JP5240049B2 - Character identification program, character identification device - Google Patents

Description

  The present invention relates to a character identification program and a character identification device, for example, a character identification program and a character identification device that recognize a similar character with high accuracy.

  In general, optical character recognition (hereinafter referred to as OCR: Optical Character Recognition) is widely known as a method for converting characters included in a document captured by an image scanner into text data.

  In OCR, it is required to recognize character recognition with high accuracy, and in particular, it is required to identify similar characters with high accuracy. Examples of the similar characters include “dog” and “large”. If a character that should be recognized as “dog” is recognized as “large”, it is misread.

  As an example of a method for identifying this similar character, the Mahalanobis distance and the projection distance method and the difference in characteristics indicated by each similar character are used, and when any character is given, it is similar to which similar character. There is a mixed identification method for determining whether or not.

  The above-mentioned “difference in characteristics of similar characters” is obtained based on information extracted from the character category of each similar character. This “character category” is defined using a feature vector whose parameters are, for example, character fluctuation, character line inclination, and character area of a predetermined character.

  In general, since the feature vector is shown on the feature space, the above-mentioned “character category” is shown as a distribution on the feature space. This is called a character category distribution. For convenience, parameters such as character fluctuation, character line inclination, and character area are referred to as “features”.

  Next, the above-described mixed identification method will be specifically described. FIG. 12 is a diagram for explaining a conventional mixed identification method. As described above, the feature space is defined in a high dimension, but the feature space 10 in FIG. 12 is, for convenience of explanation, for example, a feature amount defined by the inclination of a character line and a feature defined by a character area. It shall be expressed in a dimension defined by quantity.

  First, the character category distribution A and the character category distribution B (hereinafter simply referred to as category distributions A and B) in FIG. 12 will be described. This category distribution A indicates a category distribution in which agreed characters having different character line inclinations, character areas, and the like are gathered. The same applies to the category distribution B. Note that “consent characters” are the same characters with different fonts and points, or images that have been scanned by the scanner, but have the same characters due to reading errors, presence of dust, etc. .

  Further, the category A and the category B are similar to each other. For example, when the category A indicates “dog”, the category B indicates “large”. Further, when category A indicates “crow”, category B indicates “bird”.

  When an arbitrary character is input, similar character identification is performed to identify whether the input character belongs to category A or category B.

  In this way, identification of which arbitrary character belongs to which of the similar character categories will be described below as “similar character identification”.

  When this “similar character identification” is performed, for example, the original function is defined by the Mahalanobis distance, and a mixed identification function in which the above-described feature difference indicated by the similar characters is added as a correction term is performed. The calculation of this correction term will be described below.

  First, it is assumed that the difference between features of similar characters is regarded as a difference vector composed of average vectors (a, b) of category A and category B, and an arbitrary character to be identified: input pattern p is given.

  In this case, a value p1 projected from the input pattern p onto the difference vector is obtained. Then, m1 and m2 obtained from p1 and the average vector of each character category are correction terms for the category A and category B discrimination functions.

Then, the discriminant function value: f _A (p) between the category A and the input pattern p is obtained by the mixed discriminant function in which the correction term of the mixed discriminant function shown in FIG. 12 is m1, and similarly, the correction term is m2. The discriminant function value: f _B (p) between the category B and the input pattern p is obtained by the mixed discriminant function.

  Then, based on each calculated identification function value, it is identified whether the input pattern p belongs to category A or category B, and the category to which the input pattern p belongs is determined. This is equivalent to identifying the input pattern p using the difference in characteristics of similar characters.

  In addition, as an example of identification using a difference in characteristics of similar characters, there is a method of identifying similar characters using an emphasis vector in addition to a difference vector. The enhancement vector indicates a vector composed of an input pattern value and a vector in the normal direction on the identification boundary surface of similar characters (see, for example, Patent Document 1).

JP 2002-183664 A

  However, the above-described conventional technique has a problem in that it is not possible to identify a difference in characteristics of similar characters with high accuracy.

  For example, in the example illustrated in FIG. 12, the difference between the features of similar characters is calculated using a difference vector composed of the average vectors of category A and category B. However, the difference in features of the similar characters spreads over the feature space. In addition, other information indicated by vectors other than the difference vector exists in the feature space.

  Therefore, even if the difference between the features of similar characters is calculated using the difference vector, other information indicated by the vector other than the difference vector is not used. As a result, since other information necessary for identification is lacking, there is a limit to recognizing similar characters with high accuracy.

  The disclosed technology has been made in view of the above, and an object of the present invention is to provide a character identification program and a character identification device that identify differences between similar character categories with high accuracy.

  The character identification program disclosed in the present application specifies, for each character to be compared, a character category having a different shape and including an agreed character on the feature space, and obtains information common to each character category as a common distribution processing procedure A correction procedure for correcting each character category distribution based on the common distribution and each character category distribution, and when a character to be identified is acquired, the position of the character in the feature space, and the correction procedure The computer is caused to execute a character identification procedure for identifying the character based on the character category distribution corrected by the above.

  According to the character identification program disclosed in the present application, there is an effect that a difference between similar character categories is identified with high accuracy.

FIG. 1 is a diagram for explaining the outline of the embodiment. FIG. 2 is a diagram illustrating the character identification device according to the first embodiment. FIG. 3 is a diagram for explaining an example of the data structure of the similar character table. FIG. 4 is a diagram for explaining an example of the data structure of the character recognition dictionary. FIG. 5 is a diagram for explaining similar character identification processing. FIG. 6 is a flowchart illustrating the process of the character identification device according to the first embodiment. FIG. 7 is a diagram illustrating the character identification device according to the second embodiment. FIG. 8 is a diagram illustrating an example of the data structure of the similar character recognition dictionary. FIG. 9 is a diagram illustrating an example of an interface. FIG. 10 is a flowchart illustrating the process of the character identification device according to the second embodiment. FIG. 11 is a diagram illustrating a hardware configuration of a computer constituting the character identification device according to the first embodiment. FIG. 12 is a diagram for explaining a conventional mixed identification method.

  Embodiments of a character identification program and a character identification device disclosed in the present application will be described below with reference to the drawings. Note that the present invention is not limited to the embodiments.

  First, an outline of the character identification device according to the first embodiment will be described. The character identification device according to the first embodiment is initially given image data including an arbitrary character to be recognized, and determines whether the arbitrary character belongs to the character category A or the character category B. To identify any character. Note that the character category A and the character category B are similar to each other.

  In this way, identification of which arbitrary character belongs to which of the similar character categories will be described as “similar character identification”.

  When this “similar character identification” is performed, the character identification device according to the first embodiment obtains information common to similar character categories as a common distribution on the feature space, and each similar character indicates the obtained common distribution. Convolution is performed on the distribution, and identification is performed using the convolved distribution. Hereinafter, it demonstrates concretely using figures.

  FIG. 1 is a diagram for explaining the outline of the embodiment. In the example shown in FIG. 1, an arbitrary input pattern p is given on the feature space 10 shown in FIG. 12, and it is discriminated whether it is similar to category A or B using a mixed discriminant function.

  The categories A and B indicate the character categories described in the related art, and are defined using feature vectors whose parameters are character fluctuation, character line inclination, character area, etc., in a given character. The

  In general, since this feature vector is shown on the feature space, the above-mentioned “character category” is shown as a distribution on the feature space defined by the feature vector. This is called a character category distribution. For convenience, parameters such as character fluctuation, character line inclination, and character area are referred to as “features”.

  First, a common distribution 50 having a common shape is obtained from the distribution on the feature space indicated by each similar character category regardless of the value of the input pattern p (step S10). Note that “c” in the common distribution 50 indicates a vector in which the average vectors of category A and category B overlap.

  Then, when the input pattern p is regarded as the spread of the Gaussian distribution, the distribution of the input pattern p is assumed using the common distribution 50 obtained in step S10 (step S20).

  Then, the distribution of category A and the distribution of category B are transformed by the common distribution 50 by performing convolution integration for each category with respect to each distribution indicated by category A and category B (step S30). .

  In the above steps S10 to S30, the common distribution 50 is assumed around the vector of the input pattern p, so that the identification considering the distribution shape of the input pattern p is performed. That is, this is equivalent to performing identification in consideration of each feature amount of the input pattern p.

  When the common distribution 50 assumed as the distribution indicated by the input pattern p and the distributions indicated by the categories A and B are assumed to be normal distributions, the common distribution 50 is determined with respect to the distribution indicated by the categories A and B. Perform the convolution integral used.

  This is because, when the common distribution 50 is regarded as a Gaussian distribution, a difference between the similar character categories is less likely to appear in a direction with a large spread, and a difference between the similar character categories is likely to appear in a direction with a small spread. Think and focus on the direction of small spread. Therefore, the identification that emphasizes the feature of each similar character is led.

  Based on the distribution of category A deformed using the common distribution 50, a mixed discriminant function for category A (for example, a mixed discriminant function A) is obtained, and an discriminant function value A is calculated from the mixed discriminant function A. Is done.

  On the other hand, a mixed discriminant function for category B (for example, a mixed discriminant function B) is obtained based on the distribution of category B deformed using the common distribution 50, and the discriminant function value B is calculated from the mixed discriminant function B. Is done.

  Then, the input pattern p is identified based on the identification function value A and the identification function value B. In the example shown in FIG. 1, the discriminant function value between the category A and the input pattern p is “0.6”, and the discriminant function value between the category B and the input pattern p is “0.4”.

  As described above, since the identification is performed with an emphasis on the direction in which the spread is small, the smaller the calculated identification function value, the higher the similarity, and the input pattern p is similar to the category B. Become.

  In this way, arbitrarily given input characters are regarded as a common distribution, the common distribution is subjected to convolution integration for each similar character with respect to the distribution on the feature space indicated by the similar characters, and the convolution distribution is used. By performing the identification, the difference between similar characters can be identified with high accuracy.

  Next, the configuration of the character identification device according to the first embodiment will be described. When identifying a similar character, the character identification device 100 according to the first embodiment obtains a common distribution between similar characters, and performs convolution integration on the distribution in the feature space indicated by each similar character.

  Further, eigenvalues and eigenvectors are calculated from values calculated by convolution integration, and pseudo Bayes identification is performed using the calculated eigenvalues and eigenvectors. This will be specifically described with reference to the drawings. FIG. 2 is a diagram illustrating the character identification device according to the first embodiment.

  The character identification device 100 illustrated in FIG. 2 includes an image input unit 101, a processing unit 102, and a recognition result output unit 103. The image input unit 101 is an input device that optically reads a character image including character data, and the character image read by the image input unit 101 is output to the processing unit 102.

  The character data included in the character image described above corresponds to, for example, the input pattern p shown in FIG.

  The processing unit 102 is a processing unit that performs character recognition on the character image input from the image input unit 101 using conventional techniques, and further performs the above-described “similar character identification” on the acquired recognition result. , A similar character table 102a, a character recognition dictionary 102b, and a character recognition processing unit 102c.

  The similar character table 102a is a table that stores a list of similar characters, and stores similar characters for a predetermined character. FIG. 3 is a diagram illustrating an example of the data structure of the similar character table.

  ID: 0 shown in FIG. 3 indicates the character category “bird”, the number of similar characters corresponding to this “bird” is “1”, and the corresponding similar character category is “crow”. Become. Hereinafter, the similar character category corresponding to the character category “乎” is “flat”, “pa” corresponds to “ba”, and “large” corresponds to “dog” and “thick”.

  The number of similar characters included in the similar character category described above is also symmetrical to the number of similar characters corresponding to the character category. For example, the number of characters similar to “crow” is one, and it is stored that it is “bird”.

  In addition, the number of characters similar to “dog” is two, and it is stored that they are “large” and “fat”.

  Returning to FIG. 2, the character recognition dictionary 102b will be described. The character recognition dictionary 102b is a table that stores an average vector of each character category, eigenvalues and eigenvectors corresponding to each character category. This will be specifically described with reference to the drawings.

FIG. 4 is a diagram for explaining an example of the data structure of the character recognition dictionary. The character recognition dictionary 102b in FIG. 4 includes an “average vector”, “eigenvalue”, and “eigenvector” that each similar character has.
Is remembered.

  Taking the categories A and B shown in FIG. 1 as an example, an average vector “a” corresponding to the category A, an eigenvalue “k1” corresponding to the distribution of the category A, and an eigenvector “x1” are stored.

  On the other hand, for category B, an average vector “b” corresponding to category B, an eigenvalue “k2” corresponding to the distribution of category B, and an eigenvector “x2” are stored.

  Next, returning to the description of FIG. 2, the character recognition processing unit 102c will be described. The character recognition processing unit 102c is a processing unit that recognizes characters included in the input character image and performs similar character identification on the acquired recognition result.

  First, a character recognition process using conventional technology is performed on an input character image (for example, including an input pattern p), and a plurality of character recognition candidates for the input pattern p (hereinafter simply referred to as character candidates). ) To get.

  Then, in the plurality of acquired character candidates, it is determined whether or not the first character candidate is registered in the similar character table 102a. Based on the determination result, the first character candidate is registered in the similar character table 102a. If not, the first character candidate is taken as the character recognition result of the input pattern p.

  On the other hand, when the first character candidate is registered in the similar character table 102a, it is a character candidate character included in the second to Nth characters, and is similar to the first character candidate described above. It is determined whether or not the character to be included is included in the similar character table 102a.

  If the character similar to the first character candidate is not registered in the similar character table 102a according to the determination result, the first character candidate is output as the character recognition result of the input pattern p.

  On the other hand, if a character similar to the first character candidate is registered in the similar character table 102a, the highest character candidate is selected from the character candidates included in the second to Nth characters, Similar character identification determination is performed between the selected character candidate and the first character candidate.

  The process of the character recognition processing unit 102c described above will be described with a specific example. For example, when “cat” is acquired from the input pattern p as the first character candidate of the input pattern p, the acquired “cat” is searched from the similar character table 102a.

  Then, if “cat” is not registered in the similar character table 102a based on the search result, the character recognition processing unit 102c sets “cat” as the character recognition result of the input pattern p.

  On the other hand, when “cat” is registered in the similar character table 102a, the second to Nth character candidates for the input pattern p are referred to and the first candidate “cat” is selected. It is determined whether similar characters are registered in the similar character table 102a.

  If no character similar to the first candidate “cat” is registered as a result of the determination, the first candidate “cat” is set as the character recognition result of the input pattern p.

  On the other hand, among the 2nd to Nth character candidates, “Draw” was acquired as the 2nd character candidate, and “Ikari” was acquired as the 6th character candidate. In this case, when the acquired “draw” and “Ikari” are registered in the similar character table 102a, the following similar character identification is performed.

  In this case, among the 2nd to Nth positions, the second highest character candidate “Egaku”, which is the highest candidate, and the similar character relating to the first candidate “cat” Identify. As a result of the processing, if the input pattern p is highly similar to the first candidate “cat”, the first candidate “cat” is set as the character recognition result of the input pattern p.

  On the other hand, if the degree of similarity with the second-ranked character candidate “Draw” is high, the second-ranked character candidate “Draw” is set as the character recognition result of the input pattern p.

  Note that even if either “cat” or “draw” is used as the character recognition result, similar character identification for the sixth candidate character “Ikari” is not performed, When “Ikari” is the second candidate character, similar character identification with the first candidate “Cat” is performed.

  Next, similar character identification described above will be described with reference to the drawings. FIG. 5 is a diagram for explaining similar character identification processing. A case will be described in which the character category A or the character category B is identified when an arbitrary input pattern p is given.

First, the character recognition processing section 102c is making a similar character identification, to calculate a common distribution sigma _C category A and category B (step S50). As an example for obtaining the common distribution sigma _C, it will be described as an example a case obtained by using a simple average method.

The covariance matrix of the category A and C _A, if the covariance matrix of the category B was C _B, the sum of the covariance matrix as a sum of C _A and C _B: Request C _X. Then, the top n eigenvalues and eigenvectors are calculated from the obtained _CX .

Then, the calculated eigenvalues and projection matrix from eigenvectors: by obtaining a P _X, and covariance matrix corresponding to the common distribution sigma _C category A and category B. Incidentally, heretofore a common distribution sigma _C is determined by "common operation" below in addition to the simple average of the methods described above will be described in detail later in this common operation.

Then, assuming a common distribution sigma _C determined in step S50 around the vector indicated by the input pattern p, by using the common distribution sigma _C, for each distribution shown the category A and category B, and convolution The deformation of each category distribution is calculated (step S51).

For example, the input when the feature vector of the pattern p and x, relates to a point x in the feature space, the expected value of the discriminant function value of the secondary discriminant function category C to a common distribution D _p centered on the x and density function And

In this case, when a normal secondary discriminant function is defined by the following equation (1), the following equation (2) is further obtained from this equation (1).

In Equation (2), m _C represents an average vector of category C, Σ _C represents a covariance matrix of category C, and Σ represents a covariance matrix of common distribution D _p . This, in normal quadratic discriminant function, corresponds to replacing the sigma _C to Σ + Σ _C.

Then, the character recognition processing unit 102c, based on the equation (2), and covariance matrix shown the common distribution sigma _C, calculates the sum of the covariance matrix shown the distribution of categories A, new eigenvalues and eigenvectors calculate.

On the other hand, the character recognition processing unit 102c, similarly to Category A with respect to Category B, using equation (2), the common distribution sigma _C indicates a covariance matrix and the covariance matrix shown the distribution of Category B The sum is calculated, and the eigenvalue / eigenvector is newly calculated.

  Next, the character recognition processing unit 102c determines whether the input pattern p is similar to the category A or the category B from the pseudo Bayes identification with the category A and the category B based on the eigenvalue / eigenvector obtained in step S51. The determination result is input to the recognition result output unit 103 (step S52).

  For example, as shown in FIG. 1, when the discrimination function value A corresponding to the category A is “0.6” and the discrimination function value A corresponding to the category B is “0.4”, the calculated discrimination function The smaller the value is, the higher the degree of similarity is. Therefore, the input pattern p is similar to the category B.

  In this way, the input pattern p can be considered as one vector in the feature space. However, by assuming a common distribution around the input pattern vector, identification based on the shape of each distribution indicated by category A and category B is possible. Will do.

  Then, assuming that the common distribution of the input pattern p and the shapes of the distributions indicated by the categories A and B are normal distributions, the convolution integral by the common distribution is performed for each category on the secondary discriminant function of each category distribution.

  This corresponds to a quadratic discriminant function in which the function obtained by performing convolution integration on category A is replaced with the sum of the covariance matrix indicated by the common distribution and the covariance matrix indicated by category A.

  On the other hand, the function obtained by performing convolution integration on category B corresponds to a quadratic discriminant function replaced with the sum of the covariance matrix indicated by the common distribution and the covariance matrix indicated by category B.

  Next, returning to the description of FIG. 2, the recognition result output unit 103 will be described. The recognition result output unit 103 is a processing unit that outputs a character recognition result based on similar character identification performed by the processing unit 102.

  As described above, as an example of obtaining the common distribution, the case of obtaining using the simple average method has been described as an example, but the common distribution is also calculated by the “common operation” described below.

  When the common distribution is calculated using the simple average method, when the similar character category A and the similar character category B are given, the deformation (common distribution) common to the category A and the category B is the covariance matrix of the category A And the category B covariance matrix are obtained, and the top n eigenvalues and eigenvectors are calculated.

  However, since the top n eigenvalues and eigenvectors may include differences between similar character categories, the other distribution is an example of a method for obtaining a common distribution that does not include differences between similar character categories. There is a method of cutting ingredients that are not included.

  For example, when the input pattern p has an extreme spread with respect to the category A side, the common distribution calculated by the simple average method becomes a common distribution influenced by the contribution of the category A side.

  Therefore, there is a common operation as a method for calculating a common distribution that does not include a difference between similar character categories. Specifically, a more accurate common distribution is calculated by cutting a component that is not included in the other distribution. Hereinafter, the common operation will be described in detail.

First, common conversion using the distribution of category B is performed on the distribution of category A. At this time, if the covariance matrix of category A is C _A and the covariance matrix of category B is C _B , first, a transformation matrix T _CA based on eigenvalues / eigenvectors is calculated from C _A.

Then, a transformation matrix T based on the eigenvalue / eigenvector is obtained. In this case, the eigenvalue of the covariance matrix C is λ _i (i = 1, 2,..., K (? N)), and the eigenvector is Φ _i (i = 1, 2,..., K (? N)). Then, if Φ _{k + 1} ... Φ _n is taken to be an orthonormal basis with Φ _i (i = 1, 2,..., K (? N)), the transformation matrix T is expressed by the following formula (3 ).

Then, a portion common to the category A distribution is extracted from the category B distribution using the following equation (4).

Next, the common conversion using the category A distribution is performed on the category B distribution. In this case, by performing the same processing as described for category A, the following expression (5) is obtained.

Then, based on C ′ _B shown in Expression (4) and C ′ _A shown in Expression (5), the upper n eigenvalues and eigenvectors of C ′ _B + C ′ _A are calculated. Then, the common distribution is calculated using the calculated eigenvalue and eigenvector.

  Next, processing of the character identification device 100 shown in the first embodiment will be described. FIG. 6 is a flowchart illustrating the process of the character identification device according to the first embodiment.

  First, the character identification device 100 acquires a character image, and performs normal character recognition processing on the acquired character image (for example, including the input pattern p) (step S100). Then, a plurality of character recognition candidates for the input pattern p are acquired (step S101).

  Then, it is determined whether or not the first character candidate is registered in the similar character table 102a among the plurality of character candidates, and if the first character candidate is not registered according to the determination result (step) S102, No) sets the first character candidate as the result of character recognition.

  On the other hand, if the first character candidate is registered in the similar character table 102a (step S102, Yes), the first character candidate from the second character to the nth character acquired in step S101 is the first character candidate. It is determined whether or not a character similar to the character candidate is registered in the similar character table 102a (step S103).

  As a result of the determination, if the similar character of the first character candidate is not registered in the similar character table 102a (No in step S104), the character candidate registered in the n + 1st character is acquired (step S105). ).

  If all candidate characters acquired in step S101 are processed (No in step S106), the first character candidate is output as a recognition result.

  On the other hand, when processing has not been performed for all candidate characters acquired in step S101 (step S106, Yes), the process proceeds to step S104.

  Further, according to the determination result in step S103, characters that are character candidates included in the second to Nth positions and similar to the first character candidate described above are registered in the similar character table 102a. If yes (step S104, Yes), similar character identification is performed between the top candidate character and the first character candidate included in the second to Nth character candidates (step S107).

  According to this flowchart, the character identification device 100 obtains the difference between features indicated by similar characters such as “bird” and “bird” on the feature space, thereby obtaining each similar character. Can be identified with high accuracy.

  Next, the character identification device shown in Example 2 will be described. FIG. 7 is a diagram illustrating the character identification device according to the second embodiment. The character identification device 200 shown in FIG. 7 is a character identification device that performs similar character identification based on similar characters designated by the user.

  This “similar character identification” indicates that an arbitrary character belongs to one of character categories similar to each other as described in the first embodiment.

  The character identification device 200 includes an image input unit 201, a processing unit 202, a recognition result output unit 203, and an interface 204.

  Moreover, the character identification device 200 has substantially the same function as the character identification device 100 shown in FIG. 2, and detailed description of the functional unit having the same function as the character identification device 100 is omitted.

  Specifically, the difference from the character identification device 100 is that the processing unit 202 has a similar character recognition dictionary 202 c and an interface 204. Hereinafter, the similar character recognition dictionary 202c and the interface 204 will be described.

  FIG. 8 is a diagram illustrating an example of the data structure of the similar character recognition dictionary. The similar character recognition dictionary 202c shown in FIG. 8 stores two character categories similar to each other.

  When the input pattern p is given to the characters registered in these two character categories, the user can correctly identify the input pattern p and the characters registered in the corresponding two character categories. Indicates the character data desired to be

  “Number of samples” indicates the number of similar character data similar to the corresponding character category. For example, the number of samples corresponding to “bird” and “crow” is “2”. Show.

  The same number of vectors as the “number of samples” are stored as common vectors. The detailed data structure of the “number of samples” described above and the “common vector” shown in FIG. 8 will be described later with reference to FIG.

  Next, the interface 204 illustrated in FIG. 7 will be described with reference to the drawings. FIG. 9 is a diagram illustrating an example of an interface. The interface 204 shown in FIG. 9 is an interface in which the user designates a character category that the user desires to be correctly identified and similar character data similar to the character category, each as a set of images.

  In the example shown in FIG. 9, “designation of characters” at the top of the window indicates an interface through which a user inputs characters using a keyboard or the like (not shown). Then, “BA” is input to the character category A, and “PA” is input to the character category B.

  Therefore, when an arbitrary input pattern p is given, the user wants to accurately identify whether the input pattern p belongs to “pa” or “ba”.

  When the input pattern p is identified as belonging to “pa” or “ba”, similar characters corresponding to the “pa” and “ba” pairs registered in “character designation” are identified. The data is registered in “image designation” shown at the bottom of the interface 204.

  The number indicated by the “sample ID” in the “designation of image” corresponds to the “number of samples” in FIG. The “similar character data” described above indicates data registered by “designation of image”.

  The character recognition processing unit 202d calculates a common vector based on the set of images registered in the “designation of image”, and registers the calculated common vector in the similar character recognition dictionary 202c in FIG.

The processing performed by the character recognition processing unit 202d for calculating the common vector will be specifically described. First, assuming that feature vectors obtained from two given images are x _A and x _B , a common vector is calculated by the following equation (6).

In the above equation (6), ((x _A −x _B ) · x _A ) represents an inner product of vectors of x _A −x _B and x _A, and (x _A −x _B ) / | x _A −x _B | represents a unit vector.

  Moreover, each common vector shown by Formula (6) will have the same value, and if any one common vector is calculated, the corresponding common vector will be calculated | required.

  Then, the character recognition processing unit 202d registers the common vector calculated using Equation (6) in the similar character recognition dictionary 202c shown in FIG.

  Next, similar character identification performed by the character recognition processing unit 202d will be described. First, when an arbitrary input pattern p is given, the character recognition processing unit 202d searches whether a set of character categories similar to the input pattern p is registered in the similar character recognition dictionary 202c.

Then, if registered based on the search result, the character recognition processing unit 202d acquires a common vector corresponding to a set of character categories similar to the input pattern p, and uses the following equation (7) to obtain a common distribution. to add.

In Expression (7), “C” corresponds to the common distribution (for example, Σ _C ) shown in the first embodiment, and a new common distribution corresponding to this common distribution is C ′. The method for obtaining the common distribution sigma _C, as shown in Example 1, may be calculated using the simple average method may be calculated by using the "common operation", detailed calculation Since the method is the same as that of the first embodiment, a description thereof will be omitted.

  After that, the character recognition processing unit 202d uses the common distribution obtained by Expression (7) to categorize the character category “A” (for example, “BA”) and the character category “B” (for example: “PA”). On the other hand, convolution integration is performed. Then, pseudo Bayes identification is performed based on eigenvalues and eigenvectors obtained by convolution integration.

  In the example shown in FIG. 9, the similar character data is shown as image data. However, for convenience of explanation, this is merely an example of image data, and this embodiment can be used even if it is replaced with other data. It is assumed that the function shown in (1) is obtained.

  Next, processing will be described for the character identification device 200 shown in the second embodiment. FIG. 10 is a flowchart illustrating the process of the character identification device according to the embodiment.

  In advance, the user registers a character category that the user wants to recognize correctly through the interface 204 and similar character data in which a similar character category similar to the character category is a set of images.

  Then, the character recognition processing unit 202d calculates a common vector based on the registered similar character data, and registers the calculated common vector in the similar character recognition dictionary 202c (step S200).

  The character identification device 200 acquires a character image, performs normal character recognition processing on the acquired character image (step S201), and acquires a plurality of character recognition candidates (step S202).

  Subsequently, it is determined whether or not the first character candidate is registered in the similar character table 202b among the plurality of character candidates. If the first character candidate is not registered based on the determination result ( In step S203, No), the first character candidate is set as a result of character recognition.

  On the other hand, when the first character candidate is registered (step S203, Yes), among the character candidates from the second character to the nth character, a character similar to the first character candidate is a similar character. It is determined whether or not it is registered in the table 202b (step S204).

  If a character similar to the first character candidate is not registered in the similar character table 202b based on the determination result (No in step S205), the character registered in the n + 1st character is acquired (step S206). .

  If all candidate characters acquired in step S201 are processed (No in step S207), the first character candidate is output as a recognition result.

  On the other hand, when processing has not been performed for all candidate characters acquired in step S201 (step S207, Yes), the process proceeds to step S205.

  If characters similar to the first character candidate from the second character to the nth character are stored in the similar character table 202b (Yes in step S205), the similar character is identified. This is performed (step S208).

  In step S208, when the character recognition processing unit 202d performs similar character identification processing, it is searched whether or not the common vector registered in step S200 is registered in the similar character recognition dictionary 202c.

  If registered according to the search result, the character recognition processing unit 202d adds the referenced common vector to the common distribution. Then, similar character identification is performed based on the newly obtained common distribution.

  According to the above-described embodiment, the feature information common to similar character categories is extracted, and the features common to similar character categories are suppressed, and the feature differences are emphasized. It is possible to identify similar characters with higher accuracy by utilizing the difference in characteristics between the expanded similar character categories.

  By the way, among the processes described in the present embodiment, a part of the processes described as being automatically performed can be manually performed, or all of the processes described as being manually performed or A part can be automatically performed by a known method. In addition, the information including the processing procedure, control procedure, specific name, and various data shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

  Further, each component of the character identification device 100 shown in FIG. 2 is functionally conceptual and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured. Furthermore, each processing function performed by each device may be realized by a CPU and a program that is analyzed and executed by the CPU, or may be realized as hardware by wired logic.

  FIG. 11 is a diagram illustrating a hardware configuration of a computer constituting the character identification device according to the first embodiment. As shown in FIG. 11, the computer 300 includes an input device 301, a display 302, a RAM (Random Access Memory) 303, a ROM (Read Only Memory) 304, an HDD (Hard Disk Drive) 305, and a CPU (Central Processing Unit) 306. The medium reader 307 is connected by a bus 308.

  The HDD 305 stores a character recognition program 305a that performs the same function as the similar character identification function described above. When the CPU 306 reads and executes the character recognition program 305a, the character recognition process 306a is activated. Here, the character recognition process 306a corresponds to the character recognition processing unit 102c shown in FIG.

  The RAM 303 stores various data 303a including various data stored in the processing unit 102 and data used by the character recognition process 306a. The CPU 306 identifies similar characters based on the average vector, eigenvalue, and eigenvector of each character category included in the various data 303a.

  By the way, the character recognition program 305a shown in FIG. 11 does not necessarily have to be stored in the HDD 305 from the beginning. For example, a “portable physical medium” such as a flexible disk (FD), a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card inserted into a computer, or a hard disk drive (HDD) provided inside or outside the computer. Are stored in the “fixed physical medium”, and “another computer (or server)” connected to the computer via a public line, the Internet, a LAN, a WAN, or the like. 305a may be read and executed.

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

(Supplementary note 1)
A common distribution processing procedure for identifying a character category distribution having different shapes and including agreed characters on the feature space for each character to be compared, and obtaining information common to each character category distribution as a common distribution,
A correction procedure for correcting each character category distribution based on the common distribution and each character category distribution;
When a character to be identified is acquired, a character identification procedure for identifying the character is executed based on the position of the character in the feature space and each character category distribution corrected by the correction procedure. Character recognition program

(Supplementary Note 2) The correction procedure performs convolution integration using the common distribution on the character category distribution to correct the character category distribution, and the character identification procedure is based on the corrected character category distribution. The character identification program according to supplementary note 1, for identifying the character.

(Additional remark 3) When the said common distribution process procedure calculates | requires the said common distribution, the difference between the said character category distribution is reduced, and the said common distribution is calculated | required based on the information common between the said character category distributions. Or the character identification program of 2.

(Additional remark 4) The said common distribution process procedure calculates the information common to the said image data based on the image data input by the user, adds the calculated information to the said common distribution, and calculates | requires a new common distribution The character identification program as described in any one of appendix 1-3.

(Supplementary Note 5) A common distribution processing unit that specifies a character category distribution having different shapes and including agreed characters on the feature space for each character to be compared, and obtaining information common to each character category distribution as a common distribution;
A correction unit for correcting each character category distribution based on the common distribution and each character category distribution;
A character identification unit that identifies the character based on the position of the character in the feature space and each character category distribution corrected by the correction unit when a character to be identified is acquired; Character identification device.

(Additional remark 6) The said correction part performs the convolution integral using the said common distribution with respect to the said character category distribution, and corrects the said character category distribution, The said character identification part is based on the said corrected character category distribution. The character identification device according to attachment 5, which identifies the character.

(Additional remark 7) When the said common distribution process part calculates | requires the said common distribution, the difference between the said character category distributions is reduced, and the said common distribution is calculated | required based on the information common between the said character category distributions. Or the character identification device of 6.

(Additional remark 8) The said common distribution process procedure calculates the information common to the said image data based on the image data input by the user, adds the calculated information to the said common distribution, and calculates | requires a new common distribution The character identification device according to any one of appendices 5 to 7.

(Supplementary note 9) A method by which a character identification device identifies a character,
The character identification device is
A common distribution processing step for identifying a character category distribution having different shapes and including agreed characters on the feature space for each character to be compared, and obtaining information common to each character category distribution as a common distribution;
A correction step of correcting each character category distribution based on the common distribution and each character category distribution;
When a character to be identified is acquired, a character identification step for identifying the character based on the position of the character in the feature space and each character category distribution corrected by the correction step is executed. Character identification method.

DESCRIPTION OF SYMBOLS 10 Feature space 50 Common distribution 100, 200, 300 Character identification device 101, 201 Image input part 102, 202 Processing part 102a, 202b Similar character table 102b, 202a Character recognition dictionary 103, 203 Recognition result output part 202c Similar character recognition dictionary 204 Interface 301 Input device 302 Display 303 RAM
303a Various data 304 ROM
305 HDD
305a Character recognition program 306 CPU
306a Character recognition process 307 Media reader A, B Character category distribution

Claims (5)

On the computer,
A common distribution processing procedure for identifying a character category distribution having different shapes and including agreed characters on the feature space for each character to be compared, and obtaining information common to each character category distribution as a common distribution,
A correction procedure for correcting each character category distribution based on the common distribution and each character category distribution;
When a character to be identified is acquired, a character identification procedure for identifying the character is executed based on the position of the character in the feature space and each character category distribution corrected by the correction procedure. Character recognition program   The correction procedure corrects the character category distribution by performing convolution integration using the common distribution on the character category distribution, and the character identification procedure is based on the corrected character category distribution. The character identification program according to claim 1 for identifying a character.   3. The method according to claim 1, wherein when the common distribution processing procedure obtains the common distribution, the difference between the character category distributions is reduced, and the common distribution is obtained based on information common to the character category distributions. The character identification program described.   The common distribution processing procedure calculates information common to the image data based on image data input by a user, adds the calculated information to the common distribution, and newly obtains a common distribution. 4. The character identification program according to any one of 3. On the feature space, a common distribution processing unit that specifies a character category distribution that includes differently shaped and consensus characters for each character to be compared, and obtains information common to each character category distribution as a common distribution;
A correction unit for correcting each character category distribution based on the common distribution and each character category distribution;
A character identification unit that identifies the character based on the position of the character in the feature space and each character category distribution corrected by the correction unit when a character to be identified is acquired; Character identification device.

Images