JP2011065646A - Apparatus and method for recognizing character string - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently recognize a character string constituted of various types of characters. <P>SOLUTION: A method for recognizing a character string includes the steps of: dividing a character string image into a plurality of segments; performing OCR recognition on the plurality of segments to acquire candidate characters; acquiring statistical information of candidate characters and/or statistical information of character combinations formed from the candidate characters; and merging the statistical information with reliability of OCR recognition on the candidate characters to decide a candidate character string. An apparatus for recognizing a character string is also disclosed. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a character string recognition device and a character string recognition method for recognizing a character string from a character string image.

  Currently, it is common to recognize various types of character information using OCR technology. For example, after a user writes a series of characters on paper or a touch panel, the written character string is converted into a character string image by scanning, photographing, magnetic induction, or the like. Then, the recognition system recognizes the converted character string image and outputs the recognized character string.

  A character string composed of various characters is input to the recognition system as an image. For example, the character string includes a character string only, a character string mixed with a numeral, a character string including “.” As a delimiter (delimiter), a character string including “@” as a delimiter, and the like.

  In addition, with the rapid development of the Internet, when information is transmitted, a character string composed of various characters is often used. In particular, the recognition system recognizes a handwritten character string. Is becoming more important. For example, the recognition system may recognize an Email address handwritten on the touch panel by the user.

JP-A-9-330377 JP 2002-183664 A JP 2010-140162 A Chinese Patent Application No. 2008080080950 Specification

W.K. Pratt, John Wiley, Sons Inc, Digital Image Processing (4th Edition), 2007 Danian Zheng, “Handwritten Email Address Recognition with Syntax and Lexicons”, ICFHR2008 F. Kimura, K. Takashina, S. Tsuruoka, Y. Miyake, “Modified Quadratic Discriminant Functions and the Application to Chinese Character Recognition”, IEEE Trans, Pattern Analysis and Machine Intelligence, January 1987, Vol. 9, Vol. No., p. 149-153

  However, the above-described conventional technique has a problem that it cannot efficiently recognize a character string composed of various characters.

  The disclosed technology has been made in view of the above, and an object thereof is to provide an apparatus and a method for recognizing a character string. The outline of the disclosed technology will be described below so as to deepen the basic understanding of the disclosed technology. However, this summary description does not fully describe the disclosed technology. Further, it does not limit key points or important parts of the disclosed technology, and does not limit the scope of the disclosed technology. However, by giving the concept in a concise form, it is used as a front part of a more detailed description to be described later.

  According to one aspect of the disclosed technology, the character string recognition device divides the character string image into a plurality of segments, performs OCR recognition on the plurality of segments, acquires candidate characters, Characters including a step of acquiring statistical information and / or statistical information of a character combination formed by candidate characters, and a step of determining candidate character strings by integrating statistical information and reliability of OCR recognition of candidate characters A column recognition method is provided.

  According to another aspect of the disclosed technology, a segment division unit that divides a character string image into a plurality of segments, an OCR recognition unit that performs OCR recognition on the plurality of segments and obtains candidate characters, and candidate character statistics A statistical information acquisition unit that acquires statistical information of a character combination formed by information and / or candidate characters, and a first character that determines the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate characters A character string recognizing device including a column determining unit is provided.

  According to an embodiment of the disclosed technology, a computer program for realizing a character string recognition method is provided.

  In addition, according to an embodiment of the disclosed technology, a computer program product for realizing a character string recognition method is recorded, and a computer program product in a computer-readable form is provided.

  According to one aspect of the character string recognition device and the character string recognition method disclosed in the present application, there is an effect that a character string composed of various characters can be efficiently recognized.

FIG. 1 is a flowchart schematically showing a method according to the first embodiment of the disclosed technique. FIG. 2 is a flowchart schematically showing a method according to the second embodiment of the disclosed technique. FIG. 3 is a diagram illustrating an example of a character string recognition result. FIG. 4 is a diagram showing a sorting result after sorting the result of OCR recognition on the right side, and a corresponding path on the left side. FIG. 5 is a diagram showing a general syllable composed of two characters and / or a syllable composed of three characters. FIG. 6 is a diagram showing syllable simulation and statistical results. FIG. 7 is a flowchart schematically showing a method according to the third embodiment of the disclosed technique. FIG. 8 is a diagram showing a statistical result of a three-dimensional set composed of numbers and letters. FIG. 9 is a flowchart schematically showing a method according to the fourth embodiment of the disclosed technique. FIG. 10 is a flowchart schematically showing a method of correcting the OCR recognition result using the already grasped database according to the fifth embodiment of the disclosed technology. FIG. 11 is a flowchart schematically showing a method according to the sixth embodiment of the disclosed technique. FIG. 12 is a diagram illustrating a character string recognition apparatus according to a seventh embodiment of the disclosed technology. FIG. 13 is a diagram illustrating a character string recognition apparatus according to an eighth embodiment of the disclosed technology. FIG. 14 is a diagram illustrating a character string recognition apparatus according to a ninth embodiment of the disclosed technology. FIG. 15 is a diagram illustrating a delimiter recognition unit of the character string recognition apparatus according to the tenth embodiment of the disclosed technique. FIG. 16 is a diagram illustrating a character string recognition apparatus according to an eleventh embodiment of the disclosed technology. FIG. 17 is a block diagram illustrating a computer that implements an embodiment of the disclosed technology.

  Hereinafter, typical examples of the disclosed technology will be described with reference to the drawings. In the description, not all features of actual embodiments are described in the specification for clarity and brevity. In the process of putting any example into practical use, a decision specific to the embodiment to be put into practice is made so as to realize the specific purpose of the developer. Also, such determinations are different for different embodiments.

  The disclosed technology can be better understood in the following part by referring to the description through the drawings. In all the drawings, the same or similar parts are indicated by the same or similar reference numerals. The drawings are included in and constitute a part of this specification with the following detailed description, and further illustrate the preferred embodiment of the invention by way of example and to interpret the operation and effect of the invention. Used for.

  Here, in order to prevent the disclosed technology from becoming unclear in the drawings, only the structure and / or processing steps of the apparatus that are directly related to the technical idea related to the disclosed technology are shown, and others that are not directly related to the disclosed technology. It is supplemented that the detailed explanation is omitted.

  Hereinafter, the disclosed technology will be described in detail based on specific examples.

  A general OCR recognition technique itself is a well-known technique, but the process of OCR recognition will be briefly described for convenience of explanation and understanding later.

  The input character string image to be recognized may be a character string image obtained by being written on paper or a touch panel by a user and then converted by scanning, photographing, magnetic induction, or the like. For example, an image of a character string in which “hanashiro” is input by handwriting on the touch panel may be used.

  Usually, after the input character string image is acquired, preprocessing, for example, binarization processing is performed on the character string image input by handwriting. For example, a simulated image, a color digital image, and a gradation level image are converted into a binarized image. Thereafter, the connected region is analyzed. A connected area is an image area formed by similar pixels (for example, foreground pixels) adjacent to each other. In the case of a binarized image, the similar pixel can indicate a foreground pixel (for example, a black pixel) and can also indicate a background pixel (for example, a white pixel). In the case of a gradation level image, a similar pixel can indicate a pixel whose gradation level falls within a certain range. The normal connection area includes a connection area of 4 adjacent areas and a connection area of 8 adjacent areas. Since the connection region is a concept well known to those skilled in the art, the description thereof is omitted here. Moreover, a connection area | region can be detected by various well-known methods, for example, all the connection area | regions in a character image can be detected by the analysis of the connection area | region of 8 adjacent areas.

  In the analysis of the connected region of the eight adjacent regions, first, one foreground point is detected and used as a starting point, and foreground points that are not accessed are detected in the eight adjacent regions of the starting point, and these are set as new starting points. To do. This detection process is repeated recursively until the next new starting point cannot be detected, and the detection is terminated and all accessed foreground points are output as one connected region. A new non-accessed foreground point can be detected, and other connected regions can be detected from that point. This process is performed until all points are accessed. For example, Non-Patent Document 1 can be referred to for the analysis of the connection region of the eight adjacent regions.

  Then, features can be extracted based on the result of analysis of connected regions (characters can be divided), and OCR recognition can be performed by the classification device.

  However, a desired result cannot be obtained only by OCR recognition for a character string alone. The inventors of the disclosed technology have discovered that they often contain a certain combination of rules in a string. By obtaining statistics for a large number of character strings and obtaining this rule, and using this rule for character string recognition, the accuracy of the recognition rate can be greatly improved.

  In one embodiment of the disclosed technology, a method for specifying a character string to be recognized by integrating character combination statistical information and OCR recognition is proposed.

  In many cases, the character string to be recognized is a character string that is generally or routinely used. For example, for the Japanese character string “hanashiro”, if statistics are performed for a large amount of Japanese characters, the probability that the character combination “na” will appear after the character combination “ha” at the beginning appears. Can be obtained. Then, when recognition is performed, a better result can be obtained by integrating the result of OCR recognition and this statistical information.

  FIG. 1 is a flowchart schematically showing a method according to the first embodiment of the disclosed technique.

  In step S110, the character string recognition device divides the character string image into a plurality of segments. This division step can be performed by various existing techniques. Taking “hanashiro” as an example, division and recognition can be performed based on the analysis result of the connected region. First, the foreground connected area can be divided into several adjacent segments, and one to three adjacent segments can form one complete character. The character string recognizing device usually detects all possible dividing points by a double elastic matching algorithm and then detects an optimum dividing point by dynamic programming. Regarding the division of the connection region, specifically, Patent Document 4 and Non-Patent Document 2 can be referred to.

  In step S120, the character string recognition apparatus performs OCR recognition on the plurality of divided segments to obtain candidate characters. It is obvious to those skilled in the art that when a character image is recognized, the character image to be recognized has a probability of being acquired as a certain character. For example, the character image “h” may be recognized as “h” or “b” in OCR recognition, but the probability of being recognized as “h” or “b” is different. This probability is called recognition reliability. Thus, each segment corresponds to at least one candidate character.

  In step S <b> 130, the character string recognition apparatus acquires statistical information of segment candidate characters and / or statistical information of character combinations formed by the segment candidate characters. For example, for the character combination “ha”, the character string recognition device performs statistics on a large amount of vocabulary in advance, and acquires the probability of including the character combination “ha”. Similarly, the character string recognition device can acquire the probability that the character combination “han” appears in the vocabulary. Here, it is natural that the range of vocabulary for performing statistics can be limited. For example, the character string recognition device can obtain the corresponding probabilities by statistically analyzing all company names, statistically analyzing all first and last names, and the like. When the character string recognition device recognizes that a character string belongs to a certain field or a certain range when recognizing a character string, the character string recognition device can obtain a better result by using the corresponding probability. Here, the statistical process is completed in advance, and the statistical result can be directly used as an input in step S130.

  The character string recognition device can obtain the following results based on the statistical information of each character combination described above. When a certain character or character combination appears, the character string recognition device can obtain a result of how many probabilities that a certain character or character combination appears after that. For example, when the probability that the character “ha” appears and the probability that “han” appears is also known, the character string recognition device P (B | A) = P (AB) / P (A ), When “ha” appears, the probability of “n” appearing after that can be obtained. Conversely, the character string recognizing device can also obtain how many probabilities that “h” appears before “an”.

  In step S140, the character string recognizing device integrates the statistical information and the reliability of OCR recognition of the candidate character to determine the candidate character string as a character string. For example, the character string recognition apparatus selects candidate characters having a high appearance probability in the current context based on statistical information for a plurality of OCR candidate characters in the same segment. Further, the character string recognition device selects candidate characters by adding a certain weight to the statistical probability and the reliability of OCR recognition. Further, the character string recognition device selects from among candidate characters whose OCR reliability exceeds a certain threshold based on the statistical information.

  According to the method according to the present embodiment, the character string recognition apparatus uses the statistical information of the candidate character combinations of a plurality of segments in addition to the reliability of OCR recognition, so that the recognition rate accuracy can be further improved.

  Here, it is important to explain that statistical information includes one character, character type and / or character combination, one character, character type and / or character combination, and one other character, character type. And / or the probability that a character combination will appear together. For example, for the character string “hanashiro123”, the character string recognition device has the probability of “h” and “a” appearing together, the probability of “ha” and “n” appearing together, “h” and “an” ”, The probability of“ ha ”and“ na ”appearing together, the probability of“ ro ”and“ 12 ”appearing together, the probability of the letter and number appearing together, It is possible to statistic the probability that three) numbers appear consecutively, the probability that multiple consecutive letters and multiple consecutive numbers appear. That is, all character recognition devices perform statistics based on the character / character type / character combination that appears before (or after), and the character / character type / character combination that appears after (or before) appears. The probability of doing is obtained and used for recognition.

  When the statistical information relates to the character combination, the character string recognition device can integrate the OCR reliability and determine candidate character combinations of a plurality of segments related to the character combination, and the method is the same as the above method. is there. Further, the character string recognition apparatus may convert the probability related to the character combination into the probability related to one character and use it.

  Based on the first example, the inventor has found that syllables have a statistical association in all languages. Taking Japanese user names as an example, there are usually several syllables, and these syllables have a statistically constant relationship, so a better recognition effect is realized by using such a relationship. be able to. For example, in the case of the Japanese user name “hanashiro”, four syllables “ha”, “na”, “shi”, and “ro” are included. Can be configured. For example, a character combination can include one syllable such as “ha” and “na”, two syllables such as “hana” and “nashi”, and three syllables such as “hanashi”. It can also contain two syllables. Therefore, in order to recognize a character string, the character string recognition device detects, for example, an n-dimensional set composed of adjacent syllables from an n-dimensional thesaurus. When the character string recognition device detects an n-dimensional set of adjacent syllables, this syllable obtains a higher score. On the other hand, if the character string recognition device does not detect an n-dimensional set of adjacent syllables, this syllable is just an OCR recognition probability. In the following second embodiment, “hanashiro” will be taken as an example, and an n-dimensional group method based on this syllable will be specifically described (three-dimensional group method is taken as an example). FIG. 2 is a flowchart schematically showing a method according to the second embodiment of the disclosed technique.

  In step S210, as described in the first embodiment, the character string recognition device first divides the image of the character string “hanashiro” into a plurality of segments. Here, it is assumed that the image is divided into “h”, “a”, “n”, “a”. Thereafter, in step S220, the character string recognition device performs OCR recognition on each segment, and determines the probability of OCR recognition of each character in “hanashiro”. FIG. 3 shows the result of the recognition. In FIG. 3, the probability of OCR recognition of the character corresponding to each line is shown. For example, the character string recognition apparatus recognizes the first image segment as “k” with an OCR recognition probability of 0.114, “n” as 0.101, and “k”. The probability is 0.101, and the probability of recognizing “m” is 0.074 or the like. For the second image segment, the character string recognition device has a probability of recognizing “a” of 0.132 and a probability of recognizing “u” of 0.082.

In step S230, the character string recognition device sorts the result of OCR recognition for each character. The character string recognition device performs selection according to the principle of removing recognition results with low recognition reliability and leaving recognition results with high recognition reliability. For example, the character string recognizing device performs selection using the expression Cer _i / Max (Cer _i )> T. Here, i indicates the number of the recognition result, and in the recognition result shown in FIG. 3 taken as an example, ten candidate characters are listed for each character. 10 is assumed. Ceri is the OCR probability of the i-th candidate character. T is a threshold and can be set to 0.75, for example. According to this equation, the character string recognition device determines that a character to be recognized cannot be a candidate character if the reliability of a candidate character is much smaller than the maximum reliability of all candidate characters. To do. And since a character string recognition apparatus does not calculate this candidate character in the calculation which continues, the amount of calculation can be reduced significantly. In such a form, the sorting result shown on the right side of FIG. 4 is obtained. As shown in FIG. 4, three candidate characters “h”, “n”, and “k” are left as the first character, and only one candidate character “a” is left as the second character. As a result, three candidate characters “n”, “h”, and “m” are left as the third character. Note that, for example, the selection is performed by directly determining a threshold value that is not related to the maximum reliability, or the selection is performed by another selection mode that directly uses a predetermined amount of candidate characters having the maximum reliability. You can also.

  In step S240, the character string recognition device combines candidate characters with syllables composed of two characters and / or syllables composed of three characters based on the spelling law of Japanese. Therefore, the character string recognition apparatus performs analysis on a large number of Japanese user names in advance, and acquires, for example, the result shown in FIG. A general two-character syllable is shown on the left side of FIG. 5, and a general three-character syllable is shown on the right side of FIG. Thus, since the first candidate character is “h”, “n”, and “k” and the second candidate character is “a”, the character string recognition device has the first candidate syllable It can be determined that “ha”, “na”, and “ka”, and can be similarly estimated thereafter. The character string recognition device still retains candidate characters not included in the syllable as independent characters. For example, the character string recognizing apparatus has candidate characters for the fifth to seventh characters “s”, “h” / “k”, “i”, and combinations of these are “shi”, “s” "-" Hi "or" s "-" ki ".

  In step S250, the character string recognition device lists all possible paths (that is, possible combinations of candidate characters) based on the processing result based on the above syllables. For each path, there are several candidate recognition results for “syllables” or “characters” corresponding to each node of the path, so the character string recognition device can combine several candidate character strings of the same length. it can. The left side of FIG. 4 shows a diagram including all paths for recognizing “hanashiro”.

  Subsequently, in step S260, the character string recognition apparatus calculates a score for each node shown in FIG. 4 by integrating the statistical information and the reliability of OCR recognition. Since there is no other node before the first node, the character string recognition device directly uses the probability of appearing independently and the probability of OCR recognition for the first node. Calculate the score directly. Since the character string recognition device has a statistical relationship with the previous node from the second node, the score of the previous node and the current node when the previous node appears The recognition probability of the current node is calculated based on the probability that the node appears and the probability of OCR recognition of the current node. The character string recognition device calculates the probability for each node using this technical idea.

For example, the character string recognition device calculates a score for each node of each path according to the following equation (1). Here, the expression (1) is “Score (S _p ) = Score (S _p−1 ) + logP _nlo (S _p | h _p−1 ) + logP _ocr (S _p )”.

In Equation (1), S indicates a syllable and is composed of characters. p indicates a syllable number. For example, the first syllable ha is p = 1. In equation (1), the first item Score (S _p−1 ) indicates the score of the node before the current node in the path, and the first node in the path is The term is zero. P _nlp (S _p | h _p-1 ) in the second item is the probability that the current syllable obtained through statistics in the natural language when history information h _p-1 is present (ie, before that) The probability of the current node appearing when the node appears). Since the first node in the path has no history information if there is no previous node, this term directly indicates the probability of appearance of the first node in the natural language. P _ocr (S _p ) indicates the probability of OCR recognition of a syllable, and is obtained, for example, by multiplying the probability of OCR recognition of each character constituting the syllable. The score of the last node of each path calculated by this method is the score of this path, and the path having the maximum score can be selected as the character string recognition result of the character string having the highest reliability. Note that the logarithmic value of probability is used in Equation (1), and addition is used when combining the score of the previous node and the probability of natural language with the probability of OCR recognition. For those skilled in the art, it is obvious that the probability and the score of the previous node may be used in any mathematical form, considering the above factors together in the score of each node and each path.

The probability that each syllable is _unique and some syllables, that is, P _nlp in the context, can be obtained by various methods, and the method according to the present embodiment is directly used as an external input. Can do. For example, the string recognizer uses the SRILM toolkit to perform simulations and statistics on a large number of related strings to obtain a single probability of each syllable and the probability that some syllables are together can do. For example, FIG. 6 shows simulation and statistical results for syllables. FIG. 6 shows statistics using user names of hundreds of thousands of valid Japanese Email addresses. As shown in FIG. 6, the statistical result includes the probabilities of a single syllable (1-grams), a two-syllable group (2-grams), and a three-syllable group (3-grams). The logarithm calculation is performed with a base of 10, and the calculation result is hereinafter referred to as a probability directly). For ease of explanation, only specific statistical results for the probability of three syllables are shown. Thus, the probability that “ka-ha-ta” appears in the user name is −1.8873356, and the probability that “ke-ha-ra” appears is −0.001828611. The SRILM toolkit utilized as an example is a language statistics tool well known to those skilled in the art. Specifically, the SRILM toolkit can be referred to “http://www.speech.sri.com/projects/srilm/” and will not be described in detail here. In the present embodiment, the character string recognition apparatus performs calculation by directly using the corresponding probability value in the statistical result.

The character string recognition device calculates the score Score (S1) of the first node for the first node “ha” in the first path “ha-na-shi-ro” shown in FIG. Here, the character string recognition apparatus can use the natural language probability of the syllable “ha” and the probability information of OCR recognition in various forms. For example, character string recognition devices can be combined by a weighting method. Here, the above-described equation (1) is calculated as an example. Since p = 1 and there is no previous node, the character string recognition device adds the probability of the natural language of the syllable “ha” and the probability of OCR recognition according to the equation (1) to the first node. Score (S1) is obtained. For the second node in this path, ie the syllable “na”, the score is the score of the previous node, the probability that the current node will appear if the previous node appears, the probability of OCR recognition of the current node is connected with. Therefore, the character string recognition apparatus can specify the score of this node similarly using various combinations or calculation rules. Here, equation (1) is still taken as an example. As can be seen from equation (1), the syllable score is not only associated with the current node itself, but also with the previous syllable, ie “ha”. The character string recognition device can easily calculate the conditional probability logP _nlp (S _p | h _p−1 ) therein based on the statistical result of the SRILM toolkit. This technical idea is that if the previous node is determined to be “ha”, for example, if the previous node is “ha” by using the statistical results for a large number of examples to determine the underlying laws of the language, the current node is “na”. Indicates how many a certain probability is. Here, hp _-1 does not indicate that only history information of the previous syllable is used, but indicates that history information of more previous syllables can be used. In the current method of the disclosed technology, since the history information of the two previous syllables is used from the third node, it is called a three-dimensional composition method based on syllables. Naturally, the character string recognition apparatus can use the statistics of more syllables together for calculation. Score (S _p-1 ) indicates the score of the node before the current node in the related path, and in the disclosed technique, only the score of one node before that can be used.

  According to Expression (1), the score of each syllable in the first pass can be calculated and obtained, so the character string recognition device can also obtain the score of the first pass. Similarly, the character string recognition device obtains scores of the second pass, for example, “ha-ha-shi-ro”, and the third pass, for example, “na-na-shi-ro”. The maximum path can be determined. In step S270, the character string recognizing device can determine a candidate character string having the highest recognized probability.

  In this embodiment, the character string recognition device divides the character strings to be recognized based on the syllables, and uses the correlation probability of these syllables in the natural language based on statistics, so the accuracy of the recognition rate is improved. It can be greatly improved.

  In practice, there are often situations where letters and numbers are mixed, such as “hanashiro123”. In such a case, an effective recognition method is not disclosed in the existing technology. The inventor has discovered that this situation often appears in self-defined names by users, such as the user name “Email”. By statistically analyzing a large number of such character strings, it is possible to obtain a certain law in which combinations of letters and numbers exist. For example, if the preceding a characters are letters, then the probability that the following b letters are letters or numbers can be statistics, or if the preceding c letters are numbers, The probability that d characters are letters or numbers (a, b, c, d are all natural numbers) can be statistics. Based on the idea of the first embodiment of the disclosed technology, a method for solving this problem by an n-dimensional character combination method is proposed. FIG. 7 is a flowchart schematically showing a method according to the third embodiment of the disclosed technology (a three-dimensional composition method is taken as an example).

  First, in step S710, the character string recognition device divides the input character string image into a plurality of segments as in the above-described embodiment.

  In step S720, the character string recognition device first performs OCR recognition on each segment in a character string in which a character and a number are mixed to determine the probability of OCR recognition. In the same manner as shown in FIG. 3, the character string recognition apparatus can obtain a series of probabilities for each character.

In step S730, the character string recognizing device sorts the result of OCR recognition of each character. The principle of selection is to remove recognition results with low recognition reliability and leave recognition results with high recognition reliability. For example, the character string recognition device can perform selection using the formula Cer _i / Max (Cer _i )> T. I in this equation indicates the number of the recognition result, and in the recognition result shown in FIG. 3 taken as an example, ten candidate characters are listed for each character. 10 is assumed. Ceri is the OCR probability of the i-th candidate character, and T can be set to a threshold value of 0.75, for example. According to this equation, if the reliability of a certain candidate character is much smaller than the maximum reliability of all candidate characters, the character string recognition device determines that the character to be recognized cannot be a candidate character. To do. And since a character string recognition apparatus does not calculate this candidate character in the calculation which continues, the amount of calculation can be reduced significantly. With such a form, the character string recognition device can obtain a similar one to the sorting result shown on the right side of FIG. Note that the character string recognition device performs other selection, for example, by directly determining and selecting a threshold value not related to the maximum reliability, or directly using a predetermined amount of candidate characters having the maximum reliability. Sorting can also be performed according to the form.

  In step S740, the character string recognition device lists all possible paths based on the selection result. For each path, the character string recognition device has a few candidate recognition results for “characters” corresponding to each node of the path, so that recognition words having the same length can be combined. Therefore, the character string recognition apparatus obtains a path diagram similar to the path diagram shown in FIG.

In step 750, for each node in the path diagram, the character string recognition device integrates the statistical information and the reliability of OCR recognition to calculate a score of the node. Here, the following equation (2) is used. Here, the expression (2) is “Score (C _p ) = Score (C _p−1 ) + logP _nlo (C _p | h _p−1 ) + logP _ocr (C _p )”.

The difference between the expression (2) and the expression (1) is that the expression (1) is calculated based on the probability of the syllable S, whereas the expression (2) is based on the probability of the character or the character combination C. To calculate. As can be seen from the equation, in the method for recognizing character strings mixed with letters and numbers proposed by the disclosed technology, the probability P _nlp by statistics in natural language, the probability P _{OCR of} OCR recognition, and the history information h _{p− 1} is used.

  When performing the calculation for each node, the character string recognition apparatus performs simulation and statistics on a large number of similar character strings using the SRILM toolkit, and thus the probability that each character string appears alone, Probability of various combinations of letters and numbers can be obtained. Similarly, the character string recognition apparatus can obtain this probability by various existing methods, and the disclosed technique may directly use the statistical result. However, in order to facilitate understanding, a simulation and statistical process will be described below as an example.

  First, a large number of email address user names are statistically analyzed as a sample library. For example, the inventors of the disclosed technology performed statistics using 696818 valid Email address usernames. Using “mp2003 @ abc ... xyz.ac.jp” as an example, the user name “mp2003” was extracted.

  Thereafter, all user names were replaced as follows.

  Replace the letter in the username with "a"

  The number in the user name was replaced with “0”.

  For example, “mp2003” is replaced with “aa0000”.

  Subsequently, statistical analysis was performed on all samples using the SRILM toolkit.

  The result of statistical analysis includes statistical results such as a one-dimensional group, a two-dimensional group, and a three-dimensional group. FIG. 8 shows only the three-dimensional group statistical results. In FIG. 8, “<s>” indicates start and “</ s>” indicates end. This statistical result shows that the probability of three numbers appearing together (“000”) is −0.4126034, and the probability of ending after two numbers appearing together (“00 </ s>”) is −0. .4345168 or the like.

According to this statistical result, the character string recognition device can obtain the following conditional probabilities, for example. That is, if the first character is a letter / number, there is a probability that the second character is a letter / number, the first character is a letter / number, and the second character is a letter / number. If so, there is a probability that the third character is a letter / number. That is, the term logP _nlp (C _p | h _p−1 ) in the equation (2) can be obtained. As described in the equation (1), _hp-1 does not use only the history information of one character before that, but can use history information of more characters before that. Show what you can do. In the present method according to the present application, since the history information of the two characters before the third character is used, it is called a three-dimensional composition method based on characters. A person skilled in the art can easily conceive that the history information of more syllables before that can be used, which is based on the idea proposed in the disclosed technology.

  Since the other calculation by Formula (2) is the same as the calculation by Formula (1), description is abbreviate | omitted here.

  Finally, in step S760, the character string recognizing device obtains each character's score and the path having the maximum score, and determines the candidate character string having the highest recognized probability.

  In this example, a three-dimensional set based on characters was used. Here, the character string recognizing device uses the probability that a combination of a letter and a number appears based on statistics, that is, uses information on a large number of user naming habits. Therefore, the character string recognition apparatus has greatly improved the accuracy of the recognition rate of character strings in which numbers and letters are mixed. Here, in the present example, only various pieces of probability information of combinations of normal numbers and letters are statistically calculated. Those skilled in the art will readily conceive that the probability information of various combinations of different numbers and different letters may be specifically statistics. For example, it is possible to statistic the combined information of “ab1” and “12b”, not just the usual information such as two letters and one number or two numbers and one letter. This does not depart from the idea proposed for the disclosed technology.

  The inventor has found that the delimiter “.” Appears more frequently in the character string to be recognized. With the spread of the Internet, both homepage addresses and Email addresses include such a delimiter “.”. The delimiter “.” Serves to divide the domain name at each level in the domain name portion, and frequently appears in Email. Therefore, recognition of such delimiters is becoming increasingly important. According to a fourth embodiment of the disclosed technique, a method for recognizing a delimiter is proposed. FIG. 9 is a flowchart schematically showing a method according to the fourth embodiment of the disclosed technique.

  In step S <b> 910, the character string recognition device performs a connected region analysis on the input character string image. The character string recognition device can obtain each parameter of the connected region for each connected region CC based on the analysis of the connected region of the character string image. For example, the character string recognition device determines the coordinates of the position of the connected area, the number of pixels in the connected area, and the like.

  In step S920, the character string recognition apparatus determines a threshold value in order to recognize the delimiter. For example, the character string recognition device selects a plurality of connected regions in order from the connected region having the smallest number of pixels in the connected region based on the number of pixels in the connected region, and calculates an average value of the number of pixels in the selected connected region. Then, a threshold is set based on the calculated average value. For example, the character string recognizing device performs calculation using the three previous connected regions, and obtains an average value Av3. Subsequently, the character string recognition device sets T1 = aAv3 as a threshold value. a is a parameter that can be selected and adjusted according to the situation, and its purpose is to obtain the best recognition effect. For example, the character string recognition device can set a = 3. It is obvious that the determination of the threshold value is not limited to the above form. For example, the threshold value can be directly determined based on a result of simulation performed on a large number of samples, or can be determined based on a result obtained by performing simulation using a sample of the current user.

  Thereafter, in step S930, the character string recognition device determines whether or not the number of pixels in each connected region is less than T1. When the number of pixels is less than T1, the character string recognition device determines that the connected region is a candidate delimiter “.” In step S950. Since the writing habits for each user are different, there is a possibility that a large difference exists in the writing of the user whose delimiter “.” Size is different. The method of calculating the average value of the number of pixels according to the minimum number of connected regions is based on the “self-adaptation” form, taking into account the difference in writing habits of different users (the size of points written by different users is different) Recognized delimiter “.”.

  Since the point “.” Also exists in the characters “i” and “j”, in step S940, the character string recognition apparatus further determines whether or not the recognized point “.” Is positioned below the character line. Determine. When the recognized point “.” Is located at the lower part of the character line, the character string recognizing device determines that this point is a delimiter. The character string recognition apparatus can perform this determination based on the coordinate parameters of each connected area.

  When the inventor of the disclosed technology knows which database a character string to be recognized belongs to, it uses the similarity between the character string candidate character string and the character string in the predefined database. And found that the accuracy of the recognition rate can be further improved. For example, based on the information acquired from the outside (for example, information provided by the user, applying the environment of this embodiment, etc.) or the structural characteristics of the character string itself, which database (for example, the character string to be recognized is , Personal names, company names, names of universities and research institutions, etc.). For example, when the character string to be recognized is “fujitsu”, the character string is determined to be the name of a company in Japan by other means, for example, information proposed by the user. The similarity between the character string and “fujitsu” in the database must be the highest by calculating the similarity between each character string in the database and the candidate character string using a database of predefined Japanese company names. And recognize this as “fujitsu”. This embodiment will be described in detail below.

  There are several candidate character strings for the character string to be recognized, and each character in each character string has a certain reliability. Therefore, in the method according to the present embodiment, the character string recognition apparatus uses the reliability of OCR recognition in the candidate character string when calculating the similarity between the candidate character string and the character string in the database.

  FIG. 10 is a flowchart schematically showing a method of correcting the OCR recognition result using the already grasped database according to the fifth embodiment of the disclosed technology.

  In step S1010, the character string recognition device detects a character string similar to the result of OCR recognition in a predefined database based on the result of OCR recognition of the character string. For comparison and detection of character strings, various known existing techniques can be used. For example, TDAG (three-dimensional directed acyclic diagram) can be used. TDAG is a method for detecting a character string similar to a character string from a database, and since it is a technique known to those skilled in the art, description thereof is omitted here.

  In step S1020, the character string recognizing apparatus calculates the similarity between some results obtained through OCR recognition and similar special domain names in the database detected in step S1010. For example, the character string recognizing device calculates the similarity using an LD (Levenshtein Distance) algorithm. The character string recognition device does not always compare the recognition results of all candidates from the viewpoint of efficiency. For example, the character string recognition apparatus can perform calculation using two candidate character strings having the maximum probability of OCR recognition. The character string recognition apparatus can form another character string by inserting a character into one character, deleting a character, or replacing a character. If character string A is converted to character string B, the least number of executions in the three types of operations is called the AB LD distance, and the AB distance is obtained by dividing the LD distance by the length of the character string. Can do. In the existing technology, the standard formula for calculating the LD distance is as follows.

  In equation (3), LD (i, j) is the i-th character C (i) (or i-th number of characters) in the first character string (which can correspond to the candidate character string recognized by OCR). A part of the character string up to the character) and the corresponding jth character C (j) in the second character string (which can correspond to the character string to be compared in the database) (or jth character string) (Part of the character string up to the character)), i and j are natural numbers. For the last character of the first character string and the second character string, LD (i, j) indicates the distance between the two character strings. The first expression in Expression (3) is obtained by comparing the first character string (up to the i-th character) and the second character string when compared with the second character string after deleting the i-th character in the first character string. The distance from the character string indicates that 1 is added to LD (i-1, j). The second expression in the expression (3) is contrary to the first expression, that is, when the jth character in the second character string is deleted and then compared with the first character string, the second character string ( The distance between the first character string and the first character string indicates that 1 is added to LD (i, j-1). Conversely, the second expression corresponds to inserting one character into the first character string. The third equation in the equation (3) is obtained by comparing the i-th character in the first character string or the j-th character in the second character string with another character and then comparing them with the distance LD ( This indicates that COST is added to i-1, j-1).

  Here, there are many types of second character strings obtained by correcting the first character string. For each type of form, any LD (i, j) value can be obtained, and the smallest value among them is taken as the final LD value.

  In Expression (4), C (j) represents the jth character of the second character string, and CG (i) represents all of the image segments corresponding to the ith character in the first character string. A set of candidate characters for OCR recognition. Formula (4) is the jth character in the character string to be compared in the database (that is, the second character string) (that is, the target for replacing the i-th character in the first character string): If it exists in the set of candidate characters of the image segment corresponding to the i-th character of the first character string, it means that the replacement price is 0. Equation (4) means that the price is 1 if it does not exist in the set of image segment candidate characters corresponding to the i-th character in the first character string.

  However, such an existing technique considers only whether or not it exists in a set of candidate characters for character OCR recognition, and does not consider the reliability of character OCR recognition. The distance to a plurality of character strings in the database is the same (or, conversely, the plurality of candidate character strings have the same distance to a certain character string in the database), which is disadvantageous for optimization of recognition using the distance.

  In view of this problem, according to the present embodiment of the disclosed technology, the COST function is improved. Specifically, the value corresponding to the recognition reliability that the segment corresponding to the candidate character is recognized by the character at the corresponding position in the character string to be compared in the database from the distance of the character string corresponding to a certain candidate character Decrease. For example, the COST function is corrected to the following equation (5).

  The above formula for the COST function is as follows: 1) The i-th character C (i) of the first character string recognized by OCR (that is, a certain candidate character CG (i, k) in CG (i), k is | CG ( i) | the following natural number) is the same as the jth character C (j) of the second character string in the database. That is, if C (i) = CG (i, k) = C (j), it is determined that the first character string is most likely the second character string in the database. Then, the OCR recognition reliability p (CG (i, k)) of one candidate character is subtracted. That is, it means p (C (j)) or p (C (i)). 2) The i-th character C (i) of the first character string recognized by OCR (that is, a candidate character code CG (i, k) in CG (i)) is the j-th character of the second character string in the database. Different from the characters. That is, C (i) = CG (i, k)! = C (j), but the jth character of the second character string is also present in CG (i). That is, if C (j) εCG (i), from the corresponding distance, the corresponding segment of the candidate character is the character at the corresponding position in the second character string to be compared in the database (ie, C (j )) Means that the value P (C (j)) corresponding to the recognition reliability recognized is decreased. In the case of (1), C (j) = CG (i, k) is one of C (j) εCG (i), so both (1) and (2) This is the case when C (j) εCG (i). That is, the value corresponding to the reliability of the segment corresponding to the candidate character recognized by the character at the corresponding position in the character string to be compared in the database is reduced from the distance of the character string corresponding to a certain candidate character.

  In the following expression of the COST function in Expression (5), the j-th character C (j) of the second character string does not exist in CG (i) (at this time, the i-th character of the first character string recognized by OCR This means that the number of characters C (i) (that is, a certain candidate character CG (i, k) in CG (i) is different from C (j)) is calculated according to an existing algorithm.

  When Equation (3) and Equation (5) are integrated, the corresponding OCR recognition probability is taken into account when calculating the distance between the first character string recognized by OCR and the second character string in the database. The higher the probability of OCR recognition that the i-th character in the first character string recognized by the OCR recognition is recognized by the character at the corresponding position in the second character string to be compared in the database, the more the character is in the database. Indicates that it may be a corresponding character in the second character string. Therefore, the distance between them is small.

  In other optimal embodiments, equation (5) can be improved, for example, improved to equation (6) below.

  Equation (6) can also produce the effect of Equation (5). In the following formula (6), since it is (A), it can be recognized that the OCR reliability of C (j) is zero. Therefore, the following expression can be unified with the above expression, and the following expression (7) can be obtained. Here, Expression (7) is “COST = 1−p (C (j))”.

  In other optimal embodiments, further improvements can be made to the COST function. Specifically, one candidate character of any one segment in the candidate character string (that is, the i-th character C (i) = CG (i, k) in the first character string) is compared in the database. When the character string corresponding to the candidate character is different from the character at the corresponding position (that is, the jth character C (j) of the second character string), the distance of the character string corresponding to the candidate character is The value P (C (i)) corresponding to the degree can be increased. From the viewpoint of OCR recognition, this technical idea is less likely to be replaced with the corresponding character in the character string in the database as the reliability of recognition of a candidate character is higher. The distance between the columns needs to be increased. When C (i) = C (j), since there is no character different from C (j) in the first character string, it can be recognized that “different character” is “empty”, and its OCR It can be recognized that the recognition probability is zero. Therefore, if the further corrected COST function is Expression (8), it can be rewritten into Expression (9).

  Here, P (C (i)) indicates the reliability of OCR recognition of C (i) when C (i) in the first character string is different from C (j) in the second character string. P (C (i) | (C (i) ≠ C (j)) adds a P (C (i)) value only when C (i) ≠ C (j), and C (i) When = C (j), C (i) different from C (j) does not exist, which means that 0 is taken.

  Equation (3) can be further simplified. When inserting into the first character string and when deleting from the first character string, the first character string is missing a character (that is, there is an “empty” character at the corresponding position). It is understood that the character string of 2 is missing characters (that is, there is an “empty” character at the corresponding position). With respect to empty characters, in equation (9), both p (C (i)) and p (C (j)) are 0. Then, the calculation of the LD distance at the time of insertion, deletion, and replacement operation in Expression (3) can be unified with the following Expression (10) of the COST expression based on Expression (9). Here, Expression (10) is “LD (i, j) = LD (i−1, j−1) + COST”.

  In step S1030, the character string recognition apparatus determines whether the distance between the two character strings is less than a predetermined threshold value. If the character string recognition device determines that the distance is less than the predetermined threshold, in step S1040, the character string recognition device replaces the OCR recognized character with the character string in the corresponding database. If the character string recognition device determines that the distance is not less than a predetermined threshold value, in step S1050, the character string recognition device outputs the result using the result of OCR recognition.

  Since the character string recognition apparatus performs correction using a database including a character string to be recognized, the accuracy of the recognition rate of such a character string can be further improved.

  The first to fifth embodiments can be arbitrarily combined. One optimal form can combine the methods according to each embodiment. For example, a character string recognition apparatus efficiently and accurately recognizes a character string divided into a plurality of segments by a delimiter and having at least a part of fields having a fixed mode, for example, an Email address, a network address, and the like. In particular, efficient and accurate recognition can be performed on a handwritten character string.

  Hereinafter, the recognition of the Email address “hanashiro @ abc ... xyz.or.jp” will be described as an example.

  A valid email address, for example, “hanashiro @ abc ... xyz.or.jp” is the username “hanashiro”, the middle “@” character, and the domain name “abc ... xyz.or.jp”. It consists of three parts. Among these, domain names belong to a hierarchical structure, and are usually divided into domain names and special domain names. The domain name is generally used and includes, for example, “or” and “jp”. The special domain name indicates an organization or a group having the domain name, for example, “abc ... xyz”. The dot “.” Character is typically used as a delimiter between multiple segments in the username and domain name.

  FIG. 11 is a flowchart schematically showing a method according to the sixth embodiment of the disclosed technique.

  As is known to those skilled in the art, when actually recognizing, a concatenated area is analyzed with respect to a character string of an Email address, but since it has already been described, description thereof is omitted here.

  Based on the result of the analysis of the connected region, in step S1110, the character string recognition device recognizes a delimiter in Email. The character string recognition device regards the “@” character other than the “.” Character as a delimiter because it divides the user name and the domain name. Since the character string recognition apparatus can recognize the character “.” According to the method according to the fourth embodiment of the disclosed technology, the description thereof is omitted here. The character string recognition device recognizes the “@” character according to all current character recognition methods. However, due to the peculiarity of the “@” character, Non-Patent Document 3 proposes a method for recognizing the “@” character, and this method will be briefly described below.

  First, the character string recognition apparatus detects a segment that may be “@”, but since the size of “@” is large, a segment that is not likely to be “@” is removed from the size. Specifically, the character string recognition device determines whether or not the width and height of the segment exceed the scheduled threshold value, respectively, and if the width or height does not exceed the planned threshold value, the segment is “@”. It is determined that there is no possibility. Thereafter, the character string recognition device tests the output values of the “@” modified quadratic discriminant function (MQDF) for the passed segments and converts them into reliability (class condition probability). In this case, the character string recognizing device naturally simulates one modified secondary discriminant function using the “@” simulation sampling library. Since the email address includes one and only one “@” character, the character string recognition apparatus determines the segment corresponding to the maximum value as the “@” character in all the reliability levels.

  Through this processing, the character string recognition apparatus obtains important delimiters “@” and “.” In the Email address. These delimiters divide the Email address into different parts. For example, for “hanashiro @ abc ... xyz.or.jp”, the user name part “hanashiro”, the special domain name part “abc ... xyz”, and the normal domain name part Divided into “or” and “jp”.

  Subsequently, the character string recognition apparatus performs processing on each part with the delimiter as a boundary. The character string recognizing apparatus can perform recognition for each character in each part, can also perform overall recognition, and can use each embodiment individually or in combination. During processing, the character string recognizing device uses the a priori knowledge of the recognition target to determine the property of each part divided by each delimiter, and in the step corresponding to each embodiment, a dictionary or database And / or corresponding statistical data can be used appropriately. For example, the character string recognizing device, for the mail address, from the back to the front, “.” Indicates a normal domain name (usually 1 to 2 layers), a special domain name (user domain name), and thereafter “ It is divided into the user name after “@”. Therefore, each of the character string recognition devices uses a normal domain name dictionary (for example, a top level domain name dictionary and / or a second level domain dictionary), a special domain name database, a user name database, and the like. The same rule applies to homepage addresses.

  As an example, the character string recognition apparatus can perform processing from the back to the front. Thus, for example, when recognizing a mail address, the character string recognition device recognizes a normal domain name in step S1120. Here, the character string recognition apparatus can also recognize the whole and can recognize each character.

  In step S1120, for example, when the character string recognition apparatus scans the connected area from the back to the front and detects the first delimiter “.”, The entire character string recognition apparatus uses all the connected areas after the delimiter as sub-images. Perform recognition. When the character string recognition device detects the second delimiter “.”, The character string recognition device performs overall recognition using the connection area between the first delimiter and the second delimiter as a sub-image. The character string recognizing device can use, for example, a conventional modified secondary classifier MQDF (which has already been described above) for this recognition.

  A domain name usually has a level structure (a top level domain name and a second level domain), and has certain inherent laws. Therefore, in the recognition of the normal domain name by the disclosed technique, the character string recognition device can further improve the recognition accuracy rate and the recognition processing speed by using such a rule. For example, if the recognition result of the last domain name, that is, the first recognition result is one of “com”, “edu”, “org”, and “net”, the domain name definition rule is used. Since it is always the user domain name before that, the domain name recognition process normally ends. If the result of the first recognition is, for example, a national or regional domain name, such as “jp”, then the normal domain name to be recognized is always “ac”, “ad”, “co”, “ne” ”,“ Or ”, etc., one of the second level domains. This is also determined based on the domain name definition rule. Accordingly, the recognized domain name has a probability that it is one of “ac”, “ad”, “co”, “ne”, “or”, and exceeds a threshold value Tr, for example, Tr = 0.7. Then, recognition of the domain name usually ends. This is because this domain name is always followed by a user domain name.

  Here, a priori knowledge such as the rule that defines the domain name is used, so the overall recognition result is clearly superior to the normal recognition result. Note that the character string recognition device may perform recognition for each character without performing overall recognition. Similarly, the character string recognition apparatus uses the dictionary (or database) selected based on a priori knowledge and / or the corresponding statistical data, alone or in combination with the first to sixth embodiments. Needless to say, it can be used.

  Subsequently, in step S1130, the character string recognition device recognizes the special domain name and the user name. Here, recognition can be performed according to the methods according to the first, second and third embodiments.

  Finally, in step S1140, the character string recognition device can correct the recognition result using an already existing database. For example, in the case of an email address in Japan, if the recognition result of a normal domain name is “.ac.jp”, the special domain name before that is always the domain name of a university or research organization. Therefore, the character string recognition apparatus performs the spelling inspection and correction of the recognition result using the domain name database of the university or research organization. Then, the character string recognition device performs inspection and correction on the recognition result of the special domain name using a database of general user names (including company names). This inspection and correction can be processed in accordance with the method according to the fifth embodiment. For example, the character string recognition apparatus calculates the distance between the character string recognized by OCR and the character string in the database. On the other hand, it goes without saying that in this inspection and correction, the character string recognition device can be applied to any part of a character string, for example, an Email address, as long as there is an available database.

  According to the method related to the disclosed technique, the character string recognition device is divided into a plurality of fields by a delimiter, and at least a part of the fields has a fixed mode, for example, an e-mail address, a homepage address, etc. are recognized efficiently and accurately. be able to. In particular, the character string recognition device can recognize a handwritten character string efficiently and accurately.

  The seventh embodiment of the disclosed technique corresponds to the first embodiment of the disclosed technique.

  FIG. 12 is a diagram illustrating a character string recognition apparatus according to a seventh embodiment of the disclosed technology. As illustrated in FIG. 12, the character string recognition device includes a segment division unit 1202, an OCR recognition unit 1204, a statistical information acquisition unit 1206, and a first character string determination unit 1208.

  The segment dividing unit 1202 is installed so as to divide the character string image into a plurality of segments. This division can be performed based on various existing technologies. Taking the “hanashiro” image as an example, the segment dividing unit 1202 can recognize the divided image based on the analysis result of the connected region. The segment dividing unit 1202 can obtain a plurality of image segments “h”, “a”, “n”, “a”,.

  The OCR recognition unit 1204 is installed to perform OCR recognition on a plurality of segments and acquire candidate characters. It is self-evident that a person skilled in the art obtains the probability that a character image to be recognized is a character when the character is recognized. For example, the letter “h” may be recognized as “h” or “b” in OCR recognition, but their probabilities differ and are also referred to as recognition confidence. Thus, each segment corresponds to at least one candidate character.

  The statistical information acquisition unit 1206 is installed so as to acquire statistical information of segment candidate characters and / or statistical information of character combinations configured by segment candidate characters. For example, with respect to the character combination “ha”, statistics for a large amount of vocabulary are obtained in advance, and the probability of including the character combination is obtained. Similarly, the probability that the character combination “han” appears in the vocabulary can also be obtained. Here, the range of vocabulary for statistics can be limited. For example, the corresponding probabilities can be obtained by limiting all company names or statistics to all first and last names. When recognizing a certain character string, the statistical information acquisition unit 1206, when it is found that the character string to be recognized belongs to a certain field or range (see the description of the first to sixth embodiments), the corresponding probability Can be used to obtain even better results. Here, the statistical process may be prepared in advance, or the statistical information acquisition unit 1206 may directly use the statistical result.

  The first character string determining unit 1208 is installed so as to determine the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate character. For example, the first character string determination unit 1208 selects a candidate character that has a high probability of appearing in the current context based on statistical information for a plurality of OCR candidate characters of the same segment, or the statistical probability and the OCR. A candidate character is selected by giving a certain weight to the reliability of recognition, or selected from candidate characters whose reliability of OCR recognition exceeds a certain threshold based on statistical information.

  Since the character string recognition apparatus according to the present embodiment uses statistical information of candidate character combinations of a plurality of segments in addition to using the reliability of OCR recognition, the accuracy of the recognition rate can be further improved.

  The eighth embodiment of the disclosed technique corresponds to the second embodiment of the disclosed technique.

  FIG. 13 is a diagram showing a character string recognition apparatus according to an eighth embodiment of the disclosed technology. As shown in FIG. 13, the character string recognition apparatus includes a segment division unit 1302, an OCR recognition unit 1304, and a selection unit. 1306, a syllable combination unit 1308, a path generation unit 1310, a score calculation unit 1312, and a first character string determination unit 1314.

  The segment dividing unit 1302 is installed so as to divide the character string image into a plurality of segments. For example, when the character string “hanashiro” is taken as an example as in the second embodiment, the segment dividing unit 1302 divides the character string “hanashiro” image into a plurality of segments. Here, a case where the character string recognition apparatus divides the character string “hanashiro” into image segments “h”, “a”, “n”, “a”,... Will be described as an example.

  The OCR recognition unit 1304 is installed so as to perform OCR recognition for each segment and to determine the probability of OCR recognition for each image segment in the “hanashiro” image. The recognition result corresponding to FIG. 3 is shown. For a specific description, the second embodiment can be referred to, and the description is omitted here.

The sorting unit 1306 is installed so as to sort the result of OCR recognition of each character. The selection is performed according to the principle of removing recognition results with low recognition reliability and leaving recognition results with high recognition reliability. For example, sorting can be performed according to the formula Cer _i / Max (Cer _i )> T, where i indicates the number of the recognition result, and in the recognition result shown in FIG. Since 10 candidate characters are listed, i takes 1 to 10. Ceri indicates the OCR probability of the i-th candidate character, and T is a threshold value, which can be set to 0.75, for example. According to this formula, if the reliability of a certain candidate character is considerably smaller than the maximum reliability of all candidate characters, it is determined that there is no possibility that the character to be recognized is the candidate character, and the candidate in the subsequent calculation Since the characters are not considered, the amount of calculation can be greatly reduced. In such a form, the sorting result shown on the right side of FIG. 4 is obtained. Similarly, the second embodiment can be referred to for a specific description. Similarly, it is natural that sorting can be performed by other sorting modes. For example, it is possible to directly select based on a threshold value not related to the maximum reliability, or to directly select using a predetermined amount of candidate characters having the maximum reliability.

  The syllable combination unit 1308 is installed so as to combine candidate characters with a syllable composed of a plurality of characters (usually composed of two or three characters). Therefore, the syllable combination unit 1308 analyzes a large number of Japanese user names in advance and obtains the result shown in FIG. The left side of FIG. 5 shows a syllable composed of two general letters, and the right side shows a syllable composed of three general letters. Similarly, the second embodiment can be specifically referred to.

  The path generation unit 1310 lists all possible paths (that is, possible combinations of candidate characters) based on the processing result by the syllable, and for each path, a “syllable” or “ Since “character” has several candidate recognition results, several candidate character strings having the same length can be combined. FIG. 4 shows a diagram including all paths for recognizing the “hanashiro” image on the left side.

  The score calculation unit 1312 is installed so as to calculate the score for each node in FIG. 4 by integrating the statistical information and the reliability of OCR recognition. Among them, the statistical information is information on a single probability of each syllable and a probability of several syllables appearing together. Specifically, the description of the second embodiment can be referred to. Since there is no other node before the first node, the score calculation unit 1312 directly uses the probability of statistically appearing alone and the probability of OCR recognition for the first node. calculate. Since the score calculation unit 1312 has a statistical relationship with the previous node from the second node, the score calculation unit 1312 presents the score of the previous node and the current node under the situation where the previous node appears. The recognition probability of the current node is calculated based on the probability of the occurrence of the current node and the probability of OCR recognition of the current node. Based on the technical idea described above, the score calculation unit 1312 can calculate the probability for each node. For a specific score calculation process, the second embodiment can be referred to, and the description is omitted here.

  The first character string determination unit 1314 is installed so as to determine the candidate character string having the highest recognized probability based on the score calculation result.

  According to the character string recognition apparatus according to the present embodiment, the character strings to be recognized are divided based on syllables, and the probability that these syllables are related to each other in natural language is obtained based on statistics. Therefore, the accuracy of recognition can be greatly improved.

  The ninth embodiment of the disclosed technique corresponds to the third embodiment of the disclosed technique.

  FIG. 14 is a diagram illustrating a character string recognition apparatus according to the ninth embodiment of the disclosed technology. The character string recognition apparatus includes a segment division unit 1402, an OCR recognition unit 1404, a selection unit 1406, and a path generation unit 1408. And a score calculation unit 1410 and a first character string determination unit 1412. As described in the third embodiment, this character string recognition apparatus is applied to, for example, “hanashiro123”, particularly when the characters and numbers are mixed.

  The segment dividing unit 1402 is installed so as to divide the input character string image into a plurality of segments.

  The OCR recognition unit 1404 is installed to perform OCR recognition on each segment in a character string in which a character and a number are mixed to determine the probability of OCR recognition. As shown in FIG. 3, a series of probabilities for each character can be similarly obtained here.

The sorting unit 1406 is installed so as to sort the OCR recognition result of each character. Sorting is performed according to the principle of removing recognition results with low recognition reliability and leaving recognition results with high recognition reliability. For example, sorting can be performed according to the formula Cer _i / Max (Cer _i )> T. Here, i indicates the number of the recognition result, and in the recognition result shown in FIG. 3 taken as an example, since 10 candidate characters are listed for each character, i can take 1 to 10 it can. Ceri indicates the OCR probability of the i-th candidate character, and T is a threshold value, which can be set to 0.75, for example. With such a configuration, the sorting unit 1406 can obtain the sorting result shown on the right side of FIG. Here, it goes without saying that sorting can be performed by other sorting modes. For example, a threshold value unrelated to the maximum reliability can be directly determined and selected, or can be selected using a predetermined amount of candidate characters with the maximum reliability.

  The path generation unit 1408 is installed to list all possible paths based on the selection result. For each path, the “character” corresponding to each node of the path has several candidate recognition results, so that the path generation unit 1408 can combine several recognition terms having the same length. Therefore, the path generation unit 1408 can obtain a diagram similar to the path diagram in FIG.

  The score calculation unit 1410 calculates the score of the node by integrating the statistical information and the reliability of OCR recognition for each node in the path diagram. The statistical information here is statistical information on the probability that various combinations of letters and numbers appear. The process of obtaining the statistical information and calculating the score can refer to the third embodiment, and will not be described here.

  The first character string determination unit 1412 is installed so as to determine a candidate character string having the highest recognized probability based on the obtained score of each character and the path having the maximum score.

  According to the character string recognizing apparatus according to the present embodiment, the number and the letter are mixed using the probability that a certain combination of the letter and the number appears based on statistics, that is, using a large number of user naming habits. Significantly improve the accuracy of string recognition.

  The tenth embodiment of the disclosed technique corresponds to the fourth embodiment of the disclosed technique.

  FIG. 15 shows a delimiter recognition unit according to the tenth embodiment of the disclosed technique. The delimiter recognition unit includes a connected region analysis unit 1502 and a delimiter determination unit 1504.

  The connected area analysis unit 1502 is installed so as to analyze the connected area for the input character string image. The connected region analysis unit 1502 can obtain each parameter of the connected region for each connected region CC based on the analysis of the connected region with respect to the character string image. For example, the connected area analysis unit 1502 can determine the coordinates of the position of the connected area, the number of pixels in the connected area, and the like.

  The delimiter determination unit 1504 is installed to determine the delimiter based on the analysis result of the connected area. Among them, the delimiter determination unit 1504 first determines a threshold value in order to determine the delimiter. For example, the delimiter determination unit 1504 selects a plurality of connected regions in order from the connected region having the smallest number of pixels in the connected region based on the number of pixels in the connected region, and calculates an average value of the number of pixels in the selected connected region. Then, a threshold is set based on the calculated average value. For example, here, the average value Av3 is obtained by calculating using the top three connected regions. Subsequently, T1 = aAv3 is set as a threshold value. a is a parameter that can be adjusted according to the situation, and its purpose is to realize a further excellent recognition effect. For example, a = 3 can be selected. Here, it goes without saying that the determination of the threshold is not limited to the above-described form. For example, the threshold value can be determined directly based on the result of simulation using a large number of samples, or the threshold value can be determined based on the result of simulation using a sample of the current user. Etc.

  Subsequently, the delimiter determination unit 1504 determines whether or not the value of the number of pixels in each connected region is less than T1, and if the value of the number of pixels is less than T1, the connected region has a candidate delimiter “.”. It is determined that Since the writing habits for each user are different, there is a great difference between users with different delimiter “.” Sizes. The method of calculating the average value of the number of pixels based on a plurality of minimum connected areas is considered to be “self-adaptation” in consideration of the writing habits of each user being different (and thus the size of the point written by the user is different). In the corresponding form, the delimiter “.” Was recognized.

  The characters “i” and “j” also have “.”. Therefore, the delimiter determination unit 1504 further determines whether or not the recognized “.” Is positioned at the bottom of the character line. If the recognized “.” Is positioned at the bottom of the character line, the recognized “.” Is determined. Confirm that it is a delimiter. Such a determination can be made based on the coordinate parameters of each connected region.

  The eleventh embodiment of the disclosed technology corresponds to the sixth embodiment of the disclosed technology.

  FIG. 16 is a diagram illustrating a character string recognition apparatus according to an eleventh embodiment of the disclosed technology. The character string recognition device can efficiently and accurately recognize a character string that is divided into a plurality of fields by, for example, a delimiter, and at least some of the fields have a fixed mode, such as an Email address, a homepage address, and the like. it can. In particular, the character string recognition device can efficiently and accurately recognize a handwritten character string. The character string recognition device includes a delimiter recognition unit 1602, a normal domain name recognition unit 1604, a special domain name and user name recognition unit 1606, and a second character string determination unit 1608.

  Hereinafter, the recognition of the email address “hanashiro @ abc ... xyz.or.jp” will be described as an example.

  Before actually recognizing, a person skilled in the art uses a known method to analyze a linked region for a character string of an Email address, but since this has been described in the above embodiment, description thereof is omitted here. .

  Based on the analysis result of the connected area, the delimiter recognition unit 1602 is installed so as to recognize the delimiter in the character string. The character string recognition device regards an “@” character as a delimiter in addition to a “.” Delimiter for an Email address. The “.” Character recognition process has been described in the fourth embodiment of the disclosed technology, and the “@” character recognition process has been described in the sixth embodiment of the disclosed technology. To do.

  These delimiters divide the Email address into different parts. For example, for “hanashiro @ abc ... xyz.or.jp”, this is the user name part “hanashiro”, the special domain name part “abc ... xyz”, and the normal domain name part. Divide into “or” and “jp”.

  The normal domain name recognition unit 1604 is installed so as to recognize the normal domain name. The normal domain name recognizing unit 1604 can recognize the entire normal domain name and can also recognize each character. Usually domain names have a hierarchical structure, and there are certain inherent laws in them, so that the recognition rate and speed of recognition are improved by using such rules in recognition of ordinary domain names by the disclosed technology. be able to. The specific content of domain name recognition can be referred to the sixth embodiment of the disclosed technology and will not be described here.

  The special domain name / user name recognition unit 1606 is installed to recognize the special domain name and user name. Here, since the recognition can be performed according to the method according to the first, second and third embodiments, it will not be described.

  The second character string determination unit 1608 is installed to check and correct the recognition result using an existing database. For example, for a Japanese email address, if the recognized normal domain name is “.ac.jp”, it indicates that the special domain name before that is always a university or research institution domain. Therefore, it is possible to perform spelling inspection and correction on the recognition result by using the domain name database of the university or research organization. Then, for a general special domain name such as “fujitsu” or a general user name, the recognition result can be inspected and corrected using a corresponding database. This inspection and correction can be performed according to the method according to the fifth embodiment. For example, it can be executed by calculating the distance between the character string recognized by OCR and the character string in the database, and the description thereof is omitted here.

  According to the character string recognizing apparatus according to the present embodiment, efficient and accurate recognition is performed on a character string divided into a plurality of fields by a delimiter and at least a part of segments having a fixed mode, such as an Email address, a homepage address, and the like. It can be performed. In particular, according to the character string recognition apparatus of the present embodiment, it is possible to efficiently and accurately recognize a handwritten character string.

  On the other hand, the various examples and examples described in this specification are merely examples, and are not intended to limit the disclosed technology. In this specification, “first”, “second”, and the like are for distinguishing features so as to more clearly describe the disclosed technology, and should not be regarded as limiting the disclosed technology. .

  Each unit in the apparatus can be arranged in the form of software, firmware, hardware, or a combination thereof. Arranging in specific means or forms available is known to those skilled in the art and will not be described here. When implemented by software or firmware, a program that configures the software is mounted from a storage medium or a network to a computer having a dedicated hardware structure (for example, the general-purpose computer 1700 shown in FIG. 17), and various programs are mounted to the computer. If so, various functions can be realized.

  In FIG. 17, a central processing unit (CPU) 1701 executes each process according to a program stored in a read-only storage device (ROM) 1702 or a program loaded from a storage unit 1708 to a random access memory (RAM) 1703. . In the RAM 1703, data used by the CPU 1701 for executing each process is stored as necessary. The CPU 1701, the ROM 1702, and the RAM 1703 are connected to each other via a bus 1704. An input / output interface 1705 is also connected to the bus 1704.

  Communication with an input unit 1706 (including a keyboard, a mouse, etc.), an output unit 1707 (including a display such as a cathode ray tube (CRT) and a liquid crystal display (LCD) and a speaker), and a storage unit 1708 (including a hard disk) A unit 1709 (including a network interface card such as a LAN card and a modem) is connected to the input / output interface 1705. A communication unit 1709 performs communication processing via a network such as the Internet. A driver 1710 is also connected to the input / output interface 1705 as needed. A removable medium 1711 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory is mounted on the driver 1710 as necessary, and a computer program read from the medium 1711 is mounted on the storage unit 1708 as necessary. To.

  When executing a series of processing via software, a program constituting the software is loaded from a network such as the Internet or a storage medium such as a removable medium 1711.

  It is obvious to those skilled in the art that the storage medium is not limited to the removable medium 1711 that stores the program shown in FIG. 17 and distributes the program separately from the apparatus to provide a program to the user. Examples of the removable medium 1711 include a magnetic disk (including a floppy (registered trademark) disk), an optical disk (including an optical disk ROM (CD-ROM) and a digital versatile disk (DVD)), and a magneto-optical disk. (Mini Disc (MD) (registered trademark)) and semiconductor memory. Alternatively, the storage medium may be a hard disk or the like included in the ROM 1702 and the storage unit 1708, in which a program is stored and distributed to users together with devices including the hard disk and the like.

  The disclosed technology provides a program product in which a command code that can be read by the apparatus is stored. When the command code is read and executed by the apparatus, the method according to the embodiment of the disclosed technique can be executed.

  Correspondingly, a storage medium in which a program product in which a command code readable by the apparatus is stored is also disclosed by the disclosed technology. The storage medium includes, but is not limited to, a floppy (registered trademark) disk, an optical disk, a magneto-optical disk, a memory, a memory stick, and the like.

  Finally, “including”, “comprising” or other synonymous forms of expression indicate “including non-exclusively”. In addition, a process, method, or apparatus that includes a series of elements includes not only those elements, but also other elements that are not explicitly described, and includes elements that are specific to these processes, methods, or apparatuses. . Unless further limited, an element defined by "including" does not exclude the case where other similar elements exist in the process, method, or apparatus that includes the element.

  The embodiments of the disclosed technology have been described in detail above with reference to the drawings. However, the embodiments described above are for explaining the disclosed technology and do not limit the disclosed technology. Those skilled in the art can make various corrections and modifications to the embodiments within the scope of protection of the disclosed technology. Therefore, the scope of the disclosed technology is limited by the meanings of the claims and the like.

  Moreover, the Japanese user name in the said Example is only an example. The user name is not limited to the example in the embodiment, and may be a user name in any language.

  The following supplementary notes are further disclosed with respect to the embodiments including the above examples.

(Supplementary note 1) The character string recognition device is
Dividing the string image into a plurality of segments;
Performing OCR recognition on the plurality of segments to obtain candidate characters;
Obtaining statistical information of the candidate characters and / or statistical information of character combinations formed by the candidate characters;
And integrating the statistical information and the reliability of OCR recognition of the candidate character to determine a candidate character string.

(Supplementary Note 2) The statistical information includes one character, character type and / or character combination, one character, character type and / or character combination, and one character, character type and / or character combination. The character string recognition method according to appendix 1, including the probability of appearing in

(Additional remark 3) The said character combination is a character string recognition method of Additional remark 1 or 2 which is the combination of the character which comprises a syllable, or the combination of a character name and a number.

(Supplementary Note 4) The character string in the character string image includes a delimiter,
The character string recognition method according to appendix 1 or 2, further comprising the step of recognizing the delimiter.

(Supplementary Note 5) The step of recognizing the delimiter includes
Performing a connected region analysis on the character string image to obtain the number of connected region of foreground pixels;
The character string recognition method according to supplementary note 4, comprising: determining a delimiter based on the number of pixels in the connection area.

(Supplementary Note 6) The step of determining the delimiter based on the number of pixels in the connected region is as follows:
A plurality of connected regions are selected in order from the connected region with the smallest number of pixels, and a threshold value is determined by calculating an average value of the number of pixels in the selected connected region, and the number of pixels in the connected region is less than the threshold value. The character string recognition method according to appendix 5, further comprising a step of determining the connection area as a delimiter when the connection area is positioned below the character string image.

(Supplementary note 7) The character string recognition device is
The character string recognition method according to supplementary note 1, further comprising a step of calculating a distance between the candidate character string and a character string in a predefined database based on the reliability of OCR recognition and determining the candidate character string.

(Supplementary Note 8) Corresponding to the reliability that the segment corresponding to the candidate character is recognized as the character at the corresponding position in the character string to be compared in the database from the distance between the candidate character and the character string to be compared in the database The character string recognition method according to appendix 7, wherein a value to be subtracted is subtracted.

(Supplementary note 9) When any one candidate character of the segment in the candidate character string is different from the character at the corresponding position in the character string to be compared in the database, the candidate character and the character string to be compared in the database The character string recognition method according to appendix 7 or 8, wherein a value corresponding to the recognition reliability of the candidate character is added to the distance to the character string.

(Supplementary Note 10) a segment dividing unit that divides a character string image into a plurality of segments;
An OCR recognition unit that performs OCR recognition on the plurality of segments to obtain candidate characters;
Statistical information acquisition unit for acquiring statistical information of the candidate characters and / or statistical information of character combinations formed by the candidate characters;
A first character string determination unit that determines the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate character;
A character string recognition device.

(Supplementary Note 11) The statistical information includes one character, character type and / or character combination, one character, character type and / or character combination, and one character, character type and / or character combination. The character string recognition device according to appendix 10, including the probability of appearing in

(Additional remark 12) The said character combination is a character string recognition apparatus of Additional remark 10 or 11 which is the combination of the character which comprises a syllable, or the combination of a character name and a number.

(Supplementary note 13) The character string in the character string image includes a delimiter,
The character string recognition device according to appendix 10 or 11, further comprising a delimiter recognition unit that recognizes the delimiter.

(Supplementary Note 14) The delimiter recognition unit
A connected region analysis unit that performs a connected region analysis on the character string image and obtains the number of connected region of foreground pixels;
The character string recognition device according to appendix 13, including a delimiter determination unit that determines a delimiter based on the number of pixels in the connection region.

(Supplementary Note 15) The delimiter determination unit
When determining the delimiter based on the number of pixels in the connected region, a plurality of connected regions are selected in order from the connected region having the smallest number of pixels, and the threshold value is determined by calculating an average value of the number of pixels in the selected connected region. Then, when the number of pixels in the connected area is less than the threshold value and the connected area is located at the lower part of the character string image, the character according to appendix 14, which is arranged to determine the connected area as a delimiter Column recognition device.

(Supplementary Note 16) A second character string determination unit that calculates a distance between the candidate character string and a character string in a predefined database based on the reliability of OCR recognition, and determines the candidate character string Furthermore, the character string recognition apparatus of Additional remark 10 containing.

(Supplementary Note 17) From the distance between the candidate character and the character string to be compared in the database, the second character string determining unit corresponds to the character string to which the segment corresponding to the candidate character is to be compared in the database. Item 17. The character string recognition device according to supplementary note 16, arranged to subtract a value corresponding to the degree of reliability recognized for the character at the position.

(Additional remark 18) When the said 2nd character string determination part differs in the character of the corresponding position in the character string used as the comparison object in the said database in any one candidate character of the said candidate character string, this candidate 18. The character string recognition device according to appendix 16 or 17, arranged to add a value corresponding to the recognition reliability of the candidate character to the distance between the character and the character string to be compared in the database.

(Supplementary note 19)
Split a string image into multiple segments,
Performing OCR recognition on the plurality of segments to obtain candidate characters;
Obtaining the statistical information of the candidate characters and / or the statistical information of the character combination formed by the candidate characters,
A character string recognition program that executes a process of determining the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate character.

(Supplementary note 20)
Split a string image into multiple segments,
Performing OCR recognition on the plurality of segments to obtain candidate characters;
Obtaining the statistical information of the candidate characters and / or the statistical information of the character combination formed by the candidate characters,
A storage medium storing a character string recognition program for executing a process of determining the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate character.

1202 Segment division unit 1204 OCR recognition unit 1206 Statistical information acquisition unit 1208 First character string determination unit 1302 Segment division unit 1304 OCR recognition unit 1306 Selection unit 1308 Syllable combination unit 1310 Path generation unit 1312 Score calculation unit 1314 First character string Determination unit 1402 Segment division unit 1404 OCR recognition unit 1406 Selection unit 1408 Path generation unit 1410 Score calculation unit 1412 First character string determination unit 1502 Connection region analysis unit 1504 Delimiter determination unit 1602 Delimiter recognition unit 1604 Normal domain name recognition unit 1606 Special Domain name and user name recognition unit 1608 Second character string determination unit 1700 General-purpose computer 1701 CPU
1702 ROM
1703 RAM
1704 Bus 1705 Input / output interface 1706 Input unit 1707 Output unit 1708 Storage unit 1709 Communication unit 1710 Driver 1711 Removable medium

Claims (10)

The string recognition device
Dividing the string image into a plurality of segments;
Performing OCR recognition on the plurality of segments to obtain candidate characters;
Obtaining statistical information of the candidate characters and / or statistical information of character combinations formed by the candidate characters;
And integrating the statistical information and the reliability of OCR recognition of the candidate character to determine a candidate character string. A segment dividing unit that divides the character string image into a plurality of segments;
An OCR recognition unit that performs OCR recognition on the plurality of segments to obtain candidate characters;
Statistical information acquisition unit for acquiring statistical information of the candidate characters and / or statistical information of character combinations formed by the candidate characters;
A character string recognition device comprising: a first character string determination unit that determines the candidate character string by integrating the statistical information and the reliability of OCR recognition of the candidate character.   The statistical information includes a probability that one character, character type and / or character combination, one character, character type and / or character combination, and one character, character type and / or character combination appear together. The character string recognition device according to claim 2, including:   The character string recognition device according to claim 2 or 3, wherein the character combination is a combination of characters constituting a syllable or a combination of a letter and a number. The character string in the character string image includes a delimiter,
The character string recognition device according to claim 2, further comprising a delimiter recognition unit that recognizes the delimiter. The delimiter recognition unit
A connected region analysis unit that performs a connected region analysis on the character string image and obtains the number of connected region of foreground pixels;
The character string recognition apparatus according to claim 5, further comprising: a delimiter determining unit that determines a delimiter based on the number of pixels in the connection area. The delimiter determination unit
When determining the delimiter based on the number of pixels in the connected region, a plurality of connected regions are selected in order from the connected region having the smallest number of pixels, and the threshold value is determined by calculating an average value of the number of pixels in the selected connected region. When the number of pixels in the connected area is less than the threshold value and the connected area is located at the lower part of the character string image, the connected area is disposed so as to be determined as a delimiter. Character string recognition device.   And a second character string determination unit configured to calculate a distance between the candidate character string and a character string in a predefined database based on the reliability of OCR recognition, and to determine the candidate character string. 2. The character string recognition device according to 2.   From the distance between the candidate character and the character string to be compared in the database, the second character string determination unit converts the corresponding segment of the candidate character to the character at the corresponding position in the character string to be compared in the database. The character string recognition device according to claim 8, wherein the character string recognition device is arranged to subtract a value corresponding to the recognized reliability.   The second character string determining unit, when any one candidate character of a certain segment in the candidate character string is different from a character at a corresponding position in a character string to be compared in the database, the candidate character and the database The character string recognition device according to claim 8 or 9, wherein the character string recognition device is arranged so as to add a value corresponding to the recognition reliability of the candidate character to a distance from the character string to be compared.

Images