WO2018003153A1

WO2018003153A1 - Recognition device and recognition method

Info

Publication number: WO2018003153A1
Application number: PCT/JP2017/001418
Authority: WO
Inventors: 昭森口
Original assignee: 株式会社日立ソリューションズ
Priority date: 2016-06-30
Filing date: 2017-01-17
Publication date: 2018-01-04
Also published as: JP2018005462A

Abstract

A recognition device is provided with: a processor for executing a program; and a storage device for storing the program, and has a recognition model for determining whether a character string extracted from a form is an item value row including an item value. The recognition model is generated by converting information relating to the rows including character strings in the form into histograms, analyzing a histogram of a row including a table heading and a histogram of a row including an item value, and machine-learning the relevance of the row structures. The recognition model extracts information relating to rows including character strings from a form to be recognized, converts the extracted information relating to the rows into histograms, and, by using, as a feature quantity, the relevance of row structures obtained by comparing the histogram of a row including a table heading and the histogram of a different row, determines whether the different row is an item value row.

Description

Recognition device and recognition method

Import by reference

This application claims the priority of Japanese Patent Application No. 2016-129997, which was filed on June 30, 2016, and is incorporated herein by reference.

The present invention relates to a recognition device that recognizes a table structure from a document such as a form.

Companies will exchange sales slips, invoices, receipts and other forms with other companies during their economic activities. Technology that converts documents in a form into electronic data using OCR (Optical Characterize Recognition, optical character recognition) in order to enter these business forms into a company's business system or account system, and perform shipping and deposit processing. Is being used. After the form is digitized using OCR, data in which neighboring character strings are associated is registered in the system. For example, when there is a character string “March 29, 2016” in the vicinity of the character string “form issue date”, the item name is “form effective date” and the item value is “March 29, 2016”. sign up. Further, the table structure in the form, that is, the item name of the table heading and the cell of the item value corresponding to the table heading are recognized using the ruled lines, and these are associated and registered in the system.

In Japanese Patent Laid-Open No. 2013-205974, a table structure is recognized by ruled lines, an item name is identified using an item name candidate database, and an item name and an item are determined based on the positional relationship between the item name and other item value candidate cells. A method is disclosed in which the likelihood of correspondence with a value is calculated, and the item name and the item value are associated with each other so that the likelihood is highest in the entire table structure.

Japanese Patent Laid-Open No. 2013-190993 discloses a ruled line in which the ruled line becomes a boundary between an item name and an item value due to differences between items described across the ruled line, such as differences in background color, font size, font type, and the like. Is described, and a method for estimating an item name and an item value in a table structure and a correspondence relationship thereof is described.

In U.S. Pat. No. 8,214,733, item names and item values have similarities in the horizontal start position and end position in the form, and there are characters between the line containing the table heading and the item value. A method of associating a table headline with a line including an item value and associating an item name with an item value is described using the fact that similarity is found in the appearing coordinate positions.

In the methods described in Japanese Patent Laid-Open Nos. 2013-205974 and 2013-190993, ruled lines are used as a clue for recognizing the table structure, but cannot be used for recognizing the table structure of a form on which no ruled line is described. .

Furthermore, depending on the form, a character string that is not related to the table heading may be described between the table heading and the line including the item value (hereinafter referred to as the item value line) or between the item value lines. . For example, in the case of invoices and receipts, the item value line contains the product name and price, but if there is a shortage of inventory and more time is required for delivery of the product, the period and reason for delay in delivery, etc. The supplementary information is described at the top or bottom of the item value line. In addition, information on discounts for product purchases during the sales promotion period and product purchases in bulk is described near the item value line. In particular, in the method described in Japanese Patent Laid-Open No. 2013-205974, since the likelihood is calculated between adjacent items, if an item is divided by an irrelevant character string, the item name cannot be correctly associated with the item value. . Further, in the method described in Japanese Patent Application Laid-Open No. 2013-190993, since the boundary between the item name and the item value is identified using the feature between nearby items, it is difficult to identify the boundary due to the division by the supplementary information. .

Further, in the method described in US Pat. No. 8,214,733, the start position and end position of character strings are compared, and a line (character string line) including a character string in a form is represented by coordinates where a character exists. Is converted to binary data with 1 being blank and 0 being blank, and by calculating the Hamming distance between the binary data of the table header and the binary data of the character string row, the table header, the item value row, and other character string rows To distinguish. However, the start position and end position of the character string are not necessarily the same between the table heading and the item value line, and the number of character strings in the table heading and the number of character strings in the item value line may be different. For this reason, the hamming distance between the table heading and the item value line becomes larger than the hamming distance between the table heading and a line including another character string, making the association difficult.

For this reason, it is necessary to associate a table heading with an item value even in a form that has no ruled line and a character string that is not related to the table heading appears in the table structure.

A typical example of the invention disclosed in the present application is as follows. That is, a recognition apparatus comprising a processor that executes a program and a storage device that stores the program, and a recognition model that determines whether a character string extracted from a form is an item value line including an item value The recognition model converts line information including a character string in a form into a histogram, analyzes a line histogram including a table heading and a line histogram including an item value to determine the relevance of the line structure. The recognition model is generated by machine learning, and the recognition model extracts line information including a character string from a form to be recognized, converts the extracted line information into a histogram, and includes a line including a table heading. It is determined whether the other row is an item value row using the relationship of the row structure obtained by comparing the histogram of the other and the histogram of the other row as a feature amount.

According to one aspect of the present invention, it is possible to accurately associate table headings with item values. Problems, configurations, and effects other than those described above will become apparent from the description of the following embodiments.

It is a block diagram of the in-form table | surface structure recognition system of the Example of this invention. It is a block diagram which shows the physical structure of a recognition server. It is a figure which shows an example of the form which a recognition server recognizes. It is a flowchart of the process by an item value line learning program. It is a figure which shows an example of the form for learning. It is a figure which shows an example of the histogram produced | generated from the table | surface heading of the form. It is a figure which shows the recognition model of the neural network which performs horizontal direction item value learning. It is a figure which shows the example of a neighborhood line feature-value production | generation process. It is a figure which shows the structural example of a near neighbor line feature-value table. It is a figure which shows a vertical direction item value line recognition neural network model. It is a flowchart of the process by an item value line recognition program and an item value recognition program. It is a figure which shows the method of matching an item name and an item value. It is a figure which shows the structural example of an item name / item value database.

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a configuration diagram of the in-form table structure recognition system according to the embodiment of the present invention.

The in-form table structure recognition system according to the present embodiment includes a recognition server 100 that extracts item names and item values from a form. The recognition server 100 is connected to a reading device 112 that digitizes a paper form 111 received by mail from a business partner. The recognition server 100 is connected to a network (for example, the Internet 114), and receives an electronic form from the customer company PC 113.

The recognition server 100 includes a form receiving unit 109, an item value line learning program 101, an item value line recognition program 102, and an item value recognition program 103. Further, the recognition server 100 has an item name database 105 in which item names to be acquired from the form are registered.

The form receiving unit 109 stores the electronic form received via the reading device 112 or the Internet 114 as a learning form 104 or a recognition target form 106 together with a supplier company name. The item value line learning program 101 uses the line including the item name registered in the item name database 105 as a table headline, and the correspondence between the table headline and the item value line from the learning form 104 where the position of the item value line is known. The relationship is machine-learned to generate an item value line recognition model 107 (see FIG. 3). The item value line recognition program 102 recognizes and extracts an item value line in the recognition target form 106 using the item value line recognition model 107 (see FIG. 10). The item value recognition program 103 associates the item value in the item value row with the item name of the table header, and stores it in the item name / item value database 108 shown in FIG. 11 (see FIG. 10).

FIG. 1B is a block diagram showing a physical configuration of the recognition server 100.

The recognition server 100 of this embodiment is configured by a computer having a processor (CPU) 1, a memory 2, an auxiliary storage device 3, and a communication interface 4.

The processor 1 executes a program stored in the memory 2. The memory 2 includes a ROM that is a nonvolatile storage element and a RAM that is a volatile storage element. The ROM stores an immutable program (for example, BIOS). The RAM is a high-speed and volatile storage element such as DRAM (Dynamic Random Access Memory), and temporarily stores a program executed by the processor 1 and data used when the program is executed.

The auxiliary storage device 3 is configured by a large-capacity and non-volatile storage device such as a magnetic storage device (HDD) or a flash memory (SSD), for example, and stores a program executed by the processor 1 and data used when the program is executed. Store. That is, the program is read from the auxiliary storage device 3, loaded into the memory 2, and executed by the processor 1.

The communication interface 4 is a network interface device that controls communication with other devices (reading device 112, customer company PC 113) according to a predetermined protocol.

The recognition server 100 may have an input interface 5 and an output interface 8. The input interface 5 is an interface to which an input from an operator is received, to which a keyboard 6 and a mouse 7 are connected. The output interface 8 is an interface to which a display device 9 or a printer is connected, and the execution result of the program is output in a form that can be visually recognized by the operator.

The program executed by the processor 1 is provided to the recognition server 100 via a removable medium (CD-ROM, flash memory, etc.) or a network, and stored in the nonvolatile auxiliary storage device 3 that is a non-temporary storage medium. For this reason, the recognition server 100 may have an interface for reading data from a removable medium.

The recognition server 100 is a computer system configured on a single computer or a plurality of computers configured logically or physically, and operates in a separate thread on the same computer. Alternatively, it may operate on a virtual machine constructed on a plurality of physical computer resources.

In the recognition server 100, all or part of the functional blocks implemented by the program may be configured by a physical integrated circuit (for example, Field-Programmable Gate Array).

FIG. 2 is a diagram illustrating an example of a form recognized by the recognition server 100.

The form shown in FIG. 2 is an invoice from Company A to Company B. The products and prices purchased by Company B are listed in the form in a table structure, and the table heading 201 includes the number of products (Quantity), the product number (Item No.), the description of the product (Description), and the unit price (UNIT). The item names of “PRICE” and total price (PRICE) are described. In the

item value rows

202, 204, and 206, item values corresponding to the item names of the table headings are described. Further,

supplementary information

203 and 205 for supplementing the item value line is described between the

item value lines

202, 204 and 206. Further, an Invoice Number 207 that uniquely identifies the form is assigned to the form for each business partner company. The learning form 104 sets the rectangular coordinates of the table heading 201 and the

item value lines

202, 204, and 206 as correct data for machine learning.

FIG. 3 is a flowchart of processing by the item value line learning program 101.

First, the item value line learning program 101 receives an input of the learning form 104 (step S301).

Next, the rectangular coordinates of the character string row are extracted from the learning form 104 (step S302). In step S <b> 302, a rectangle as shown in FIG. 4 is extracted from the learning form 104.

Thereafter, the learning form 104 is subjected to OCR processing, and character information and the coordinates of the character are extracted (step S303). Then, from the OCR result, a character that matches the item name registered in the item name database 105 is specified, and the coordinates of the specified character on the form are specified as the position of the table heading (step S304).

A histogram of character pixels in the rectangle is generated for all the character string rows extracted as a rectangle in step S302 (step S305). This histogram represents the structural features of the rows in the horizontal direction. Specifically, after dividing a rectangle of a character string row by a certain number in the horizontal direction, the number of black pixels contained in characters in the divided area is set as the frequency of the histogram. A histogram generated from the table heading 201 of the form shown in FIG. 2 is shown in FIG.

Next, horizontal item value learning is performed (step S306). The horizontal item value learning is a process in which the neural network learns the relationship between the table header and the structure of the item value row from the horizontal histogram representing the pixel distribution generated in step S305. Table headings and item value rows are: (1) the number of character strings is the same or close, (2) character strings exist at a common position in the horizontal direction, and (3) item values are indicated by item names in the table headings. There are patterns such that the character string length is greater than or equal to a predetermined value or less, and the neural network learns this pattern. For example, the character string length of the item value corresponding to the item name Description tends to be long, and the character string length of the item value corresponding to the item name Quantity tends to be short.

FIG. 6 is a diagram showing a recognition model of a neural network that performs horizontal item value learning.

The horizontal direction item value line recognition neural network model 610 shown in FIG. 6 takes a table header histogram 601 and a character string line histogram 602 as input values. The table heading histogram 601 is a histogram generated in step S305 for the rectangle of the table heading specified in step S304. The character string row histogram 602 is a histogram generated in step S305 for a character string rectangle other than the table header extracted in step S302.

The horizontal item value row recognition neural network model 610 includes a feature amount extraction layer A611 that extracts the feature amount of the structure of the table header histogram 601 and a feature amount extraction layer B612 that extracts the feature amount of the structure of the character string row histogram 602. , And a comparison layer 613 that compares the two feature amounts. In the feature quantity extraction layer A611, learning is performed so that the position of the character string in the table header, the number of character strings, and the position of a specific item name (for example, Description) are extracted as the feature quantity. In the feature amount extraction layer B612, learning is performed so that the position of the character string in the character string row, the number of character strings, and the length of the character string are extracted as feature amounts. The comparison layer 613 evaluates the likelihood that the structure of the character string row histogram 602 is likely to be the structure of the item value row corresponding to the table header histogram 601 from the two feature amounts. Specifically, the position of the character string in the character string row, the number of character strings, and the length of the character string with respect to the table heading correspond to each of the position of the character string of the table heading, the number of character strings, and the item name. Likelihood is learned. The output of the comparison layer 613 is the item value row probability 614.

For the horizontal item value line recognition neural network model 610, for each character string line extracted from the form, the output is 1 when the table heading histogram 601 and the item value line histogram of the learning form 104 are input, and the learning form. The learning is executed by a known neural network learning method (for example, error back-propagation method) so that the output when inputting the histogram of the table header histogram 601 and the character string row other than the item value row becomes zero. To do.

In step S306, the item value row can be estimated from the structural features of the table header and the item value row.

Subsequently, a neighboring line feature value generation process for generating a feature value that can be input to the neural network from information in the peripheral space of the item value line is performed (step S307). When information on the space around the item value row is used as an additional feature amount, the item value row can be estimated with higher accuracy. Specifically, the peripheral space information includes ruled lines, blanks, and similar character string rows. Depending on the form, a ruled line is described between the table heading and the item value line, or at the end of the table structure. Therefore, the ruled line is effective information for determining the existence range of the item value line. In addition, depending on the form, a certain amount or more of space is provided between the table structure and the non-table structure. Therefore, the space is effective information for determining the existence range of item value rows. Further, when there are a plurality of item value rows in the table structure, row structures having similar feature quantities repeatedly exist within a certain range, and the relative position of the similar row structure is information useful for determining an item value row. Therefore, it is possible to improve the recognition accuracy of the item value line by causing the neural network to learn information in which ruled lines, blanks, and similar character string lines exist.

7A and 7B are diagrams illustrating an example of the neighborhood line feature value generation process.

In the example shown in the figure, feature amounts are generated from the top and bottom 10 lines as the space around the character string line 701 of the form 700. Specifically, the target range is 10 neighboring

rows

702 and 703 in which each character string row is one row, a blank portion having the same height as the character string row 701 is one row, and a ruled line is one row.

7B includes a neighboring row feature amount table 710 including neighboring

row numbers

704 and 711 assigned to each neighboring row and a feature amount 712 of each neighboring row. The feature quantity 712 is a value calculated by the horizontal direction item value line recognition neural network model 610 generated in step S306. The probability that each character string line is an item value line (Possibilities), whether it is blank (Blank), a ruled line (Line) or table header (Header). For example, Possibilities compare the row structures of the rows and determine that a row having the same or similar row structure is likely to be an item value row.

Next, vertical direction item value row learning is performed using the neighboring row feature quantity generated in step S307 as an input (step S308). As shown in FIG. 8, the vertical item value line recognition neural network model 802 generated by the vertical item value line learning is the same as the horizontal item value line recognition neural network model 610 with the neighboring line feature 801 as an input. The item value row probability 803 is output. For each character string row extracted from the form, learning is performed using the inverse error propagation method so that 1 is output when the character string row 701 is an item value row and 0 is output when the character string row 701 is a non-item value row. To do.

FIG. 9 is a flowchart of processing by the item value line recognition program 102 and the item value recognition program 103.

First, the item value line recognition program 102 acquires the recognition target form 106 together with the business partner company name (step S901).

The processing from step S902 to step S905 is the same as the processing from step S302 to step S305 by the item value line learning program 101.

In step S906, the table header histogram 601 and the character string row histogram 602 generated by the processing up to step S905 are input for each character string row of the recognition target form 106, and the horizontal direction item value line recognition generated in step S306 is input. The probability that the character string row is the item value row is calculated by the neural network model 610 (step S906).

Using the probability of the item value line calculated in step S906, a neighboring line feature amount is generated for each character string line of the recognition target form 106, similarly to step S307 by the item value line learning program 101 (step S907). .

The probability that the character string line is the item value line is calculated from the neighboring line feature amount generated in step S907 by the vertical direction item value line recognition neural network model generated in step S308 (step S908).

Specifically, after a predetermined number of blank lines continue, it is determined that there is a low possibility that the character string line is an item value line. Further, it is determined that a row having the same or similar row structure is likely to be an item value row. Further, it is determined that the character string line between the two ruled lines is highly likely to be an item value line, and after the bottom ruled line, it is determined that the possibility of being an item value line is low.

The character string row having the probability of being the item value row calculated in step S908 is determined to be the item value row, and the item name in the table header is associated with the item value in the item value row. A method of associating the item name with the item value is shown in FIG. Of the item names stored in the item name database 105, the number of item names included in the table heading is calculated. The item name database 105 includes Quantity, Item No. , Description, UNIT PRICE, and PRICE. At this time, it can be determined that the table heading 1001 includes five item names. Note that UNIT PRICE in the table heading 1001 corresponds to UNIT PRICE and PRICE in the item name database 105, but item names having a long character string length are preferentially used. Subsequently, the number of character strings is calculated by dividing the character string in the item value line by the minimum space. If the number of character strings is different from the number of item names in the table heading 1001, the length of the space separating the character strings is increased, and the number of character strings is calculated again. Until the number of item names in the table heading becomes equal to the number of character strings in the item value row, the process is repeated with the blank length increased to determine the item value. For example, in the item value row 1002, the character string is divided with a blank between Office and Chair, and the number of character strings is 6. When the blank length between P000115 and Office is used for character string division, the number of character strings is 5 (1003). That is, small blanks are excluded so that the number of items in the item value row is the same as the number of items in the table header. Therefore, in the case shown in FIG. 10, 4, P000115, Office Chair, $ 40.00, and $ 160.00 are item values. The obtained item values are associated with the item names in the table header in order from the left (step S909).

Next, the form number is extracted (S910). Specifically, the Invoice Number is extracted from the OCR result extracted in Step S903. The Invoice Number is generally a character string that includes a numerical value that exists immediately to the right of or directly below the character string Invoice Number on the form, so that it can be easily distinguished from other character strings in the form. In the form shown in FIG. 2, 111111 on the right side of the character string Invoice Number is extracted.

The item value recognition program 103 stores the supplier company name acquired in step S901, the item name and item value associated in step S909, and the Invoice Number extracted in step S910 in the item name / item value database 108. (Step S911).

FIG. 11 is a diagram showing a configuration example of the item name / item value database 108.

The item name / item value database 108 stores the item value 1103 corresponding to the supplier company name 1101, the Invoice Number 1102, and the item name (Quantity, Item No., Description, Unit Price, Price). In the forms shown in FIGS. 2 and 10, as shown in the bottom row of FIG. 11, Company A as Company, 111111 as Invoice Number, 4, 111 as Quantity, Item No. P000115 is stored as the description, Office Chair as the description, 40 as the unit price, and 160 as the price.

As described above, according to the embodiment of the present invention, the item value line recognition model 610 extracts line information including a character string from a form to be recognized, converts the extracted line information into a histogram, Analyzing the histogram of the row including the table heading and the histogram of the other row and using the relationship of the row structure as a feature amount to determine whether the other row is an item value row. The value can be accurately associated.

In addition, since the line information is rectangular information determined to include a character string, position information of the rectangle, and character information recognizing the character string, the area to be analyzed in the form is limited, and calculation is performed. The amount can be reduced.

In addition, since the histogram is configured to represent the number of black pixels included in a character in an area obtained by dividing a rectangle defined to include the character string in the row into a predetermined number in the horizontal direction, the sum of characters in the row is represented. You can quantify the position of the characters.

Also, it extracts line information including text strings from the form, converts the extracted line information into a histogram, analyzes the line histogram including the table header and the line histogram including the item value, and relates the line structure. Since the item value line recognition model 610 is generated by machine learning using the characteristic as a feature quantity, a value suitable for machine learning, which is a quantitative value representing a structural characteristic of a line, is used rather than inputting a character itself. To generate models for analyzing forms.

In addition, the item value line recognition model 610 uses at least one of the ruled line, the blank, and the position of the character string line having the same structure included in the form to be recognized, and the other line is an item value line. Since it is determined whether it exists, the accuracy which recognizes an item value line can be improved.

In addition, since the item value line recognition model 610 determines that there is a low possibility of being an item value line after a predetermined number of blank lines continue, the item value line can be recognized with high accuracy even in an unknown form.

In addition, since the item value line recognition model 610 determines that there is a high possibility that lines having the same line structure are item value lines, the item value line can be recognized with high accuracy even in an unknown form.

The item value line recognition model 610 determines that a line between two ruled lines is highly likely to be an item value line, and that a line below the bottom ruled line is unlikely to be an item value line. The item value line can be recognized with high accuracy even in the form.

The present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the scope of the appended claims. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and the present invention is not necessarily limited to those having all the configurations described. A part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Moreover, you may add the structure of another Example to the structure of a certain Example. In addition, for a part of the configuration of each embodiment, another configuration may be added, deleted, or replaced.

In addition, each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware by designing a part or all of them, for example, with an integrated circuit, and the processor realizes each function. It may be realized by software by interpreting and executing the program to be executed.

Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, and an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, and a DVD.

Also, the control lines and information lines indicate what is considered necessary for the explanation, and do not necessarily indicate all control lines and information lines necessary for mounting. In practice, it can be considered that almost all the components are connected to each other.

Claims

A recognition device,
A processor that executes the program; and a storage device that stores the program;
A recognition model that determines whether a character string extracted from a form is an item value line including an item value;
The recognition model converts line information including text in a form into a histogram, analyzes the line histogram including table headings and the line histogram including item values, and performs machine learning on the relationship between the line structures. Generated by
The recognition model is
Extract line information including character strings from the form to be recognized,
Converting the extracted row information into a histogram;
Recognizing apparatus characterized by determining whether or not the other row is an item value row by using, as a feature amount, the relation of the row structure obtained by comparing the histogram of the row including the table header and the histogram of the other row. .
The recognition device according to claim 1,
The line information includes rectangular information determined to include a character string, position information of the rectangle, and character information that recognizes the character string.
The recognition device according to claim 2,
The recognition apparatus, wherein the histogram represents the number of black pixels included in a character in an area obtained by dividing a rectangle defined to include a character string in a row into a predetermined number in the horizontal direction.
The recognition device according to claim 1,
Extract line information including character strings from the form, convert the extracted line information into a histogram, analyze the line histogram including the table header and the line histogram including the item value, and relevance of the line structure The recognition device is characterized by generating the recognition model by machine learning using as a feature quantity.
The recognition device according to claim 1,
The recognition model determines whether the other line is an item value line by using at least one of a ruled line, a blank, and a character string line having the same structure included in the form to be recognized. A recognition device characterized by that.
The recognition device according to claim 5,
The recognition apparatus determines that the possibility that the recognition model is an item value line after a predetermined number of blank lines continues is low.
The recognition device according to claim 5,
The recognition apparatus determines that a line having the same line structure is likely to be an item value line.
The recognition device according to claim 5,
The recognition apparatus determines that a line between two ruled lines has a high possibility of being an item value line, and that a line below the lowest ruled line has a low possibility of being an item value line.
A recognition method executed by a recognition device,
The recognition device is
A processor for executing the program; and a storage device for storing the program;
A recognition model that determines whether a character string extracted from a form is an item value line including an item value;
The recognition model converts line information including text in a form into a histogram, analyzes the line histogram including table headings and the line histogram including item values, and performs machine learning on the relationship between the line structures. Generated by
The method
The recognition model extracts line information including character strings from the form to be recognized,
The recognition model converts the extracted row information into a histogram;
The recognition model determines whether the other row is an item value row by using, as a feature amount, the relevance of the row structure obtained by comparing the histogram of the row including the table header and the histogram of the other row. Recognition method as a feature.
The recognition method according to claim 9, comprising:
The recognition method according to claim 1, wherein the line information includes rectangular information determined to include a character string, position information of the rectangle, and character information that recognizes the character string.
The recognition method according to claim 10, comprising:
The recognition method, wherein the histogram represents the number of black pixels included in a character in an area obtained by dividing a rectangle defined to include a character string in a line into a predetermined number in the horizontal direction.
The recognition method according to claim 9, comprising:
Extract line information including character strings from the form, convert the extracted line information into a histogram, analyze the line histogram including the table header and the line histogram including the item value, and relevance of the line structure The recognition method is characterized in that the recognition model is generated by machine learning as a feature quantity.
The recognition method according to claim 9, comprising:
The recognition model determines whether the other line is an item value line by using at least one of a ruled line, a blank, and a character string line having the same structure included in the form to be recognized. A recognition method characterized by the above.
The recognition method according to claim 13, comprising:
The recognition model is characterized by determining that there is a low possibility that the recognition model is an item value line after a predetermined number of blank lines continue.
The recognition method according to claim 13, comprising:
The recognition model is characterized by determining that a row having the same row structure is likely to be an item value row.
The recognition method according to claim 13, comprising:
The recognition model is characterized by determining that a line between two ruled lines is highly likely to be an item value line, and that a line below the bottom ruled line is unlikely to be an item value line.