JP2015069256A - Character identification system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an addition learning function using a few samples, in an automatic recognition device and an automatic recognition service.SOLUTION: A character identification system includes: a character image input receiving unit which receives input of a sample character image; a character component extraction unit which extracts a character component, on the basis of the sample character image; a pseudo-character model generation unit which generates a pseudo-character model, on the basis of the character component; and an identification dictionary generation unit which generates a character identification pattern, on the basis of the pseudo-character model, to generate an identification dictionary.

Description

  The present invention relates to a system and related services for recognizing characters in documents such as handwritten characters and printed characters, and images such as videos and photographs.

  With increasing social interest in improving the efficiency of information utilization, useful information can be searched and organized from a large number of electronic documents stored on a server or personal computer (PC) owned by an individual with high accuracy. There is a need for information utilization technology to do this. The data stored in the corporate information system increases from 50% to 60% annually, while the structured data stored in the database is about 20% and the remaining 80% is unstructured such as document images / photos / videos. Data. A large amount of unstructured data stored in the company or organization is originally an information asset for the organization, but it is difficult to say that it is fully utilized at present.

  In order to utilize information, it is necessary to code character strings included in paper documents, document images, photographs, and videos. The problem at this time is the handling of external characters used for variants of first and last names and place names.

  For example, according to related laws and notices such as the Japanese Family Register Act, there are about 50,000 character types that can be used for family register. This is much more character than character code systems (first level, second level) such as JIS (JIS X 0208) and Shift JIS, which are generally used. So far, each company and local government system has built its own external character handling function. Even in UNICODE, which covers character code systems around the world, there are many differences from character sets and shapes used for family register, so they are not necessarily usable. In recent years, character systems such as character information infrastructure, family register unified characters, and basic resident register network unified characters have been studied as an attempt to improve kanji such as personal name kanji used in such administrations. It is designed to handle approximately 20,000 to 60,000 character types, although the character types that can be handled differ depending on the character system. The number will increase further if foreign character types are included.

  There is an OCR device as a device for encoding a character string included in a paper document, document image, photograph, or video. The general functions of the OCR device and the form input form using the same are outlined in Patent Document 1, Patent Document 2, and Patent Document 3. Patent Document 1 describes a basic processing flow in the OCR apparatus. When reading a form automatically, extract the character code, character line, ruled line, frame, etc. written in the form, read a specific area on the form that requires data input, and externalize it as a text file. Output to storage device. Patent Document 2 describes a method of applying morphological analysis to the recognition result of OCR as means for improving the reading accuracy of OCR. Patent Document 3 proposes a character string notation analysis process using ascending syntax analysis for a handwritten digit string. In both cases, a technique for improving the accuracy of reading data on a paper document or a document image using an OCR apparatus is proposed.

  That is, the background of the present invention is that a character recognition function in the OCR device and a character code system for strictly handling the character form of family names are being developed.

JP 06-52156 JP 05-108991 JP 2002-117374 A

  Even in an OCR apparatus that discriminates and reads a character pattern in a paper document, an image, or a video, careful design has been required for handling external characters such as first and last names. When registering an external character in an OCR apparatus, a special character code area called a user-defined area is generally provided, and reading is performed by registering a sample image of the external character. However, the problem is that the recognition accuracy is not sufficient with a few sample images.

The above-described problems include, for example, a character image input receiving unit that receives an input of a sample character image, a character component extraction that extracts a character component based on the sample character image, and a pseudo character that generates a pseudo character model based on the character component The invention is solved by a character identification system comprising: a model generation unit; and an identification dictionary generation for generating an identification dictionary by generating a character identification pattern based on a pseudo-character model.

  According to an embodiment of the present invention, an external character for registering a simple sample image by learning at least one handwritten character pattern or printed character pattern corresponding to at least one prepared external character or new character image. It becomes possible to recognize with higher accuracy than the recognition method.

1 is an overall view of a document recognition service constituting the present invention. It is the recognition dictionary production | generation apparatus which comprises this invention. It is a recognition apparatus which implement | achieves the recognition function of the automatic recognition apparatus of this invention. It is a structural example of the hardware in this invention. It is the input device part of the hardware in this invention. It is a recognition process diagram of a document in the present invention. It is a recognition process figure of the character in the present invention. It is the figure which showed the mechanism of the autonomous learning in this invention. It is the figure which showed the mechanism of the automatic differentiation in this invention. It is an example of the sample of the input document in this invention.

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

  First, the external appearance of the entire automatic recognition apparatus system to which the present invention is applied will be described. FIG. 1 shows a state in which a plurality of window input devices 0101 are connected to a recognition cloud or a host computer 0103 via a network 0102.

  FIG. 1 is a diagram showing a business flow using a document recognition service using a Web service, which is an application example of the present invention. This service has the feature that the character recognition dictionary is improved and the recognition accuracy is improved by continuing the business.

  First, the user of the document recognition service registers a document or character to be read (process 1, 0110). Document images and text images can also be performed using a computer (ie, display on a display and input using a mouse or keyboard). In this figure, an electronic pen and special paper are used, or a camera or scanner is used. An imaged document group (0101) is handled. The input business document is sent to the recognition cloud 0103 through the network 0102. Business documents collected from various regions and various authors are recorded in the database 0104 in the recognition cloud. Next, based on the document image data, a necessary part of the document is recognized by the computer (0105). At the time of recognition, a dictionary 0106 for character recognition is used. The recognition result is presented to the user of the recognition service (process 2, 0111).

  The recognition cloud has a section with three major functions. The first is the recognition unit 0105. The recognition unit recognizes individual character patterns based on the document image data. At this time, data recording the shape and distribution of the character pattern is used as a dictionary. The second is the analysis unit 0106. The analysis unit identifies a defective pattern or a new pattern to be registered among the recognition results presented to the user. The third is a learning unit. The learning unit changes the character recognition dictionary in order to read the pattern identified by the analysis unit. At this time, a dictionary necessary for learning is used. In recognition result presentation (process 2, 0111), the result of character recognition is presented to the user. If necessary, the user provides feedback (process 3, 0112) to the system according to the contents presented. Examples of feedback include recognition correction to input correct characters if the recognition result is incorrect, or addition of characters to add to the dictionary if the recognition target character is an external character that is not registered in the dictionary. There are requests. Based on these feedbacks, parameters to be corrected in the recognition dictionary are identified by the learning unit (0108), and the result is reflected in the recognition dictionary (0106) (process 4, 0113). In addition, in order to reflect this learning result, what kind of improvement proposal should be reflected in the dictionary is fed back to the recognition cloud. In this way, the learning result is fed back, the recognition dictionary is updated, and more accurate recognition can be performed.

  While the general OCR reading target character types are 3000 to 4000, the present invention presupposes recognition of tens of thousands of character types including external characters. Therefore, the correct answer rate of the first place is inevitably lower than that of a general OCR. Therefore, in the present invention, not only the first recognition rate but also the cumulative recognition rate is considered. The cumulative recognition rate is a recognition rate when a correct answer exists within the nth place (n varies depending on the application, for example, 15th place or 50th place). When the recognition result is not the first correct answer, in the recognition correction of the feedback process (0112) of the present invention, the user inputs a correct character by selecting from the candidate characters of the recognition result within the nth position. This work has a feature that the work efficiency is improved as compared with a case where a target character code is manually checked from a character code of tens of thousands of characters. In order to realize this feature, the present invention uses a dictionary that contributes to an improvement in the first-rank accuracy rate and a dictionary that contributes to an improvement in the cumulative recognition rate. Hereinafter, a method for creating each dictionary will be described.

  FIG. 2 is a recognition dictionary generation device corresponding to the learning unit 108 and the learning dictionary 109 of FIG. The storage device of the recognition dictionary generation device has a learning character pattern 0201 from which a feature vector is extracted (0202). When the recognition dictionary generation device and the recognition device are driven by the same hardware device, for example, it is assumed that a learning character pattern 0201 is stored in the external auxiliary storage unit (0411) of FIG. In general, the feature vector obtained here is learned and recorded in a dictionary, and a dictionary 1 (simple learning dictionary, 0204) for identifying a newly registered character pattern such as an external character is constructed. Dictionary 1 (0204) is a dictionary that contributes to improving the first-ranked correct answer rate, and is a basic dictionary for recognizing character patterns. Here, a feature vector created from a character pattern, a character code, its weight, and the like are stored. Since the characteristics of the character pattern used for learning are recorded, it is used to process simple identification algorithms such as the nearest neighbor method. However, if there are very many character types, such as variants of first and last names, or if new characters can be added in the character pattern shape, it may be difficult to prepare sufficient learning samples for all character types. Thus, it is difficult to obtain sufficient recognition accuracy with the dictionary 1 alone.

  The external character pattern A may have a character shape similar to the existing character pattern B. In the case of similar characters, it is difficult to discriminate between the character pattern A and the character pattern B, so that recognition accuracy can be improved by performing clustering. By performing the variant pattern clustering process (0205) in which persons who have similar character patterns are grouped together, it is possible to understand the ease of recognition of similar characters. As a result of clustering, a subset of similar characters is obtained as information. For example, if there is a character {A, B, C,...}, And it is determined by clustering that the characters O and Q are similar, a subset {O, Q} is obtained. This information is recorded as a search complement dictionary 2 (search complement dictionary, 0206). The dictionary 2 is a dictionary that contributes to improving the cumulative recognition rate. With this dictionary, when “O” is recognized, “Q” can be added to the recognition result candidate assuming misrecognition.

  In addition, a vertical or horizontal projection is taken with respect to the input character pattern image, a connected component of black pixels is obtained, and it is calculated whether there is another character pattern B that matches a part of the character pattern A. Through the above process, the character pattern parts are disassembled (0207), and the result is recorded as character pattern part disassembly data (0208). The parts disassembly pattern stores information such that the character pattern A is similar to another character pattern B that is partially similar to the character pattern A, and this is used as a dictionary 3 (search complement dictionary, 0209). Record. The dictionary 3 is a dictionary that contributes to improving the cumulative recognition rate. The dictionary 3 stores information on the deviation and habit of each character pattern. For example, “Japanese” has Nogihen and Kuchi placed on the left and right. In addition, as a storage format for the information of bias and 旁, there is another format in which “sum” is in the list of kanji characters including nogihen, and the position is on the left in the character. When this dictionary is used, interactive character code search can be performed in the feedback process (0112). For example, when the user has written a cross with a mouse or a digital pen, a character code candidate including that can be presented. Further, when the character “gold” is input, it is considered that the kanji biased general kanji is being searched, and a character code candidate including this can be presented.

  For the character pattern component, pseudo character model generation (0211) is further performed. The pseudo-character model generation is a function that generates a large amount of character patterns necessary for statistical learning based on one or a small number of sample samples for characters not registered in the dictionary or characters with a low recognition rate. First, in the handwritten character pattern part (0210), there are parts that make up the character pattern (flat knots such as Nogihen and Mongamae, and elemental kanji such as “day” and “mouth” that are often used as part of characters) Is stored in advance. In the pseudo pattern synthesis (0211), if necessary, the stroke order of characters is estimated from the character patterns to be learned, and the handwritten character pattern corresponding to the part is compared with the patterns that can be a group on the connected component or time series. It is checked from the dictionary (0210) whether a part exists. When a corresponding part is obtained, a large number of pseudo patterns can be synthesized in a combinational manner by synthesizing a part of a handwritten character pattern at that part. In addition, even if there is no corresponding handwritten character pattern component, the stroke order is estimated from the input character pattern to be learned. The pattern can be deformed. As a result, a non-linear deformation pattern that cannot be obtained by affine transformation or barrel transformation with respect to the whole is obtained, and variations in the resulting feature space are increased, enabling more accurate character recognition.

  A feature vector is extracted from the generated character pattern (2012), and high-order classifier learning (0213) is performed based on many patterns. For higher-order classifier learning, support vector machines, multilayer neural networks, parametric statistical discrimination methods, and the like can be used. By generating pseudo character patterns, characters with various patterns are created, so that higher-order discriminator learning that generally enables high-accuracy recognition is possible. The dictionary 4 (higher-order identification dictionary, 0214) is a dictionary that contributes to improvement of the first-order recognition rate generated by higher-order discriminator learning. In order to identify the added external character pattern, the dictionary 4 generates a pseudo character pattern based on one or a small number of sample samples, and stores the result of statistical learning based on this character pattern. . On the other hand, since the character patterns are synthesized in a pseudo manner, it is not always possible to realize the deformation of characters considered by humans. Therefore, instead of adding data to the character identification dictionary (0204), it is necessary to manage the character identification dictionary and the higher-order identification dictionary (0214) separately and use the respective recognition results.

  As described above, the automatic recognition apparatus having the additional learning function by the dictionary 4 generates a pseudo-generation model of a character pattern based on a small number of sample images presented as external characters, and learns this to generate a small number of samples. An automatic recognition device that can improve recognition accuracy even from a pattern can be realized.

  The recognition process of FIG. 3 is an example of the process of the recognition unit 105 of FIG. 1, and shows how recognition is performed using the dictionary generated by the previous recognition dictionary generation apparatus. The recognition process referred to here is one form of the recognition unit shown in the recognition unit 105 in FIG. First, when an unknown character pattern (0301) to be recognized is input, recognition is performed using dictionary 1 (simple learning dictionary, 0204) and dictionary 4 (higher-order identification dictionary, 0214), respectively (0302). For example, in the simple learning dictionary, the feature vector obtained from the learning character pattern 0201 is stored as it is, and the recognition candidate is calculated by calculating the nearest neighbor distance. On the other hand, the high-order classification dictionary stores a large amount of feature vectors extracted from handwritten character patterns automatically created through pseudo-character model generation, or supports secondary discriminant function parameters obtained by learning a large number of patterns. Vectors and the like are stored, and patterns that are difficult to discriminate with a simple learning dictionary are identified. In the generation of a pseudo model, learning can be performed from a small number of samples, and identification with higher accuracy is possible.

  In addition, since a plurality of identification dictionaries are used, recognition candidates can be obtained in multiples. For this, the recognition order is alternated, or the recognition likelihood (the center of each category in the feature space and the unknown pattern). The order of candidates can be made into a series by means such as rearranging in the order of (calculated by the distance to). In this way, by combining the two identification dictionaries, it is possible to increase the probability that the correct candidate will be ranked higher. Next, the character candidate is complemented using the recognition result obtained here. Completion of character candidates includes a method of adding characters that are easily misrecognized using the dictionary 2 (0206) as character candidates, or character candidates that are partially similar such as flats and wrinkles using the dictionary 3 (0209). There is a way to add. By adding this character candidate, the first recognition accuracy cannot be improved, but there is an advantage that the cumulative recognition rate can be improved. Through the above process, a highly accurate character recognition result (0304) is obtained.

  FIG. 4 shows that the document recognition system includes a form processing computer 0400. The document recognition system referred to here is one form of a recognition service / recognition system using a recognition cloud indicated by 0103 in FIG. The form processing computer 0400 includes an image input unit 0403, an input unit 0405, an output unit 0406, a communication unit 0407, a control unit 0408, an auxiliary storage unit 0409, a storage unit 0410, and an external auxiliary storage unit 0411. These are connected to each other via an internal bus 0412. A form image read by the scanner 0402 is input to the image input unit 0403. A form image may be input to the image input unit 0403 as electronic data 0404 via a network or the like without using the scanner 0402. The input unit 0405 receives input from the user. For example, the input unit 0405 is a keyboard and a mouse. The output unit 0406 outputs the result of the form processing. For example, the output unit 0406 is a display, a printer, or the like. A communication unit 0407 is an interface connected to the external network 0413. The result of the form processing may be output to the external server 0414 connected to the external network 0413. The control unit 0408 executes various processes related to the control of the form processing computer 0400, and is, for example, a CPU. The auxiliary storage unit 0409 is a storage unit other than the storage unit 0410 provided in the form processing computer 0400, and is an HDD, for example. The storage unit 0410 is a storage unit that can be directly accessed by the control unit 0408, and is, for example, a memory. The external auxiliary storage unit 0411 is a kind of auxiliary storage unit 0409, and is a storage unit provided outside the form processing computer 0400. For example, the external auxiliary storage unit 0411 is a CD-R, a DVD-R, or the like. Various programs including a program related to form processing (form processing program) are stored in the auxiliary storage unit 0409 or the external auxiliary storage unit 0411, and loaded into the storage unit 0410 when the control unit 0408 executes the various programs. The control unit 0408 executes the program loaded in the storage unit 0410. In addition, the control unit 0408 stores the form image input to the image input unit 0403 in the storage unit 0410, the auxiliary storage unit 0409, the external auxiliary storage unit 0411, and the like via the internal bus 0412. Note that the form processing computer 0400 may include at least the image input unit 0403, the control unit 0408, and the storage unit 0410, and may not include other units.

  FIG. 5 shows a partial apparatus configuration when an electronic pen is used when inputting electronic writing data 0404 as an input device in the document recognition system. The communication device 0503 is an interface that is connected to a network (not shown) and communicates with other devices (not shown) connected to the network. For example, the communication device 0503 receives data entered in a document 0501 such as an application form through the electronic pen device 0502 in the form of a wireless run or the like. By transmitting the received data, character pattern data having a stroke format can be transmitted. For example, the input in the feedback process 0112 of FIG. 1 is made by such a device.

  FIG. 6 shows a recognition processing flow of an application form recognition service which is an example of an application using the document recognition system shown in FIG. This corresponds to the details of the processing of the recognition unit 0105 in FIG.

  FIG. 10 shows an example of an application form used in the recognition process flow of FIG. Reference numeral 1001 is an application form (gift application form), and there are an entry field 1002 for a destination and a customer name 1003 as entry fields. When correctly converting a name having a large number of variants, the handwritten entry written in 1002 or 1003 is read, converted into a character code, and delivery is arranged based on this. At that time, if the converted character code is checked and it is wrong, or if multiple candidate character codes are output and candidate characters are selected, feedback from the user on the recognition result is obtained. Furthermore, it is possible to adjust the dictionary and parameters for character recognition based on this.

  Returning to FIG. 6, the outline of the processing flow of character string recognition in the application form recognition service will be described. In the character string recognition device according to the embodiment of the present invention, the OCR device captures a paper document and converts it into electronic image data. This process can be omitted when the original document is electronic image data (0601). Next, based on the electronic image data, document structure analysis such as ruled line extraction, frame structure analysis, and position estimation of the reading target frame is performed (0602). Next, in response to the result of the document structure analysis, a character line to be read is extracted (0603). Next, extraction of character pattern candidates from the character line image and character identification of each character pattern are performed (0604). The character extraction pattern and the identification result are collectively referred to as a character string hypothesis. In a document to be read, if a character notation string that can be written is determined in advance, a notation analysis is performed on the character string hypothesis (0605). As a result, the temporary character string including character extraction and character identification ambiguity is converted into character string text and output from the OCR as a read result text (0606), however, analysis with notation knowledge is sufficient. When the reliability of the read result text is low, such as when it could not be performed, the character string hypothesis is output. Both the reading result text and the reading hypothesis data hold position information on the document image in which the character string is written, if necessary. Through the above processing, the reading result text and the reading hypothesis data are output, and the document processing is generally performed based on these data. A dictionary obtained as a result of additionally learning character patterns such as external characters is reflected in the dictionary 0610. In this way, even when a sentence is used as input information, highly accurate character recognition using a small number of samples is possible.

  7 to 9 are diagrams illustrating an example of a specific learning method for realizing the learning unit 108 of FIG. However, the learning method for realizing FIG. 1 is not limited to this example.

  FIG. 7 shows the process of character recognition.

  When an image 0710 having one character pattern is input, feature extraction 0702 is performed. At this time, the direction component of the character stroke is extracted, and the character pattern image is converted into one vector. After the vector is obtained from the character pattern, it is determined what the character type is. This is referred to as category identification 0703. In category identification, it looks like a distribution using a large number of patterns in advance, so it stores in the dictionary which character type is distributed in which side of the feature space, and to which category the unknown input pattern belongs. To decide. This figure conceptually shows that information such as categories “8”, “5”, and “9” is stored. Although it is originally a high-dimensional vector, it is displayed in two dimensions to facilitate visualization. The character code 0704 is obtained by the above process.

Thus, for character recognition, it is important to know how a large amount of feature vectors obtained from a large amount of character patterns are distributed in the feature space. The variation of the feature vector on the feature space is caused by the deformation of the character pattern. For this reason, when a new external character is additionally registered, a pseudo-character generation model that generates a large number of character patterns from a small number of character patterns is important.
There are parameters to be learned in the pseudo model generation. One example of realizing learning of these parameters is automatic differentiation. The learning configuration using automatic differentiation is shown in FIGS.

The learning unit learns parameters necessary for character recognition using a mechanism such as multiple regression analysis, structural equation modeling, and automatic differentiation.
In parameter learning, a combination of automatic differentiation and regression analysis or structured modeling can be used.

FIG. 8 describes the relationship between the mechanism for learning parameters necessary for character recognition and automatic differentiation. When character recognition is performed, parameters are used to calculate the unknown pattern category by performing various vector operations, matrix operations, autocorrelation operations, convolution operations, etc. from the original signal (image) or feature vector. To do. Automatic differentiation is a basic numerical operation system that supports this calculation process.
In order to describe a more specific processing process, a mechanism for using an automatic differentiation mechanism and regression analysis together in learning will be described. Automatic differentiation defines "number" and "calculation" independently. An example of an arithmetic function used in automatic differentiation is shown in FIG. Here, it is assumed that n variables among all variables in the program are objects of partial differentiation. At this time, the structure of “number” is represented by the following vector.

Here, v is a place where a function value is held. Further, dk (k = 1 to n) is a place for holding a value when the function is partially differentiated by the kth variable. In the automatic differentiation, the number having the above structure is set as the AD number [Expression 1], and various calculations are performed based on this.
The reason why such a mechanism is introduced is to flexibly configure parameter adjustment in learning. As will be described later, an implicit parameter is used for rule calculation. For example, assuming that a secondary discriminant function is used as a high-order discrimination for character recognition, the parameters stored in the high-order discrimination dictionary are coefficients of a quadratic function expressing the distribution of each category. In the document recognition service of FIG. 1, when the recognition of the application form is processed, this higher-order identification dictionary is used as the recognition dictionary (0106). As a result, character recognition is performed, the result is presented to the user, and the feedback is obtained. As a result, it can be understood which character has made a mistake in recognition. In that case, in the learning process 0108, the dictionary is updated. If the character recognition is wrong, it can be interpreted that the likelihood that the originally correct discriminant function A is output is higher than the likelihood that the wrong discriminant function B outputs. Therefore, the mistaken discriminant function B is calculated. It is only necessary to finely modify the parameters used for the adjustment so that the likelihood decreases. In doing so, an automatic differentiation mechanism that can simultaneously process the calculation of the value and the calculation of the derivative value is useful. By the automatic differentiation mechanism, the fine adjustment amount of the parameter can be calculated from the value of the derivative value so that the calculated value (likelihood in this case) is lowered. As a result, a learning dictionary (0109) storing the finely adjusted parameters is created and reflected in the recognition dictionary (0106).

  The method of finely adjusting the parameters as described above is called a gradient method. The gradient method requires a partial derivative of the target function. In the document recognition service and application reception system, the target functions are the accuracy of character recognition and the accuracy of form processing. Here, a specific process of the gradient method will be described. Here, it is assumed that the teacher signal is given as feedback from the system user although the character recognition is correct or incorrect. For example, if a parameter effective for improving character identification accuracy is estimated for regression analysis, [Equation 2] is the target function.

This target function increases the “character recognition accuracy improvement”. Assume that the actual degree of improvement in character identification accuracy is represented by “improvement in Y character identification accuracy”. Further, it is assumed that there are a plurality of parameters related to weights such as variable 1, variable 2, and variable 3 as items that are considered to be related to the improvement of character identification accuracy. For example, the variable may be the standard blackness of the upper left area of the character, the stroke density of the right half area, the average strength of the diagonal directionality of the character outline, or the like. Further, considering the deformation of the character pattern as a parameter that can affect the character identification accuracy, the above [Equation 2] can be further regarded as a function of the character deformation parameter. That is,

It becomes.

The regression analysis parameters a1 to [Equation 2] can be easily obtained by solving a linear equation. In addition, the learning of the parameters of [Equation 3] can be performed by gradually changing the parameters by the gradient method.
In general, when implementing learning by the gradient method, a partial differential equation is derived by manual calculation from a fixed function definition expression (e.g., a calculation formula for a neuro hidden layer or a parameter superposition of a polynomial discriminant function). Implement the original learning program. However, the addition and deletion of rules can be performed dynamically. Furthermore, the target function can be changed depending on whether the importance of improving character identification accuracy is important or the number of cases. If the calculation formula of the objective function can be dynamically changed, and it is necessary to perform parameter learning by the gradient method accordingly, the partial differential equation must also be changed dynamically.

  A program (function) for obtaining a target function consists of an if statement, a for statement, a mathematical function, and a mathematical operation. Of these, the mathematical function / math operation part is described using the number structure of automatic differentiation. Since the value and the differential value can be obtained simultaneously from the defined function by using automatic differentiation, the differential value can be easily derived even when the calculation formula is changed. In addition, by combining with regression analysis, parameter adjustment can be performed focusing on rules that are considered effective for improving character identification accuracy.

0201 ... Character pattern for learning, 0204 ... Simple learning dictionary, search complement dictionary / confusion matrix information ... 0206, search complement dictionary / fuzzy / darkness information ... 0209, 0211 ... handwritten pattern synthesis, 0213 ... higher classifier learning 0601 ... Image input unit 0602 ... Document structure analysis unit 0603 ... Character line extraction unit 0604 ... Character string hypothesis creation unit 0605 ... Character string notation analysis unit 0606 ... Text output unit 0601 ... Input to a conventional document processing system Paper document

Claims (6)

A sample character image input receiving unit for receiving input of a sample character image;
A character component extraction unit that extracts a character component based on the sample character image;
A pseudo-character model generation unit that generates a pseudo-character model based on the character parts;
An identification dictionary generating unit for generating an identification dictionary by generating a character identification pattern based on the pseudo-character model;
A character identification system comprising: The character identification system according to claim 1,
A character image input receiving unit for receiving input of a character image;
An identification unit that identifies the character image using the identification dictionary and generates an identification result;
A character identification system further comprising: The character identification system according to claim 2,
An identification result output for outputting the identification result;
An identification result success / failure accepting unit that accepts input of success / failure information of the identification result;
A feedback unit that updates the character identification pattern of the identification dictionary based on the success / failure information;
A character identification system further comprising: The character identification system according to claim 1,
It further has a parts information database storing part information constituting the character pattern,
The character part includes stroke order information,
The character part extraction unit extracts the character parts based on the stroke order information. The character identification system according to claim 2,
A flat information database for storing flat information indicating the relationship between flat and characters;
A character candidate interpolation unit that extracts characters related to the identification result as character candidates using the prone information;
A character identification system further comprising: The character identification system according to claim 2,
An OCR unit that captures a document and converts it into electronic image data;
A document structure analysis unit for specifying a document structure of the document based on the electronic image data;
A character extraction unit that extracts a character image to be read based on the document structure and inputs the character image to the character image input reception unit;
A character identification system further comprising:

Images