JP2017194806A - Character recognition device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition device, method and program capable of precise character recognition on a document including various hand-written characters.SOLUTION: The character recognition device comprises: a first image recognition unit 11 that extracts feature points from image data using a multilayer neural network, and outputs a plurality of text candidates and their likelihoods; a second image recognition unit 12 that extracts feature points from the image data using a multilayer neural network, and outputs text candidates separated into prime factor texts which are factors smaller than texts, and their likelihoods; a natural language processing unit 13 that couples and decouples neighboring prime factor texts with respect to the text candidates output by the second image recognition unit 12, applies morphological analysis to the plurality of patterns of combined texts, and outputs a plurality of text candidates that are likely in terms of natural languages, and their likelihoods; and a determination unit 14 that compares the output from the first image recognition unit 11 and the output from the natural language processing unit 13.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a character recognition apparatus, method, and program for character recognition of image data.

  2. Description of the Related Art A technique for generating digital data in which input information is converted into a predetermined character code by reading a handwritten document with an image scanner or the like and performing OCR (Optical Character Recognition) processing has become widespread.

  For example, according to Patent Document 1, a character identification system that performs character recognition by machine learning is disclosed. The character recognition system includes a character image input receiving unit that receives input of a sample character image, character component extraction that extracts a character component based on the sample character image, and pseudo character model generation that generates a pseudo character model based on the character component And an identification dictionary generation for generating an identification dictionary by generating a character identification pattern based on a pseudo-character model.

  Patent Document 2 discloses an information processing apparatus that performs morphological analysis after character recognition and determines likelihood using a feature vector that includes both part-of-speech likelihood and character similarity. . The information processing apparatus includes: a morpheme analysis unit that performs morpheme analysis on a character recognition result; and a P type that is created based on a part of speech quality of a word to which a target character that is a morpheme analysis result by the morpheme analysis unit belongs A feature quantity vector creating means for creating a feature quantity vector of P + 1 elements composed of the part of speech likelihood of each part of speech and the character similarity of each character of the character recognition result, and a feature quantity vector creating means. The certainty factor calculating means for calculating the certainty factor of each character of the character recognition result from the feature amount vector.

Japanese Patent Application Laid-Open No. 2015-069256 JP 2014-120059 A

  According to Patent Documents 1 and 2 described above, it is difficult to recognize various handwritten characters (for example, handwritten characters that are handwritten and thin handwritten characters) with high accuracy. There is a desire to do.

  An object of the present invention is to provide a character recognition device, method, and program capable of performing character recognition with high accuracy on a document containing various handwritten characters.

  In order to achieve the above object, a character recognition device according to an aspect of the present invention extracts a feature point from image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate separated into a prime factor text that is a factor smaller than the text and a likelihood; and For the text candidates output by the second image recognizing unit, adjacent prime factor texts are joined and separated, morpheme analysis is performed on the texts of a plurality of patterns to be combined, and a plurality that is likely from a natural language viewpoint. A natural language processing unit for outputting a text candidate and likelihood, a text candidate output by the first image recognition unit, and the natural candidate And a determining section for comparing the text candidate output by the word processor, the determining unit outputs the text of a predetermined or more likelihoods.

  In the character recognition device according to the aspect of the present invention, the most likely text candidate output by the first image recognition unit and the most likely text candidate output by the natural language processing unit In contrast, if a likelihood greater than or equal to a predetermined value is not obtained, another candidate may be compared in order of the likelihood until a likelihood greater than or equal to a predetermined value is obtained.

  In the character recognition device according to an aspect of the present invention, the determination unit outputs a text candidate having a high likelihood among the text candidates output by the natural language processing unit, by the first image recognition unit. It may be configured to handle with priority over text candidates.

  In the character recognition device according to an aspect of the present invention, the determination unit determines that the text has a likelihood greater than or equal to a predetermined value when the difference between the two text candidates does not exceed a predetermined threshold as a result of the comparison. May be configured to output.

  In the character recognition device according to an aspect of the present invention, the character recognition device further includes a setting unit that sets a time required for the processing, and the determination unit includes the most likely text candidate output by the first image recognition unit, When the most likely text candidate output by the natural language processing unit is compared and a likelihood greater than or equal to a predetermined value is not obtained, a likelihood greater than or equal to a predetermined value is obtained within the time set by the setting unit. Until it is determined, another candidate may be compared in descending order of likelihood.

  In the character recognition device according to an aspect of the present invention, the determination unit may be configured to output text in which the likelihood that the predetermined or higher likelihood is not obtained is converted to a buff.

  In the character recognition device according to an aspect of the present invention, the first image recognition unit has a multilayer neural network suitable for each item by machine learning, searches for an item included in the image data, A feature point is extracted from the image data using a multi-layer neural network suitable for the item, a plurality of text candidates and likelihoods are output, and the second image recognition unit is a multi-layer neural network suitable for each item. Is a factor that is smaller than the text by searching for an item included in the image data, extracting a feature point from the image data using a multilayer neural network suitable for the item A configuration may be adopted in which text candidates and likelihoods separated into prime factor texts are output.

  To achieve the above object, a character recognition method according to an aspect of the present invention includes a first image recognition step of extracting feature points from image data using a multilayer neural network and outputting a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate and likelihood separated into a prime factor text that is a factor smaller than the text; and For the text candidates output by the second image recognition process, adjacent prime factor texts are joined and separated, morpheme analysis is performed on a plurality of patterns of text to be combined, and a plurality that is likely from a natural language viewpoint. A natural language processing step for outputting a text candidate and likelihood, and a text candidate output by the first image recognition step; And a determination step of comparing the text candidate outputted by the natural language processing steps, said determining step outputs a text of a predetermined or more likelihoods.

  To achieve the above object, a character recognition program according to one aspect of the present invention extracts a feature point from image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate and likelihood separated into a prime factor text that is a factor smaller than the text; and For the text candidates output by the second image recognition process, adjacent prime factor texts are joined and separated, morpheme analysis is performed on a plurality of patterns of text to be combined, and a plurality that is likely from a natural language viewpoint. A natural language processing step for outputting a text candidate and likelihood and a text output by the first image recognition step A character recognition program for realizing, by a computer, a determination step for comparing a complement and a text candidate output by the natural language processing step, wherein the determination step outputs a text having a predetermined likelihood or more It is a character recognition program.

  According to the present invention, character recognition can be performed with high accuracy.

It is a block diagram which shows the structure of a character recognition apparatus. It is a figure which shows typically the mode of the feature extraction and vector conversion by an image recognition part. It is a figure which shows typically the mode of determination of the character type by an image recognition part. It is a figure where it uses for description about the image recognition by a 1st image recognition part and a 2nd image recognition part. It is a flowchart with which it uses for description about the flow of operation | movement of a character recognition apparatus.

  Hereinafter, a character recognition device, method, and program according to embodiments of the present invention will be described with reference to the drawings. In all the drawings for explaining the embodiments, common constituent elements are denoted by the same reference numerals, and repeated explanation is omitted.

  In the following, as an example, the configuration and operation of a character recognition apparatus that recognizes characters in imaged image data by imaging a document such as a form or a questionnaire sheet containing handwritten characters with a scanner or the like will be described. Even if a document does not contain handwritten characters, that is, a document in which characters are printed using a typeface data, if the printed characters are blurred or blurred, the characters The recognition rate decreases. The character recognition apparatus according to the present embodiment may be applied to the recognition of characters of image data obtained by imaging such a document with a scanner or the like.

  Conceptually, the character recognition device 1 uses an algorithm for learning a multilayer neural network including an input layer, one or a plurality of intermediate layers, and an output layer by backpropagation (error back propagation method). Character recognition.

  Specifically, as shown in FIG. 1, the character recognition device 1 is generated by an input unit 10 that inputs image data, an image recognition unit 11 that recognizes image data and generates text, and an image recognition unit 11. A natural language processing unit 12 that performs natural language processing on the text that has been processed, and a determination unit 13 that compares the text generated by the image recognition unit 11 with the text subjected to natural language processing by the natural language processing unit 12. The character recognition device 1 also includes an output unit 16 that outputs text output from the determination unit 13. The output unit 16 may divide the text into items and output them as csv data. In addition, the character recognition device 1 causes the image recognition unit 11, the natural language processing unit 12, and the determination unit 13 to function independently, and each output result affects the output result. The text of is output.

  The input unit 10 is constituted by, for example, a scanner device, generates an image data by imaging a document, and inputs the generated image data to the storage unit 14.

  Here, the operation of the image recognition unit 11 will be described. The image recognition unit 11 performs document structure analysis such as ruled line extraction, frame structure analysis, and position estimation of the reading target frame based on the image data. Next, the image recognition unit 11 receives a result of the document structure analysis and extracts a character line to be read. Next, the image recognition unit 11 cuts out character pattern candidates from the character line image and performs character identification of each character pattern.

Next, a character identification procedure will be described. As shown in FIG. 2, the image recognition unit 11 performs feature extraction on the image data of one extracted character pattern. Then, the image recognition unit 11 extracts the direction component of the character stroke and converts the image data into one vector. In the example illustrated in FIG. 2, the image data X is input to a multi-layer neural network, and a state in which features are extracted by capturing features such as directions and positions is schematically illustrated. Further, in the example illustrated in FIG. 2, a state in which the vectors are converted into the vector X ₁ , the vector X _2, and the vector X ₃ is schematically illustrated.

  As shown in FIG. 3, the image recognition unit 11 determines what the character type is based on the converted vector. In the determination, the image recognition unit 11 refers to the dictionary data holding which character type is distributed to which side in the feature space from the state of distribution using a large number of patterns in advance, and is unknown Image data candidates with the input pattern are determined. In the example shown in FIG. 3, the dictionary data conceptually shows the information of the character type “medium”, the character type “Sen”, and the character type “10”.

  The image recognizing unit 11 acquires a plurality of text candidates (for example, Naka, Shin, and Ten) and the likelihood of each text candidate by the above process. The likelihood of each text candidate can be calculated by the distance between the center of each candidate in the feature space and image data that is an unknown input pattern.

  As shown in FIG. 1, the image recognizing unit 11 includes a first image recognizing unit 11a and a second image recognizing unit 11b.

  The first image recognition unit 11a reads image data from the storage unit 15, extracts feature points from the image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods.

  The second image recognition unit 11b reads the image data from the storage unit 15, extracts feature points from the image data using a multilayer neural network, and separates the text candidates into prime factor texts that are factors smaller than the text. And likelihood.

  The main difference between the first image recognition unit 11a and the second image recognition unit 11b is a difference in learning data used when performing machine learning for image recognition. The first image recognition unit 11a prepares learning data so that one character is output as one character, whereas the second image recognition unit 11b separates one character into a prime factor text that is a smaller factor. Learning data is prepared for output.

  Here, specific examples of image recognition by the first image recognition unit 11a and image recognition by the second image recognition unit 11b will be described. Hereinafter, a case where image recognition is performed on the first image recognition unit 11a and the second image recognition unit 11b, for example, image data A1 as illustrated in FIG. 4A will be described.

  The first image recognition unit 11a performs processing for dividing the image data A1. In this embodiment, it is assumed that the image data A1 can be divided into four pieces of image data a1, a2, a3, and a4 as shown in FIG.

  The first image recognition unit 11a extracts feature points from the image data a1 using a multilayer neural network, and generates a plurality of candidates (for example, “high”, “，”, “wealth”, “livestock”, etc.) Then, the likelihood of each candidate is calculated.

  The first image recognition unit 11a extracts feature points from the image data a2 using a multilayer neural network, and generates a plurality of candidates (for example, “school”, “core”, “梓”, “inspection”, etc.) Then, the likelihood of each candidate is calculated.

  The first image recognition unit 11a extracts feature points from the image data a3 using a multilayer neural network, and generates a plurality of candidates (for example, “hour”, “，”, “wait”, “晤”, etc.) Then, the likelihood of each candidate is calculated.

  The first image recognition unit 11a uses a multilayer neural network to extract feature points from the image data a4, and generates a plurality of candidates (for example, “price”, “addition”, “what”, “finish”, etc.) Then, the likelihood of each candidate is calculated.

  Based on the likelihood of each candidate, the first image recognizing unit 11a outputs, for example, “high school age”, “喬喬 暁” etc. to the determination unit 13 as text candidates.

  The second image recognition unit 11b performs a process of dividing the image data A1 so as to generate a prime factor text that is a smaller unit than the first image recognition unit 11a. Prime factor text is text that represents a single character with a smaller factor. For example, “tree” and “intersection” represented by small factors from the image data “school” are prime factor texts. In this embodiment, it is assumed that the image data A1 can be divided into six pieces of image data b1, b2, b3, b4, b5, and b6 by the dividing process as shown in FIG.

  The second image recognition unit 11b extracts feature points from the image data b1 using a multi-layer neural network, and a plurality of prime factor text candidates (for example, “high”, “喬”, “rich”, “livestock”). Etc.) and the likelihood of each candidate is calculated.

  The second image recognition unit 11b extracts feature points from the image data b2 using a multilayer neural network, and a plurality of prime factor text candidates (for example, “tree”, “not”, “six”, “禾”). Etc.) and the likelihood of each candidate is calculated.

  The second image recognition unit 11b extracts feature points from the image data b3 using a multi-layer neural network, and a plurality of prime factor text candidates (for example, “intersection”, “fix”, “ki”, “fill”). Etc.) and the likelihood of each candidate is calculated.

  The second image recognition unit 11b extracts feature points from the image data b4 using a multilayer neural network, and generates a plurality of prime factor text candidates (for example, “day”, “，”, “month”, etc.). Then, the likelihood of each candidate is calculated.

  The second image recognition unit 11b extracts feature points from the image data b5 using a multi-layer neural network, and a plurality of prime factor text candidates (for example, “Tera”, “圭”, “茉”, “bitter”). Etc.) and the likelihood of each candidate is calculated.

  The second image recognition unit 11b extracts feature points from the image data b6 using a multi-layer neural network, and a plurality of prime factor text candidates (for example, “price”, “addition”, “what”, “finish”). Etc.) and the likelihood of each candidate is calculated.

  Based on the likelihood of each prime factor text, the second image recognition unit 11b outputs, for example, “Takagi Kyohichijiro”, “喬喬 不定 曰 圭” etc. to the determination unit 13 as text candidates.

  The natural language processing unit 13 joins and separates adjacent prime factor texts from the text candidates output by the second image recognition unit 11b, performs morpheme analysis on a plurality of patterns of combined text, Output multiple text candidates and likelihoods that are likely from a linguistic point of view.

  Specifically, the natural language processing unit 13 separates “Takagi Kyo Nichitera” into “Takagi” and “Kohichi Terayo”, or separates it into “Taka”, “Kiko”, and “Hijitera”. Or, morphological analysis is performed on each of them, and each likelihood is calculated.

  Further, the natural language processing unit 13 performs morphological analysis on “Takagi” and “Kojitsujiro” by joining “Sun” and “Tera” to “Time” and “Takagi” and “Kyodai”. And calculate the likelihood.

  In addition, the natural language processing unit 13 joins “tree” and “crossing” into “school” for “high”, “kikko”, “hidera dai”, and combines “high school” with the previous text “high”. ”,“ High school ”,“ Hijirayo ”, perform morphological analysis, and calculate the likelihood.

  The natural language processing unit 13 joins “day” and “temple” to “time” and “high school” and “era” for “high school” and “Hijirayo”, and performs morphological analysis. Calculate the likelihood.

  In this way, the natural language processing unit 13 performs processing for joining and separating the text candidates output from the second image recognition unit 11b with all patterns. For example, a morphological analysis of the text “My Takagi Kyohijidaiha” would result in “subject”, “case particle”, “noun (surname)”, “noun”, “noun”, and “coordinate particle”. Since it is an unreasonable arrangement from the viewpoint, the text is evaluated to have a low likelihood. On the other hand, when the morphological analysis is performed on the text “My high school days”, it becomes “subject”, “case particle”, “noun”, and “participant particle”. On the other hand, it is evaluated that the likelihood is high. An error due to an image recognition error can be reduced by performing evaluation by morphological analysis on a text candidate separated into prime factor texts and combined in all patterns.

  The determination unit 14 includes a text candidate output by the first image recognition unit 11a (for example, “high school age”), a text candidate output by the natural language processing unit 13 (for example, “Takagi Koji Neraji”), Contrast.

  Based on the comparison result, the determination unit 14 outputs a text with a predetermined likelihood or higher (for example, “high school age”).

  By comparing the text from two different viewpoints, the first image recognition unit 11a that emphasizes image recognition and the second image recognition unit 11b that emphasizes natural language, the character recognition device 1 performs character recognition with higher accuracy. It can be carried out.

  In addition, the determination unit 14 compares the text candidate with the highest likelihood output by the first image recognition unit 11a with the text candidate with the highest likelihood output by the natural language processing unit 13, and exceeds a predetermined value. When the likelihood (for example, the coincidence rate is 98% or more) is not obtained, the other candidates may be compared with each other in the descending order of likelihood until a predetermined or more likelihood is obtained.

  The determination unit 14 feeds back to the image recognition unit 11 the weight of which answer is likely from all the output patterns obtained by the first image recognition unit 11a and the natural language processing unit 13. Further, the determination unit 14 feeds back “possibility of letter X” that does not exist in the output pattern. The image recognition unit 11 receives the feedback and performs image recognition based on weighting again.

  For example, the determination unit 14 includes the “highest likelihood text candidate” output by the first image recognition unit 11 a and the highest likelihood text candidate output by the natural language processing unit 13. If the difference is large and the likelihood greater than or equal to the predetermined value is not obtained, “high school” which is the next most likely text candidate output by the first image recognition unit 11a is compared. The “high school age” is output when the likelihood is more than a predetermined level by comparing the “era” with the “high school age” that is the most likely text candidate output by the natural language processing unit 13.

  Therefore, since the character recognition device 1 repeats the process for a document containing various handwritten characters until a likelihood equal to or higher than a predetermined value is obtained, highly reliable and highly accurate character recognition can be performed.

  In addition, the determination unit 14 may have a configuration in which a text candidate having a high likelihood among text candidates output by the natural language processing unit 13 is preferentially handled over a text candidate output by the first image recognition unit 11a. .

  In the case of this configuration, the character recognition device 1 gives priority to the processing by the natural language processing unit 13, and thus is suitable for natural language processing, and exhibits strength in character recognition of sentences.

  In addition, the determination unit 14 may be configured to determine that the text has a likelihood equal to or higher than a predetermined value and output it when the difference between the two text candidates does not exceed a predetermined threshold value as a result of the comparison. That is, when the text candidate output by the natural language processing unit 13 and the text candidate output by the first image recognition unit 11a substantially coincide with each other, the determination unit 14 uses a text having a likelihood greater than or equal to a predetermined value. Judge that there is.

  In the case of this configuration, the character recognition device 1 performs character recognition using the results of the two processes of the natural language processing unit 13 and the first image recognition unit 11a, and thus performs character recognition with high accuracy. be able to.

  As shown in FIG. 1, the character recognition device 1 may include a setting unit 15 that sets a time required for processing. For example, the setting unit 15 sets the processing time to 10 minutes or the like in accordance with a user instruction.

  The determination unit 14 compares the text candidate having the highest likelihood output from the first image recognition unit 11a with the text candidate having the highest likelihood output from the natural language processing unit 13, and has a likelihood greater than or equal to a predetermined value. In the case where the values are not obtained, the other candidates are compared in descending order of likelihood until a likelihood equal to or higher than a predetermined value is obtained within the time set by the setting unit 15.

  Therefore, since the character recognition device 1 repeats the process until a predetermined likelihood or more is obtained within the set time, it is difficult to recognize that no more than a predetermined likelihood is obtained no matter how many times the recognition process is performed. There is a merit that the trial is not repeated many times for the image data.

  Moreover, the determination part 14 may be the structure which outputs the text in which the likelihood more than predetermined was not obtained as a baffle.

  A prone character indicates a portion where character recognition could not be performed, such as “◯” and “Δ”. Specifically, when a likelihood greater than or equal to a predetermined value is not obtained for the “time” portion of “high school age”, the determination unit 14 outputs “high school ○ fee”.

  Therefore, the character recognition device 1 can clearly output the location where the character was identified and the location where the character could not be identified. It should be noted that the correct character (in this embodiment, the portion of “◯” is “hour”) may be fed back to the character recognition device 1 for the portion that has been turned upside down. With this feedback, the character recognition device 1 can output the image data that was previously “◯” as the correct text “school” in the next character recognition.

  The second image recognition unit 11b has a multilayer neural network suitable for each item by machine learning, searches for an item included in the image data, and uses the multilayer neural network suitable for the item. The configuration may be such that feature points are extracted from the image data, and text candidates separated into prime factor texts, which are factors smaller than the text, are output.

  For example, information relating to an application date such as “May 1, 2015” can be expected to be input to an item such as “application date”. That is, for items such as “application date”, the numbers “0-9” and the kanji characters “year”, “month”, “day” are input, and no other characters are input. Therefore, when the recognized item is “application date”, the second image recognition unit 11b outputs the numbers “0-9” and the kanji characters “year”, “month”, “day”. A feature point is extracted from image data using a simple neural network, and text candidates separated into prime factor texts, which are factors smaller than the text, are output.

  In addition, it can be expected that information related to a katakana of a name such as “Tokitataro” is input to an item such as “name_phonetic”. That is, for an item such as “name_reading”, katakana “A-N” or the like is input, and other characters (kanji, numbers, etc.) are not input. Therefore, the second image recognition unit 11b extracts feature points from the image data using a neural network that outputs katakana when the item to be recognized is “name_reading”, and is smaller than the text. Output text candidates separated into prime factor texts.

  In addition, it is expected that information related to a telephone number such as “03-3581-1111” is input to an item such as “telephone number (TEL)”. That is, for items such as “telephone number (TEL)”, numbers “0-9” and hyphen “-” are input, and other characters (kanji, hiragana, etc.) are not input. Therefore, when the item to be recognized is “telephone number (TEL)”, the second image recognition unit 11b uses a neural network that outputs numbers “0-9” and a hyphen “-”. Feature points are extracted from the data, and text candidates separated into prime factor texts, which are factors smaller than the text, are output.

  Therefore, since the character recognition apparatus 1 outputs a text candidate from image data using the multilayer neural network suitable for every item, it can perform character recognition efficiently and with high precision.

  Here, the operation flow of the character recognition device 1 will be described with reference to the flowchart shown in FIG.

  In step S1, the first image recognition unit 11a extracts feature points from the image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods.

  In step S2, the second image recognition 11b extracts feature points from the image data using a multilayer neural network, and outputs text candidates separated into prime factor text that is a factor smaller than the text.

  In step S3, the natural language processing unit 12 performs morphological analysis on the texts of a plurality of patterns to be combined by joining and separating the adjacent prime factor texts with respect to the text candidates output in the process of step S2. A plurality of text candidates and likelihoods that are likely from a natural language perspective are output.

  In step S4, the determination unit 13 compares the text candidate output in the step S1 with the text candidate output in the step S3. In this step, the determination unit 13 outputs a text with a predetermined or higher likelihood (for example, a matching rate of 98% or higher).

  Therefore, the character recognition device 1 can perform character recognition with high accuracy for a document containing various handwritten characters.

  In the present embodiment, the configuration and operation of the character recognition device 1 capable of performing character recognition with high accuracy for documents mainly including various handwritten characters have been described. However, the present invention is not limited thereto. A document including each component and including various handwritten characters may be configured as a method and a program for performing character recognition with high accuracy.

  Further, it is realized by recording a program for realizing each function constituting the character recognition device 1 on a computer-readable recording medium, causing the computer system to read the program recorded on the recording medium, and instructing it. May be.

  Specifically, the program extracts a feature point from image data using a multi-layer neural network, outputs a plurality of text candidates and likelihoods, and uses a multi-layer neural network to generate an image. A feature point is extracted from the data, a text candidate separated into a prime factor text that is a factor smaller than the text, a second image recognition step for outputting the likelihood, and a text candidate output by the second image recognition step A natural language processing step that combines and separates adjacent prime factor texts, performs morphological analysis on multiple patterns of combined text, and outputs multiple text candidates and likelihoods that are likely from a natural language perspective And the text candidate output by the first image recognition step and the text candidate output by the natural language processing step. A program for realizing the Hisuru determining step, the computer. In the determination step, text having a likelihood equal to or higher than a predetermined value is output.

  Furthermore, the “computer system” here includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a portable medium such as a flexible disk, a magneto-optical disk, a ROM, and a CD-ROM, and a hard disk built in the computer system.

  Furthermore, “computer-readable recording medium” means that a program is dynamically held for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It is also possible to include one that holds a program for a certain time, such as a volatile memory inside a computer system that becomes a server or client in that case. Further, the program may be for realizing a part of the above-described functions, and may be capable of realizing the above-described functions in combination with a program already recorded in the computer system. .

  DESCRIPTION OF SYMBOLS 1 Character recognition apparatus, 10 input part, 11 image recognition part, 11a 1st image recognition part, 11b 2nd image recognition part, 12 natural language processing part, 13 judgment part, 14 memory | storage part, 15 setting part, 16 output part

In order to achieve the above object, a character recognition device according to an aspect of the present invention extracts a feature point from image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate separated into a prime factor text that is a factor smaller than the text and a likelihood; and Adjacent prime factor texts are joined and separated from the text candidates output by the second image recognition unit, and morpheme analysis is performed on each of the combined patterns of text, which is plausible from a natural language perspective. A natural language processing unit that outputs a plurality of text candidates and likelihoods; a text candidate that is output by the first image recognition unit; Serial and a determining section for comparing the text candidate output by the natural language processing unit, wherein the determination unit outputs the text of a predetermined or more likelihoods.

To achieve the above object, a character recognition method according to an aspect of the present invention includes a first image recognition step of extracting feature points from image data using a multilayer neural network and outputting a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate and likelihood separated into a prime factor text that is a factor smaller than the text; and Adjacent prime factor texts are joined to and separated from the text candidates output in the second image recognition process, and morpheme analysis is performed on each of the combined patterns of text, which is plausible from a natural language perspective. A natural language processing step for outputting a plurality of text candidates and likelihoods, and a text output by the first image recognition step. Comprising a candidate, and a judgment step of comparing the output text candidates by the natural language processing steps, said determining step outputs a text of a predetermined or more likelihoods.

To achieve the above object, a character recognition program according to one aspect of the present invention extracts a feature point from image data using a multilayer neural network, and outputs a plurality of text candidates and likelihoods. Extracting a feature point from the image data using a multi-layer neural network, and outputting a text candidate and likelihood separated into a prime factor text that is a factor smaller than the text; and Adjacent prime factor texts are joined to and separated from the text candidates output in the second image recognition process, and morpheme analysis is performed on each of the combined patterns of text, which is plausible from a natural language perspective. A natural language processing step for outputting a plurality of text candidates and likelihoods, and the first image recognition step. A character recognition program for realizing, by a computer, a determination step for comparing a text candidate and a text candidate output by the natural language processing step, wherein the determination step includes text having a likelihood greater than or equal to a predetermined value. This is a character recognition program to output.

Claims (9)

A first image recognition unit for extracting feature points from image data using a multilayer neural network and outputting a plurality of text candidates and likelihoods;
A second image recognition unit that extracts feature points from the image data using a multi-layer neural network, and outputs a text candidate and likelihood separated into a prime factor text that is a factor smaller than the text;
For the text candidates output by the second image recognition unit, adjacent prime factor texts are joined and separated, morpheme analysis is performed on a plurality of patterns of text to be combined, and it is likely from a natural language viewpoint. A natural language processing unit that outputs a plurality of text candidates and likelihoods;
A determination unit that compares the text candidate output by the first image recognition unit with the text candidate output by the natural language processing unit;
The determination unit is a character recognition device that outputs text having a likelihood equal to or higher than a predetermined value.   The determination unit compares the text candidate with the highest likelihood output by the first image recognition unit with the text candidate with the highest likelihood output by the natural language processing unit, and has a likelihood greater than or equal to a predetermined value. The character recognition device according to claim 1, wherein if the candidate is not obtained, the other candidates are compared in descending order of likelihood until a likelihood equal to or greater than a predetermined value is obtained.   The determination unit treats a text candidate having a high likelihood among text candidates output by the natural language processing unit with priority over a text candidate output by the first image recognition unit. The character recognition device described.   4. The determination unit according to claim 1, wherein if the difference between the two text candidates does not exceed a predetermined threshold as a result of the comparison, the determination unit determines that the text has a likelihood greater than or equal to a predetermined level and outputs the text. The character recognition device described in 1. It has a setting part that sets the time required for processing,
The determination unit compares the text candidate with the highest likelihood output by the first image recognition unit with the text candidate with the highest likelihood output by the natural language processing unit, and has a likelihood greater than or equal to a predetermined value. 5. If any of the candidates is not obtained, the other candidates are compared in descending order of likelihood until a likelihood greater than or equal to a predetermined value is obtained within the time set by the setting unit. The character recognition device according to item.   6. The character recognition device according to claim 1, wherein the determination unit outputs a text for which a likelihood greater than or equal to a predetermined value is not obtained as a illegitimate character. The first image recognition unit
It has a multilayer neural network suitable for each item by machine learning,
Search for items included in the image data, extract feature points from the image data using a multilayer neural network suitable for the items, and output a plurality of text candidates and likelihoods,
The second image recognition unit
It has a multilayer neural network suitable for each item by machine learning,
Search for items included in the image data, extract feature points from the image data using a multilayer neural network suitable for the items, and separate text candidates into prime factor texts that are factors smaller than the text The character recognition device according to claim 1, wherein the likelihood is output. A first image recognition step of extracting feature points from image data using a multilayer neural network and outputting a plurality of text candidates and likelihoods;
A second image recognition step of extracting feature points from the image data using a multi-layer neural network, and outputting text candidates and likelihoods separated into prime factor text that is a factor smaller than the text;
For the text candidates output by the second image recognition process, adjacent prime factor texts are joined and separated, morpheme analysis is performed on a plurality of patterns of text to be combined, and it is plausible from a natural language viewpoint. A natural language processing step for outputting a plurality of text candidates and likelihood;
A determination step of comparing the text candidate output by the first image recognition step with the text candidate output by the natural language processing step;
The determination step is a character recognition method for outputting text having a likelihood equal to or greater than a predetermined value. A first image recognition step of extracting feature points from image data using a multilayer neural network and outputting a plurality of text candidates and likelihoods;
A second image recognition step of extracting feature points from the image data using a multi-layer neural network, and outputting text candidates and likelihoods separated into prime factor text that is a factor smaller than the text;
For the text candidates output by the second image recognition process, adjacent prime factor texts are joined and separated, morpheme analysis is performed on a plurality of patterns of text to be combined, and it is plausible from a natural language viewpoint. A natural language processing step for outputting a plurality of text candidates and likelihood;
A character recognition program for realizing, by a computer, a determination step of comparing the text candidates output by the first image recognition step with the text candidates output by the natural language processing step,
The determination step is a character recognition program that outputs text having a likelihood equal to or higher than a predetermined value.

Images