JP5974576B2 - Character recognition learning device, character recognition device, and program - Google Patents

Description

  The present invention relates to a character recognition learning device, a character recognition device, and a program.

  Patent Document 1 describes a character recognition device that switches a coordinate correction method according to a position in an input frame in recognition of a character handwritten in the input frame. In Patent Document 2, in recognition of handwritten characters, characters that may be written out of the entry frame, such as muddy and semi-voiced dots, are registered in advance, and when the registered characters are recognized, adjacent entries are entered. A character recognition device that determines the presence or absence of a cloudy point, a semi-cloudy point, or the like in a frame is disclosed. Patent Document 3 describes a handwritten character recognition device that normalizes the result of online recognition and the result of offline recognition and performs character recognition processing based on these results. Patent Document 4 describes a character recognition method in which recognition results obtained by a plurality of character recognition means are integrated and a final recognition result is generated based on the integrated results. Patent Document 5 describes a character recognition method that performs character recognition using a plurality of character recognition programs, compares the respective outputs, and outputs a character recognition result. Patent Document 6 describes a character recognition device that performs character recognition based on the time series of strokes input online and the shape of the entire characters input offline.

JP-A-8-96082 JP-A-9-190497 JP 2000-105798 A JP 2001-22883 A Japanese Patent Laid-Open No. 7-114620 JP-A-9-288717

  One of the objects of the present invention is a character recognition learning device that generates learning data using a combination of recognition results obtained by a plurality of character recognition means in a character recognition process using a plurality of character recognition means, and a plurality of character recognitions. An object of the present invention is to provide a character recognition device that performs character recognition processing using a combination of recognition results obtained by means.

  In order to achieve the above object, the invention according to claim 1 is directed to teacher information acquisition means for acquiring teacher information indicating each character type of a plurality of characters included in a learning document, and to each of the plurality of characters. The first recognition result information acquisition means for acquiring the first recognition result information indicating the result of the character recognition processing by the first character recognition means is different from the first character recognition means for each of the plurality of characters. Second recognition result information acquisition means for acquiring second recognition result information indicating a result of character recognition processing by the second character recognition means using the processing method; and each of the plurality of characters is included in the teacher information. Machine learning is performed using information indicating the character type of a character and a combination of the first recognition result information and the second recognition result information acquired for the character as input data. And a character recognition learning device characterized by comprising a learning data output means for outputting the execution result as a character recognition training data.

  Further, the invention according to claim 2 is the character recognition learning device according to claim 1, wherein the first recognition result information includes the first recognition result information for each of the plurality of characters as a recognition result of the character. Information indicating the character type recognized by one character recognition means, and information indicating the likelihood, and the second recognition result information includes, as the character recognition result for each of the plurality of characters, The character recognition learning apparatus includes information indicating a character type recognized by the second character recognition means and information indicating the likelihood.

  The invention according to claim 3 is the character recognition learning device according to claim 1, wherein the first recognition result information is predetermined as a recognition result of each of the plurality of characters. Each of the plurality of character type candidates includes information indicating the likelihood that the character type of the character is the character type candidate, and the second recognition result information includes the recognition result of the character for each of the plurality of characters. For each of a plurality of predetermined character type candidates, the character recognition learning device includes information indicating the likelihood that the character type of the character is the character type candidate.

  The invention according to claim 4 is the character recognition learning device according to claim 3, wherein the learning data output means sets a likelihood smaller than a predetermined value in the first recognition result information to 0. In the second recognition result information, the machine learning is performed by replacing the likelihood smaller than a predetermined value with 0 and executing the machine learning.

  The invention according to claim 5 is the first recognition result information acquisition for acquiring the first recognition result information indicating the result of the character recognition processing by the first character recognition means for the determination target character included in the determination target document. 2nd recognition result information which acquires the 2nd recognition result information which shows the result of the character recognition process by the 2nd character recognition means using the processing method different from a said 1st character recognition means with respect to the said determination object character Acquisition means; teacher information indicating character types of a plurality of characters included in the learning document; a recognition result of the first character recognition unit for the learning document; and the second character recognition for the learning document. A character recognition learning data acquisition means for acquiring character recognition learning data generated using a combination of the recognition result by the means, and the determination target character. Character type for executing a character type determination process for determining a character type of the determination target character using the combination of the first recognition result information and the second recognition result information acquired for the target character and the learning data for character recognition. A character recognition device including a determination unit.

  According to a sixth aspect of the present invention, there is provided a teacher information acquisition means for acquiring teacher information indicating a character type of each of a plurality of characters included in a learning document, and a first information for each of the plurality of characters. A first recognition result information acquisition means for acquiring first recognition result information indicating a result of character recognition processing by the character recognition means, and a processing method different from the first character recognition means for each of the plurality of characters. Second recognition result information acquisition means for acquiring second recognition result information indicating a result of character recognition processing by the second character recognition means to be used; and for each of the plurality of characters, the character type of the character included in the teacher information And machine learning using the combination of the first recognition result information and the second recognition result information acquired for the character as input data A learning data output means for outputting the execution result as a character recognition training data is a program for functioning as a.

  According to the seventh aspect of the present invention, the computer recognizes the first recognition result information indicating the result of the character recognition processing by the first character recognition means for the determination target character included in the determination target document. A second recognition result information indicating a result of a character recognition process performed by a second character recognition unit using a result information acquisition unit and a processing method different from the first character recognition unit for the determination target character; Recognition result information acquisition means, teacher information indicating the character types of a plurality of characters included in the learning document, recognition results by the first character recognition means for the learning document, and the second for the learning document. A character recognition learning data acquisition means for acquiring character recognition learning data generated using the combination of the recognition result by the character recognition means, and the determination target character About the character type determination process for determining the character type of the determination target character using the combination of the first recognition result information and the second recognition result information acquired for the determination target character and the learning data for character recognition. Is a program for functioning as a character type determination means for executing.

  According to the first and sixth aspects of the present invention, the character is based on the information indicating the character type included in the teacher information and the combination of the recognition result by the first character recognition unit and the recognition result by the second character recognition unit. Learning data for recognition is generated.

  According to the second aspect of the present invention, information indicating the character type recognized by the first character recognition means and information indicating the likelihood are acquired as the first recognition result information, and the second recognition result information As such, information indicating the character type recognized by the second character recognition means and information indicating the likelihood thereof are acquired.

  According to the invention of claim 3, as the first recognition result information, the character type of the character for each of a plurality of predetermined character type candidates as the character recognition result for the character included in the learning document is Information indicating the likelihood of being a character type candidate is acquired, and as the second recognition result information, for each of a plurality of character type candidates determined in advance as a character recognition result, Information indicating the likelihood that the character type of the character is the character type candidate is acquired.

  According to the invention of claim 4, the likelihood smaller than the predetermined value in the first recognition result information is replaced with 0, and the likelihood smaller than the predetermined value in the second recognition result information is replaced with 0.

  According to the inventions according to claims 5 and 7, for the determination target character, a combination of the first recognition result information and the second recognition result information acquired for the determination target character, the learning data for character recognition, Is used to determine the character type.

It is a figure which shows the structure of the character recognition learning apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the example of the 1st recognition result information which concerns on the 1st Embodiment of this invention, and 2nd recognition result information. It is a figure which shows an example of the integrated recognition result information which concerns on the 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the character recognition learning apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the structure of the character recognition apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows the example of the 1st recognition result information which concerns on the 2nd Embodiment of this invention, and 2nd recognition result information. It is a figure which shows an example of the integrated recognition result information which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows operation | movement of the character recognition apparatus which concerns on the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
FIG. 1 is a diagram showing a configuration of a character recognition learning apparatus 100 according to the first embodiment of the present invention. The character recognition learning device 100 is connected to an input reception device 200 that recognizes a user's handwriting action such as a tablet PC or an electronic pen.

  The character recognition learning device 100 includes a control unit 110, an external I / F unit 120, a first character recognition unit 130, a second character recognition unit 140, a recognition result information integration unit 150, a learning data generation unit 160, and a storage unit 170. Have. The control unit 110, the external I / F unit 120, the first character recognition unit 130, the second character recognition unit 140, the recognition result information integration unit 150, and the learning data generation unit 160 operate according to programs stored in the storage unit 170. This is realized as one function of the CPU. The storage unit 170 includes a storage device such as a hard disk or a memory.

  In the input receiving device 200, a writing tool (touch pen, electronic pen, fingertip, etc.) is lowered to a recording medium (tablet PC display, paper, etc.) and released from (pen down, start of contact) (end of pen up, contact end). The electric signal indicating the movement of the tip of the writing tool up to) is generated and output to the character recognition learning device 100.

  Here, what is input by the writing instrument is data of a written document for learning composed of, for example, several hundred to several thousand characters given in advance.

  Further, the input reception device 200 acquires teacher information indicating the actual character type of each of the plurality of characters described in the learning document and outputs the teacher information to the character recognition learning device 100. The teacher information is composed of, for example, character codes of a plurality of characters described in a learning document.

  The control unit 110 of the character recognition learning device 100 controls the operation of each unit of the character recognition learning device 100. The control unit 110 acquires input information and teacher information indicating the input content in the input receiving device 200 input from the input receiving device 200 via the external I / F unit 120. The control unit 110 stores information indicating the input content and teacher information in the storage unit 170.

  The external I / F unit 120 acquires a signal input from the input reception device 200 and outputs the signal to the control unit 110.

  The first character recognizing unit 130 and the second character recognizing unit 140 are provided for the input information indicating the document for learning composed of a plurality of characters acquired from the input receiving device 200 and stored in the storage unit 170. Character recognition processing for recognizing (determining) each character type of a plurality of characters is performed. The first character recognition unit 130 and the second character recognition unit 140 perform character recognition processing using a known character recognition technique. Here, the 1st character recognition part 130 and the 2nd character recognition part 140 perform a character recognition process by a mutually different processing system. The first character recognition unit 130 generates first recognition result information indicating the likelihood that each of the plurality of characters is a plurality of predetermined character type candidates. Moreover, the 2nd character recognition part 140 produces | generates the 2nd recognition result information which shows the likelihood that it is each of several predetermined character type candidates about each of several characters. In the present embodiment, a configuration is shown in which two character recognition units (first character recognition unit 130 and second character recognition unit 140) are provided, but three or more character recognition units may be provided.

  Here, the plurality of predetermined character type candidates include, for example, all hiragana, katakana, kanji, and symbols in Japanese, and all alphabets and symbols in English. FIG. 2 is a diagram illustrating an example of first recognition result information and second recognition result information according to the first embodiment of the present invention. The first recognition result information and the second recognition result information in FIG. 2 indicate the likelihood of recognition as a predetermined character type for a character with a learning document.

  The recognition result information integration unit 150 generates integrated recognition result information by combining the first recognition result information generated by the first character recognition unit 130 and the second recognition result information generated by the second character recognition unit 140. To do. Specifically, the recognition result information integration unit 150 generates, for each of a plurality of characters, an m-dimensional vector whose component is a value output as the likelihood of the same character type. Here, m is a number for identifying a character recognition unit, and m = 0, 1 in the present embodiment in which two character recognition units are provided. FIG. 3 is a diagram illustrating an example of the integrated recognition result information according to the first embodiment of the present invention.

  For each of a plurality of characters, the learning data generation unit 160 performs integrated recognition combining information indicating the character type of the character included in the teacher information, and the first recognition result information and the second recognition result information acquired for the character. Machine learning is executed using the result information as input data, and the execution result is output as learning data for character recognition. The learning data generation unit 160 executes learning processing using a known learning method such as Boosting, neural network, SVM (support vector machine), or the like. The learning data generated by the learning data generation unit 160 is stored in the storage unit 170 and used for character recognition processing.

  Next, operation | movement of the character recognition learning apparatus 100 is demonstrated using a flowchart. FIG. 4 is a flowchart showing the operation of the character recognition learning apparatus 100 according to the first embodiment of the present invention. Here, the number of characters described in the learning document is L, and the number of character types is N.

  First, the control unit 110 initializes an array W [l] [n] [m] stored in the recognition result information integration unit 150 that stores vectors that are integrated recognition result information (S401). Here, l is a number for identifying the character in the learning document (0 ≦ l <L), n is a number for identifying the character type (0 ≦ n <N), m is the first character recognition unit 130 and the second character recognition unit. 140 is a number for identifying 140 (0 ≦ m <2).

  Next, the control unit 110 acquires input information and teacher information input from the input receiving device 200 (S402). The control unit 110 stores the acquired input information and teacher information in the storage unit 170.

  Next, the first character recognition unit 130 and the second character recognition unit 140 execute a character recognition process (S403). The result of the character recognition processing by the first character recognition unit 130 and the second character recognition unit 140 is output to the recognition result information integration unit 150.

  Next, the recognition result information integration unit 150 combines the character recognition results input from the first character recognition unit 130 and the second character recognition unit 140 to generate integrated recognition result information (S404). Specifically, the value included in the character recognition result is assigned to the array W [l] [n] [m] described above. In the example of the integrated recognition result information shown in FIG. 3, W [0] [0] [0] = 0.8, W [0] [0] [1] = 0.7, W [0] [1] [0] = 0.005 and W [0] [1] [1] = 0.08. The recognition result information integration unit 150 assigns values to the array W [l] [n] [m] for each value of 0 ≦ l <L, 0 ≦ n <N, and 0 ≦ m <2.

  Then, the learning data generation unit 160 performs learning processing using the integrated recognition result information generated in S403 (S405). The generated learning data is stored in the storage unit 170. Thus, the operation of the character recognition learning device 100 ends.

[Second Embodiment]
FIG. 5 is a diagram showing a configuration of a character recognition device 500 according to the second embodiment of the present invention. The character recognition device 500 is connected to an input reception device 600 that recognizes a user's handwriting action such as a tablet PC or an electronic pen.

  The character recognition device 500 includes a control unit 510, an external I / F unit 520, a first character recognition unit 530, a second character recognition unit 540, a recognition result information integration unit 550, a character identification unit 560, and a storage unit 570. Control unit 510, external I / F unit 520, first character recognition unit 530, second character recognition unit 540, recognition result information integration unit 550, and character identification unit 560 operate according to a program stored in storage unit 570. This is realized as a function of the CPU. The storage unit 570 includes a storage device such as a hard disk or a memory.

  In the input receiving device 600, a writing tool (touch pen, electronic pen, fingertip, etc.) is lowered to a recording medium (tablet PC display, paper, etc.) and released from (pen down, contact start) (pen up, contact end) The electric signal indicating the movement of the tip of the writing tool up to) is generated and output to the character recognition device 500. Here, the data of the determination target document is input by the writing instrument.

  The control unit 510 of the character recognition device 500 controls the operation of each unit of the character recognition device 500. Control unit 510 acquires input information and teacher information indicating input contents in input receiving device 600 input from input receiving device 600 via external I / F unit 520. Control unit 510 stores information indicating the input content and teacher information in storage unit 570.

  External I / F unit 520 acquires a signal input from input reception device 600 and outputs the signal to control unit 510.

  The first character recognizing unit 530 and the second character recognizing unit 540 obtain characters from the determination target document with respect to the input information indicating the determination target document acquired from the input receiving device 600 and stored in the storage unit 570. Character recognition processing for recognizing (determining) the character type is performed. The first character recognition unit 530 and the second character recognition unit 540 perform character recognition processing using a known character recognition technique. Here, the 1st character recognition part 530 and the 2nd character recognition part 540 perform a character recognition process by a mutually different processing system. The first character recognition unit 530 generates first recognition result information indicating the likelihood that each character (determination target character) included in the determination target document is a plurality of predetermined character type candidates. Moreover, the 2nd character recognition part 540 produces | generates the 2nd recognition result information which shows the likelihood that it is each of several predetermined character type candidates about the character contained in the determination object document. In the present embodiment, a configuration is shown in which two character recognition units (first character recognition unit 530 and second character recognition unit 540) are provided, but three or more character recognition units may be provided.

  Here, the plurality of predetermined character type candidates include, for example, all hiragana, katakana, kanji, and symbols in Japanese, and all alphabets and symbols in English. FIG. 6 is a diagram illustrating an example of the first recognition result information and the second recognition result information according to the second embodiment of the present invention. The first recognition result information and the second recognition result information in FIG. 6 indicate the likelihood of recognition for each of the predetermined character types for the characters on the determination target document.

  The recognition result information integration unit 550 generates integrated recognition result information by combining the first recognition result information generated by the first character recognition unit 530 and the second recognition result information generated by the second character recognition unit 540. To do. Specifically, the recognition result information integration unit 550 generates an m-dimensional vector whose component is a value output as the likelihood of the same character type for each of a plurality of characters. Here, m is a number for identifying a character recognition unit, and m = 0, 1 in the present embodiment in which two character recognition units are provided. FIG. 7 is a diagram showing an example of integrated recognition result information according to the second exemplary embodiment of the present invention.

  The character identification unit 560 uses, for the determination target character, the integrated recognition result information obtained by combining the first recognition result information and the second recognition result information acquired for the determination target character, and character recognition learning data. Then, character type determination processing for determining the character type of the determination target character is executed. Here, the learning data for character recognition includes teacher information indicating each character type of a plurality of characters included in the learning document, a recognition result by the first character recognition unit 530 for the learning document, and a second character for the learning document. Character recognition learning data generated in advance using a combination with a recognition result obtained by the recognition unit 540 and stored in the storage unit 570. The learning data for character recognition may be learning data generated by the same configuration as the character recognition learning device 100 according to the first embodiment, for example. The character identification unit 560 executes character identification processing using a known identification method such as Boosting, neural network, SVM (support vector machine), or the like.

  Next, the operation of the character recognition device 500 will be described using a flowchart. FIG. 8 is a flowchart showing the operation of the character recognition device 500 according to the second embodiment of the present invention.

  First, the control unit 110 initializes an array W [n] [m] stored in the recognition result information integration unit 550 that stores vectors that are integrated recognition result information (S801). Here, n is a number for identifying the character type (0 ≦ n <N), and m is a number for identifying the first character recognition unit 530 and the second character recognition unit 540 (0 ≦ m <2).

  Next, the first character recognition unit 530 and the second character recognition unit 540 perform a character recognition process (S802). The result of the character recognition processing by the first character recognition unit 530 and the second character recognition unit 540 is output to the recognition result information integration unit 550.

  Next, the recognition result information integration unit 550 generates the integrated recognition result information by combining the character recognition results input from the first character recognition unit 530 and the second character recognition unit 540 (S803). Specifically, the value included in the character recognition result is assigned to the array W [n] [m] described above. In the example of the integrated recognition result information illustrated in FIG. 7, W [0] [0] = 0.85, W [0] [1] = 0.8, W [1] [0] = 0.006, W [1] [1] = 0.075. The recognition result information integration unit 550 assigns values to the array W [n] [m] for each value of 0 ≦ n <N and 0 ≦ m <2.

  The character identification unit 560 performs character identification processing using the integrated recognition result information generated in S803 (S804). Thus, the operation of the character recognition device 500 ends.

  With the above configuration, in character recognition processing using a plurality of character recognition means, the character recognition learning device 100 that generates learning data using a combination of recognition results obtained by the plurality of character recognition means, and recognition by the plurality of character recognition means. A character recognition device 500 that performs character recognition processing using a combination of results is provided.

  In the above embodiment, a configuration in which two character recognition means (first character recognition unit 130, 530, second character recognition unit 140, 540) are provided is shown, but the present invention is not limited to this, Two or more character recognition means may be provided.

  In the above embodiment, the configuration is shown in which the likelihood of recognition for all character types is acquired by the two character recognition means. However, the likelihood for character types whose likelihood is smaller than a preset value is shown. May be replaced with 0, or may be arranged in descending order of likelihood, and the likelihood for character types below a predetermined order may be replaced with 0. In addition, it is not always necessary to use character recognition means for acquiring likelihoods for all character types. For example, a device that outputs character recognition results only for character types whose likelihood is equal to or greater than a predetermined value may be used. With these configurations, the amount of information to be handled is reduced, the processing load is reduced, and the processing speed is improved.

  Further, in the above-described embodiment, the configuration using the input receiving devices 200 and 600 for performing input processing by so-called online processing using an electronic pen or the like has been shown. However, for example, an input device by so-called offline processing such as a scanner is used. It is good also as a structure.

  The operations of the character recognition learning device 100 and the character recognition device 500 described here are realized by operating programs stored in the storage units 170 and 570 of the character recognition learning device 100 and the character recognition device 500, respectively. The This program may be provided by communication, or may be provided by being stored in a storage medium such as a CD-ROM that can be read by a computer.

  100 character recognition learning device, 110 control unit, 120 external I / F unit, 130 first character recognition unit, 140 second character recognition unit, 150 recognition result information integration unit, 160 learning data generation unit, 170 storage unit, 200 input Reception device, 500 character recognition device, 510 control unit, 520 external I / F unit, 530 first character recognition unit, 540 second character recognition unit, 550 recognition result information integration unit, 560 character identification unit, 570 storage unit, 600 Input reception device.

Claims (5)

Teacher information acquisition means for acquiring teacher information including each character type of a plurality of characters included in the learning document;
For each of the plurality of characters, the first recognition result information obtaining means for obtaining a first recognition result information indicating the result of the character recognition processing for each of a plurality of character types contained in the instruction information by the first character recognition means,
For each of the plurality of characters, the second recognition indicating the result of the character recognition processing by the second character recognition means for each of a plurality of character types contained in the instruction information using different processing schemes and the first character recognition means Second recognition result information acquisition means for acquiring result information;
For each of the plurality of characters , information indicating the character type of the character included in the teacher information, the first recognition result information for each of the plurality of character types included in the teacher information acquired for the character, and the first Learning data output means for executing machine learning using a combination of two recognition result information as input data, and outputting the execution result as learning data for character recognition;
A character recognition learning apparatus comprising: The character recognition learning device according to claim 1,
The first recognition result information is about each of the plurality of characters.
The character recognition result includes information indicating the character type recognized by the first character recognition means, and information indicating the likelihood,
The second recognition result information is about each of the plurality of characters.
The character recognition result includes information indicating the character type recognized by the second character recognition means, and information indicating the likelihood.
A character recognition learning device characterized by that. The character recognition learning device according to claim 1,
The first recognition result information is about each of the plurality of characters.
As the recognition result of the character, for each of a plurality of predetermined character type candidates, including information indicating the likelihood that the character type of the character is the character type candidate,
The second recognition result information is about each of the plurality of characters.
For each of a plurality of predetermined character type candidates, the character recognition result includes information indicating the likelihood that the character type of the character is the character type candidate.
A character recognition learning device characterized by that. The character recognition learning device according to claim 3,
The learning data output means includes
In the first recognition result information, a likelihood smaller than a predetermined value is replaced with 0,
In the second recognition result information, the likelihood smaller than a predetermined value is replaced with 0,
Performing the machine learning;
A character recognition learning device characterized by that. Computer
Instruction information acquiring means to acquire the teaching information containing the respective character types of a plurality of characters included in the learning writing,
Wherein for each of a plurality of characters, the first recognition result information acquiring means to acquire the first recognition result information indicating the result of the character recognition processing for each of a plurality of character types contained in the instruction information by the first character recognition means,
For each of the plurality of characters, the second recognition indicating the result of the character recognition processing for each of a plurality of character types contained in the instruction information by the second character recognition means using different processing schemes and the first character recognition means second recognition result information acquiring means to acquire the result information,
For each of the plurality of characters , information indicating the character type of the character included in the teacher information, the first recognition result information for each of the plurality of character types included in the teacher information acquired for the character, and the first second recognition results using a combination of information as input data to perform machine learning, learning data output means to output the execution result as a character recognition training data,
Program to function as.

Images