” is used.

FIG. 5 is a diagram showing the total score T _ij calculated by the comprehensive evaluation unit 35 in step 115. Here, as described above, T _ij is obtained by multiplying C _ij by the word score which is the maximum value when k of S _ijk is varied.
For example, as described above, the word score that is the maximum value of the search score when using the search key including the first candidate character “shi” of the first character is S ₁₁₁ = 55. Therefore, the figure shows that the total score T ₁₁ is 5115 (T ₁₁ = C ₁₁ × maxS _11K = 93 × 55 = 5115). Similarly, the overall score is shown for the second candidate character “K”, the third candidate character “Sat”, and the fourth candidate character “K”. In addition, since S _15K is not obtained for the fifth candidate character “Yu”, C ₁₅ is set as T ₁₅ as it is.

Further, as described above, the word score that is the maximum value of the search score when the search key including the first candidate character “field” of the second character is used is S ₂₁₃ = 2150. Therefore, in the figure, the total score _{T 21} is shown to be _{204250 (T 21 = C 21 ×} maxS 21K = 95 × 2150 = 204250). Similarly, the overall score is also shown for the second candidate character “Old”. Since S _2jk is not obtained for the third candidate character “mouth”, the fourth candidate character “ten”, and the fifth candidate character “X”, C _2j is used as T _2j as it is (j = 3,4,5).

Thus, for the first character of the candidate character, become a total score T ₁₃ is the largest of the "soil", for the second character of the candidate character, the overall score T ₂₁ of the "field" is the largest. Therefore, it can be seen that the character recognition result in this division unit has the highest certainty of the result “Tsuchida”.

In the first embodiment, the user specifies one search target DB, but a plurality of search target DBs may be specified.
In the first embodiment, the character string is automatically decomposed into groups of parts of speech by morphological analysis. However, the user may set which part of the input character string is used as the search key.
Furthermore, in this first embodiment, a search key comprising a combination of candidate characters whose average character score exceeds the threshold value is generated. However, only a preset number of candidates are selected from those having a large average character score. A search key may be generated by selecting a combination of characters.

[Second Embodiment]
In 1st Embodiment, since a user must select search object DB, we are anxious also about taking time. In particular, when one database must be selected from a large number of databases, it is preferable to assist the user or automatically make a selection suitable for the user.
Therefore, in the second embodiment, the character recognition device 10 automatically selects a search target DB suitable for calculating a word score.

FIG. 6 is a block diagram illustrating a functional configuration example of the character recognition device according to the second embodiment.
As illustrated, the character recognition device 10 includes a UI (User Interface) unit 20, a processing unit 30, and a DB (Database) unit 40.
The UI unit 20 is a part that receives information input by the user and outputs information to the user. The UI unit 20 is a character image reading unit 21, a recognition result display unit 23, a user ID reception unit 24, and a confirmation. And an instruction receiving unit 25.
The processing unit 30 is a part that corrects the result of character recognition based on the information received by the UI unit 20 using the DB unit 40 and displays the result on the UI unit 20. A morpheme analysis unit 32, a search key creation unit 33, a word evaluation unit 34, a comprehensive evaluation unit 35, a DB weight holding unit 36, and a DB weight calculation unit 37.

  The DB unit 40 includes a database used by the processing unit 30, and includes document data storage units 41a, 41b, and 41c. Although the document data storage units 41a, 41b, and 41c are shown in the figure, they may be referred to as the document data storage unit 41 when it is not necessary to distinguish them. Although only three document data storage units 41 are shown in the figure, four or more document data storage units 41 may be provided. Furthermore, although the DB unit 40 is shown as a part of the character recognition device 10 in the figure, the DB unit 40 may exist outside the character recognition device 10. In the present embodiment, a DB unit 40 is provided as an example of a document set.

First, components of the UI unit 20 will be described. However, since the character image reading unit 21 and the recognition result display unit 23 are the same as those in the first embodiment, description thereof is omitted here.
The user ID reception unit 24 receives a user ID as information that is a clue for the processing unit 30 to automatically select a search target DB. In the present embodiment, a user ID receiving unit 24 is provided as an example of an identification information receiving unit that receives identification information.
The confirmation instruction receiving unit 25 receives an instruction input for finally confirming the recognition result displayed by the recognition result display unit 23. In the present embodiment, a confirmation instruction receiving unit 25 is provided as an example of a confirmation instruction receiving unit that receives an instruction to fix a specific character as a result of recognition processing.

Next, components of the processing unit 30 will be described. However, since the character recognition unit 31, the morphological analysis unit 32, the search key creation unit 33, and the comprehensive evaluation unit 35 are the same as those in the first embodiment, description thereof is omitted here.
The word evaluation unit 34 selects a search target DB by referring to the DB weight holding unit 36 using the user ID received by the user ID reception unit 24 as a key, and sets a search key search score for each selected search target DB. calculate. Then, the DB weight holding unit 36 is referred to weight the search score to calculate the word score. Here, the weighting is performed, for example, by taking a weighted average. In the present embodiment, a search score is used as an example of the usage frequency of the search word.

The DB weight holding unit 36 accumulates and holds the weight given to the database by the DB weight calculation unit 37. For example, the score given by the DB weight calculation unit 37 is added and held.
The DB weight calculation unit 37 calculates the database weight so that a search target DB having a word score suitable for the user is selected. This weight is linked to the confirmation instruction accepting unit 25, and a weighting function such as assigning a score to a database in which many of the same characters or character strings that the user has confirmed as the final character recognition result is found. have. In the present embodiment, a DB weight calculation unit 37 is provided as an example of an updating unit that updates the weight of a specific document set.

  Since the components of the DB unit 40 are the same as those in the first embodiment, description thereof is omitted here.

Here, specific contents of the DB weight holding unit 36 will be described.
FIG. 7 is a diagram showing an example of information held by the DB weight holding unit 36.
As shown in the figure, in the DB weight holding unit 36, a search target DB is set together with the weight of each database for each user.
For example, when the user with the user ID “U01” instructs character recognition, DB # 1 and DB # 2 are used as search target DBs, and the search score obtained by searching DB # 1 is doubled. The weight is set to be given to the search score obtained by searching DB # 2.

Next, the operation of the character recognition device 10 in the second embodiment will be described.
FIG. 8 is a flowchart showing an operation example of the character recognition device 10 at this time.
First, when a user inputs a handwritten character image or a printed character image into the character recognition device 10, the character image reading unit 21 reads the character image (step 151).
Further, when the user performs an operation for inputting a user ID, the user ID receiving unit 24 receives the input user ID (step 152).

When the UI unit 20 acquires the information in this way, the UI unit 20 outputs the information to the processing unit 30.
Then, first, the character recognition unit 31 performs character recognition on the character image output from the character image reading unit 21 (step 153). Thereby, the character recognizing unit 31 obtains the character code and the character score indicating the certainty (confidence, likelihood) of the predetermined number of candidate characters. The character code and the character score are stored in a memory that can be referred to from each function of the processing unit 30.
Next, the morphological analysis unit 32 decomposes the character string made up of candidate characters obtained by the character recognition unit 31 into a plurality of words by morphological analysis (step 154). At this time, parts of speech information such as nouns, verbs, particles, and auxiliary verbs are attached to each word obtained by morphological analysis.

  Next, the search key creation unit 33 determines a division unit based on the result of the morpheme analysis by the morpheme analysis unit 32 (step 155). For example, among words obtained by morphological analysis, words attached with participle or part of speech information of auxiliary verbs are connected as a division unit if the preceding and following words are nouns, adjectives, and verbs. In other cases, words obtained by morphological analysis are used as division units as they are.

After that, the search key creation unit 33 creates a search key used for searching the search target DB for each division unit.
That is, the search key creation unit 33 first focuses on one division unit (step 156).
Next, a combination of candidate characters corresponding to each character included in the division unit of interest is acquired (step 157). Normally, all of the acquired combinations of candidate characters may be used as search keys.
However, in such a search key generation method, a huge number of search keys may be generated. Specifically, assuming that the number of candidate characters is m and the number of characters in the division unit is n, m-th n search keys are generated. Therefore, in order to avoid such a situation, only combinations of candidate characters whose average character score of candidate characters is larger than a predetermined threshold are used as search keys. Accordingly, the search key creation unit 33 refers to the character score stored in step 153 and determines whether the average value of the character scores of the candidate characters is greater than a predetermined threshold (step 158).

  As a result, if the average value of the character scores is equal to or less than the threshold value, the determination in step 158 is performed for other candidate character combinations. On the other hand, if the average value of the character scores is larger than the threshold value, the candidate character combination is used as a search key (step 159), and it is determined whether there are other candidate character combinations (step 160). If there are other candidate character combinations, the processes of steps 157 to 159 are repeated. If there are no other candidate character combinations, the process proceeds to the word evaluation unit 34.

  Then, the word evaluation unit 34 receives one or more search keys created by the search key creation unit 33 and uses the search key to add the user ID received by the user ID reception unit 24 to the DB weight holding unit 36. The associated document data storage unit 41 is searched as a search target DB (step 161). Then, a search score is calculated based on the search result (step 162). Here, the search score may be calculated by, for example, the number of documents searched with the search key (number of search hits), the sum of the appearance frequencies of the search keys in the searched document, or the like.

  Thereafter, the word evaluation unit 34 determines whether there is another document data storage unit 41 associated with the user ID received by the user ID reception unit 24 in the DB weight holding unit 36 (step 163). If there is another document data storage unit 41, the processing of steps 161 to 162 is repeated. If there is no other document data storage unit 41, the word evaluation unit 34 calculates the word score by taking the weighted average of the search scores with the weight for each DB held in the DB weight holding unit 36. (Step 164).

Next, the comprehensive evaluation unit 35 integrates the character score obtained by the character recognition unit 31 in step 153 and the word score calculated by the word evaluation unit 34 in step 164 to obtain the final certainty for each candidate character. Is calculated (step 165). Then, the total score is sent to the recognition result display unit 23, and the recognition result display unit 23 adds the final ranking to the candidate characters and displays them (step 166). As this rank, for example, it is conceivable to adopt a rank such that the higher the total score calculated in step 165, the higher the rank.
Furthermore, the comprehensive evaluation unit 35 determines whether there are other division units (step 167). If there is another division unit, the control is returned to the search key creating unit 33, and the processes of steps 156 to 166 are repeated. If there is no other division unit, the process proceeds to a process of correcting the DB weight.

That is, when it is determined that the candidate characters displayed by the recognition result display unit 23 and their ranks are valid, the user inputs a confirmation instruction, and the confirmation instruction receiving unit 25 receives the input of the confirmation instruction (step 168).
Then, the fact that the input of the confirmation instruction has been accepted is transmitted from the confirmation instruction receiving unit 25 to the DB weight calculation unit 37, and the DB weight calculation unit 37 is held by the DB weight holding unit 36 based on the confirmation instruction. The weight of each database is changed (step 169), and the operation of the second embodiment ends.

  The total score calculated by the above operation example is expressed by the following equation.

_Tij is a total score of the jth candidate character of the ith character in a certain division unit.
C _ij is the character score of the j-th candidate character of the i-th character in a certain division unit.
W _m is a weight assigned to the m-th search target DB.
S _ijkm is a search score obtained when the mth search target DB is searched with the kth search key including the jth candidate character of the ith character in a certain division unit. As S _ijkm , the number of hits, the appearance frequency of search keys in all hit documents, and the like may be used. Further, even when the appearance frequency is 0, the prior frequency 1 or the like may be added to all the search keys so that the word score does not become 0.
Then, the above formula is obtained by multiplying the character score C _ij by the word score which is the maximum value when k of the total sum relating to m of (W _m × S _ijkm ) is changed to obtain the total score T _ij. It is shown that.

Here, the operation example shown in FIG. 8 will be described using a specific example.
First, the character score C _ij calculated by the character recognition unit 31 in step 153 and passed to the search key creation unit 33 and the comprehensive evaluation unit 35 via a memory (not shown) is the same as that shown in FIG.

FIG. 9 is a diagram illustrating the search score S _ijkm calculated by the word evaluation unit 34 in step 162. However, the threshold used in step 158 is 80, and the search score is shown for a combination of candidate characters whose average character score exceeds this threshold. Here, assuming that the user with the user ID “U01” shown in FIG. 7 instructs character recognition, DB # 1 and DB # 2 are used as search target DBs. Accordingly, S _ijk1 and S _ijk2 are calculated as search scores.

For example, among the search keys including the first candidate character “shi” of the first character, those having an average character score of more than 80 are “shida” and “shiji”. Therefore, in the figure, a search score S ₁₁₁₁ obtained by searching DB # 1 using “Sida” as a search key, and a search score obtained by searching DB # 2 using “Sida” as a search key. S ₁₁₁₂ , a search score S ₁₁₂₁ obtained by searching DB # 1 using “shiji” as a search key, and a search score S ₁₁₂₂ obtained by searching DB # 2 using “shiji” as a search key Is shown. Up to the first weighted average and that sum by weighting the search score _{S 1111} a search score _{S 1112,} in the second weighted average sum by weighting the search score _{S 1121} a search score _{S 1122} A certain first weighted average is passed as a word score to the comprehensive evaluation unit 35 via a memory (not shown). Similarly, among the search key including the second candidate character “K”, the search key including the third candidate character “Sat”, and the search key including the fourth candidate character “K”, the average value of the character scores is A search score when using search keys exceeding 80 is also shown. Since there is no search key including the fifth candidate character “yu” and the average character score exceeds 80, a field for storing the search score when the search key including the candidate character “yu” is used. Is not provided.

Of the search keys that include the first candidate character “da” of the second character, those with an average character score exceeding 80 are “Shida”, “Kita”, “Tsuchida”, and “ Rice field ". Therefore, in the figure, the search score S ₂₁₁₁ obtained by searching the DB # 1 as a search key "Sita", "Sita" search key as DB # 2 search-obtained search score S ₂₁₁₂ , a search score S ₂₁₂₁ obtained by searching DB # 1 using “Kida” as a search key, and a search score S ₂₁₂₂ obtained by searching DB # 2 using “Kida” as a search key And a search score S ₂₁₃₁ obtained by searching DB # 1 using “Tsuchida” as a search key, a search score S ₂₁₃₂ obtained by searching DB # 2 using “Tsuchida” as a search key, A search score S ₂₁₄₁ obtained by searching DB # 1 using “Ta” as a search key and a search score S ₂₁₄₂ obtained by searching DB # 2 using “Koda” as a search key are shown. . First weighted average and the sum are weighted the search score _{S 2111} a search score _{S 2112,} a second weighted average sum by weighting the search score _{S 2121} a search score _{S 2122,} the search score S third weighted average and the sum are weighted ₂₁₃₁ with search score _{S 2132,} the third is the fourth largest of the weighted average of the sum by weighting the search score _{S 2141} a search score _{S 2142} The weighted average is passed as a word score to the comprehensive evaluation unit 35 via a memory (not shown). Similarly, the search score when the search key including the second candidate character “Old” using the search key with an average value of the character score exceeding 80 is also shown. It should be noted that a search key including the third candidate character “mouth”, a search key including the fourth candidate character “ten”, a search key including the fifth candidate character “X”, and an average character score of 80 Since there is nothing exceeding, there is not provided a column for storing a search score when a search key including candidate characters “mouth”, “ten”, and “X” is used.

In FIG. 9, the value of the search score _{S Ijk1} obtained by searching the DB # 1, and, as the value of the search score _{S Ijk2} obtained by searching the DB # 2, the user ID in FIG. 7 A value that is equal to S _{ijk in} FIG. 4 is illustrated by taking a weighted average based on the weight for each DB set for the user of “U01”.
Therefore, the total score T _ij calculated by the total evaluation unit 35 in step 165 is the same as that shown in FIG.

The correction of the DB weight in step 168 will be described.
For example, it is assumed that the user gives an instruction to confirm “Tsuchida” as a character recognition result in step 167. In this case, it can be seen in FIG. 9 that many “Tsuchida” are found in DB # 2.
On the other hand, referring to FIG. 7, the weight of DB # 1 is 1 and the weight of DB # 2 is 2 for the user with the user ID “U01”. In such a case, the weight of DB # 2 in which more words finally determined by the user are found is changed to a value larger than 2. Here, as to how much the weight is to be increased, it is preferable to set a reference in advance and make it based on the reference.

  In the second embodiment, in step 161, the database having a weight other than 0 in FIG. 7 is determined as the search target DB. However, the present invention is not limited to this. For example, it is also conceivable that a threshold value for weight is set and only a database having a weight exceeding this threshold value is set as a search target DB even if the database has a weight other than zero. Alternatively, for example, databases may be selected from a large weight to a predetermined maximum number, and this may be used as a search target DB. Further, the character recognition device 10 may present a database to be prioritized so as to assist the user's selection instead of determining the search target DB regardless of the user's intention.

  In this embodiment, a search target DB and a weight are defined for the user ID, and a word score is calculated using the search target DB and the weight corresponding to the user ID of the user who has instructed character recognition. . This is because the possibility of covering vocabulary and the like unique to the user is increased, and an improvement in character recognition accuracy can be expected. However, in addition to the user, for example, an organization, a document, a document type, and the like can be considered as a unit for requesting individual accuracy regarding character recognition. That is, the search target DB and weight are defined for the organization, and the word score is calculated using the search target DB and the weight corresponding to the organization to which the user who instructed character recognition belongs, Alternatively, a configuration may be employed in which the search target DB and the weight are defined, and the word score is calculated using the search target DB and the weight corresponding to the character recognition target document or document type.

By the way, the correction process of the character recognition result in the present embodiment may be realized by a general-purpose computer. Therefore, the hardware configuration will be described assuming that this processing is realized by the computer 90.
FIG. 10 is a diagram illustrating a hardware configuration of the computer 90.
As shown in the figure, the computer 90 includes a CPU (Central Processing Unit) 91 as a calculation means, a main memory 92 as a storage means, and a magnetic disk device (HDD: Hard Disk Drive) 93. Here, the CPU 91 executes various types of software such as an OS (Operating System) and applications to realize the above-described functions. The main memory 92 is a storage area for storing various software and data used for execution thereof, and the magnetic disk device 93 is a storage area for storing input data for various software, output data from various software, and the like. .
Further, the computer 90 includes a communication I / F 94 for performing communication with the outside, a display mechanism 95 including a video memory and a display, and an input device 96 such as a keyboard and a mouse.

  The program for realizing this embodiment can be provided not only by communication means but also by storing it in a recording medium such as a CD-ROM.

DESCRIPTION OF SYMBOLS 10 ... Character recognition apparatus, 20 ... UI part, 30 ... Processing part, 40 ... DB part

Claims (8)

First acquisition means for acquiring a certainty factor in the recognition process of a specific character included in the character string obtained as a result of the recognition process of any one of the character recognition process and the voice recognition process;
Generating means for generating a search word including the specific character included in the character string and a character immediately before or after the specific character included in the character string;
Second acquisition means for acquiring an index relating to use of the search word by searching a set of documents prepared for a purpose other than the recognition processing using the search word generated by the generation means; ,
A recognition result comprising correction means for correcting the certainty factor of the specific character acquired by the first acquisition means based on the index acquired by the second acquisition means Correction device. An identification information receiving means for receiving identification information for identifying a unit that requires individual accuracy with respect to the recognition processing;
The recognition result correcting apparatus according to claim 1, wherein the second obtaining unit searches the document set associated in advance with the identification information received by the identification information receiving unit.   The second acquisition means includes the search word in each document set obtained by searching the plurality of document sets previously associated with the identification information received by the identification information receiving means using the search word. The recognition result according to claim 2, wherein the index is acquired based on a use frequency of the document and a weight of each document set associated in advance with the identification information received by the identification information receiving unit. Correction device. A confirmation instruction accepting unit that accepts an instruction to confirm the specific character included in the character string as a result of the recognition process;
When the confirmation instruction receiving unit receives the instruction, the weight of the specific document set is updated based on the use frequency of the search word obtained by searching the specific document set using the search word. The recognition result correcting apparatus according to claim 3, further comprising update means. A selection instruction receiving means for receiving an instruction of a user for selecting the document set;
The recognition result correcting apparatus according to claim 1, wherein the second acquisition unit searches the document set selected by the instruction received by the selection instruction receiving unit. It further comprises morphological analysis means for performing morphological analysis on the character string,
The said generation means specifies the character immediately before or after the said specific character included in the said search word based on the result of the morphological analysis by the said morpheme analysis means, The one in any one of Claim 1 thru | or 5 characterized by the above-mentioned. The recognition result correction device described. Reading means for reading the image from the recording medium on which the image is recorded;
First acquisition means for acquiring a certainty factor in character recognition of a specific character included in a character string obtained as a result of performing character recognition on the image read by the reading means;
Generating means for generating a search word including the specific character included in the character string and a character immediately before or after the specific character included in the character string;
Second acquisition means for acquiring an index relating to use of the search word by searching a document set prepared for use other than the character recognition using the search word generated by the generation means; ,
Correction means for correcting the certainty factor of the specific character acquired by the first acquisition means based on the index acquired by the second acquisition means;
An image processing apparatus comprising: a display unit configured to display the specific character included in the character string based on the certainty factor of the specific character after correction by the correction unit. On the computer,
A function of acquiring a certainty factor in the recognition process of a specific character included in the character string obtained as a result of the recognition process of any one of the character recognition process and the voice recognition process;
A function of generating a search term including the specific character included in the character string and a character immediately before or after the specific character included in the character string;
A function for obtaining an index relating to use of the search word by searching a set of documents prepared for use other than the recognition process using the search word;
A program for realizing a function of correcting the certainty factor of the specific character based on the index.

Images