JP2013190993A - Device and method for analyzing table structure - Google Patents

Abstract

PROBLEM TO BE SOLVED: To stably extract item names-data relation even when an item name word is not known in advance, and an item name dictionary cannot be fully provided in technology for extracting the item name-data relation from unspecified and a large amount of documents.SOLUTION: Form features and character string features of all pairs of adjacent frames in a table are referred to, and a differential score showing difference among them is set to a contact of the pair of frames. Next, to all rule grids in the table, the differential score set to a frame contact belonging to the rule grids is projected (such as taking the sum, taking the average), and an item name-data boundary score is calculated. The item name-data boundary score is a certainty factor representing whether or not the rule grids are ruled lines at the boundaries between an item name frame and a data frame, and set based on a policy that the contact at which difference appears in frame features is the boundary between the item name frame and the data frame. Next, the item name frame in the table is determined from a position of the item name-data boundary, and the item name-data relation is determined based on adjacent relation with other frames.

Description

  The present invention relates to a table analysis method in a document, and more particularly to a method for extracting an item name in a table in a document or a correspondence relationship between an item name and data.

  In the table analysis technology, the item name, which is a keyword representing the character string attribute, and the data representing the character string value are extracted from the character strings described in the table, and the item name-data relationship (item name and data There is a method for extracting the correspondence relationship with data). In the conventional item name-data relationship extraction method, the item name-data relationship is extracted by the user registering the item name to be read in the item name dictionary in advance.

  As an application destination of the item name-data relation extraction method, there is a form recognition system. For example, when a form recognition system is used in a financial institution or the like, information necessary for processing the form is automatically extracted from the form using the form recognition and transferred to the business system. Thereby, the operation | work which a user inputs required information into a business system manually can be automated. In such an application example, the item name-data relationship to be extracted is “transfer account number”, “delivery amount” and specific amount, or “delivery date” and specific date, etc. Often known in advance. In addition, the types of corresponding item name words are limited and often known in advance. Therefore, an item name word corresponding to the item name-data relationship to be extracted is prepared in advance as an item name dictionary by the user or system administrator.

  In the technique disclosed in Patent Document 1, the character string in the table is matched with the item name word registered in the item name dictionary, and the character string that matches the item name word is not matched with the item name or item name word. The column is determined as a data candidate. Further, the correspondence between the item name and the data is determined from the arrangement relationship between the item name and the data candidate.

  Further, Patent Document 2 extracts ruled lines from a document, extracts intersections and end points of two ruled lines, and detects the upper right corner, upper left corner, lower right corner, and lower left corner corresponding to the four corners of the rectangular frame. , A technique for detecting a frame from within a document is disclosed.

JP 2008-204226 A Japanese Patent Laid-Open No. 11-053466

  In a document management system that handles a large number of unspecified documents, using the item name-data relation extraction method, search for similar or related documents and integrate multiple tables to improve the efficiency of handling large volumes of documents. Can be achieved. However, when targeting an unspecified and large number of documents, the item name-data relationship to be extracted is not known in advance, the item name word that is the key for extracting the item name-data relationship is unknown, or part of it Even if it is known, it is difficult to complete the item name dictionary due to the existence of fluctuations in the notation. For this reason, it is difficult to apply the method of extracting the item name-data relationship using the words registered in the item name dictionary as a key to a document management system that handles a large number of unspecified documents.

  In view of the problems in the prior art as described above, the object of the present invention is to extract the item name-data relationship from an unspecified and large amount of documents. Even when it is not complete, it is possible to stably extract the item name-data relationship.

  Another object of the present invention is to enable the extraction of the item name-data relationship to be extracted without requiring the user to maintain the fluctuation of notation one by one, even in a document group in which there are many fluctuations of the item name. There is to do.

  In order to achieve the above object, according to one aspect of the present invention, there is provided a table structure analyzing apparatus including a control unit, a storage unit connected to the control unit, and an input unit to which image information is input. The control unit detects a frame region constituting a table included in the image from the image information input from the input unit, calculates a frame feature amount indicating a feature amount of the detected frame region, and sets two adjacent frames The difference between the frame feature amounts between the regions is calculated, the calculated difference between the frame feature amounts is associated with the line segment between two adjacent frame regions, and the line segment constituting the ruled line constituting the table is associated. Based on the difference in the frame feature value, a ruled line that is a boundary between the frame area where the item name exists in the table and the frame area where the data exists is extracted as a boundary line, and the item name in the table based on the boundary line Exists, or there is a frame area and data where the item name exists. Determining the correspondence between the frame area.

  More preferably, the storage unit stores an item name dictionary in which item names are registered, and the control unit stores a character string in the frame area determined as the frame area in which the item name exists in the item name dictionary. Register as

  More preferably, in the item name dictionary, a character string and likelihood information indicating the likelihood of the item name of the character string are associated and registered, and the control unit includes the likelihood information registered in the item name dictionary. The likelihood information corresponding to the character string in the frame area determined to be the frame area in which the item name exists is changed.

  In addition, in order to extract a ruled line that is a boundary between a frame area in which an item name exists in the table and a frame area in which data exists, in one aspect of the present invention, the control unit may A frame feature difference value indicating a difference between the frame feature amounts is calculated, the calculated frame feature difference value is associated with a line segment between two adjacent frame regions, and is associated with a line segment constituting a ruled line. Based on the frame feature difference value, boundary line likelihood information indicating the likelihood of the border of the ruled line is calculated, and when the border line likelihood information of the ruled line is a predetermined value or more, the ruled line is a boundary line judge. Moreover, you may determine with a predetermined number of ruled lines with large likelihood information being a boundary line among several ruled lines which comprise a table | surface.

  According to the present invention, it is possible to extract the item name and the correspondence between the item name and data from an unspecified and large amount of documents. In addition, even when the item name word is not known in advance or when there are many fluctuations in the notation of the item name, it is possible to stably extract the item name and the correspondence between the item name and the data. .

It is a processing flow figure of the item name-data relation extraction method in the example of the present invention. FIG. 5 is a basic processing flow diagram as a premise of an item name-data relationship extraction method in an embodiment of the present invention. It is a module block diagram of the item name-data relationship extraction system in the Example of this invention. It is a flowchart of the frame feature difference score calculation process in the Example of this invention. It is a flowchart of the item name-data boundary line detection process in the Example of this invention. It is an example of the frame characteristic in the Example of this invention. It is an example of calculation of a frame feature difference score in the embodiment of the present invention. It is an example of the detection of the item name-data boundary line in the Example of this invention. It is a processing flow figure of the example of application of the item name-data relationship extraction system in the Example of this invention. It is a processing flow figure of the example of application of the item name-data relationship extraction system in the Example of this invention. It is an example of the item name dictionary in the Example of this invention. It is a correction example of the item name dictionary in the Example of this invention. It is a flowchart of the item name dictionary correction process in the Example of this invention. It is an example of the extraction result of an item name-data relationship. It is an example of the extraction result of an item name-data relationship. It is a figure which shows the correspondence of the ruled line in a table | surface, and a tangent. It is an example of the item name-data boundary line score table which shows the correspondence with the score which shows the rule line and the score which shows the tangent which comprises a ruled line, and the boundary line of a ruled line. It is an example of the frame feature difference score table which shows the correspondence of a tangent and a frame feature difference score.

  The present invention is a method for extracting the correspondence between item names and data (item name-data relationship) from a table. First, the features of all the frame regions in the table (refer to the frame format feature and the character string feature in the frame, and the frame feature or the frame feature in the following also has the same meaning) are extracted. Taking a frame as one pair, a difference score indicating a difference in characteristics between two frames constituting the pair is obtained for each pair. Then, the difference score is associated with a tangent of a pair of frames (a line segment between two adjacent frames). Next, for all ruled lines in the table, the difference score set for the frame tangent belonging to the ruled line is projected (summed, averaged, etc.), and the item name-data boundary score is calculated. The item name-data boundary score is a value indicating a probability (likelihood) indicating whether the ruled line is a ruled line that is a boundary between the item name frame and the data frame. Since the tangent line with a large difference score is estimated as the boundary between the item name frame and the data frame, it is determined that the ruled line including the tangent line with the large difference score is a ruled line serving as the boundary between the item name frame and the data frame. Next, the item name frame in the table is determined from the position of the item name-data boundary line, and the item name-data relationship is determined based on the adjacency relationship with other frames.

The table is composed of a plurality of ruled lines, and has two or more frames (frame regions) including item names and data. In the present invention, a line surrounding the frame region is called a line segment, and the line segment constitutes a ruled line. Usually, the frame is a rectangle, and in this case, the frame is composed of four line segments that surround four sides. In addition, a line segment between two adjacent frame regions (a common line segment of two adjacent frames) is referred to as a tangent line (or abbreviated tangent line) between the two adjacent frames.
<First embodiment>
Hereinafter, an item name-data relationship extraction method according to an embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 3 is a block diagram showing the configuration of the table structure analysis computer 331 in which the item name-data extraction method according to one embodiment of the present invention is constructed. The table structure analysis computer 331 includes a memory 309, an external storage device 314, input / output interfaces (I / F) 311 and 312, and a CPU 310, and inputs a document to be processed via the input / output I / Fs 311 and 312. A document input device 301 and an output device 303 that outputs a processing result such as an item name-data relationship. The table structure analysis computer 331 may include a network I / F 313 and be connected to the network 315 via the network I / F 313.

The memory 309 or the external storage device 314 includes an OS, a frame detection unit (frame detection program) 307 that detects a frame from the input document, a character string detection unit (character string detection program) 308 that detects a character string from the input document, and an item name -Item name for extracting data relationship-Data relationship extraction unit (item name-data relationship extraction program) 302, Item name dictionary 304 for storing item name words to be extracted, Item name extracted from document-Item indicating data relationship The name-data relation information 305 and the input document 306 are stored. The CPU 310 executes various programs loaded in the memory 309.
The document input device 301, the CPU 310, the output device 303, the memory 309, and the external storage device 314 may be connected via a network or the like without depending on physical connection means, and may be configured on different computers. .

Hereinafter, the processing flow of item name-data relationship extraction will be described with reference to FIG. 1. First, as a premise of the present invention, the basic processing flow of the item name-data relationship extraction method will be described with reference to FIG. In the processing flow shown in FIG. 2, the item name dictionary is used to extract the character string of the item name and the frame to which the character string belongs, and the item name-data relationship is extracted from the arrangement relationship of the frame. The present invention is configured by including the frame feature difference score calculation process S140 and the item name-data boundary detection process S150 of FIG. 1 instead of the item name collation step S230 of the basic process flow of FIG.
First, in the frame detection process in step S210, a frame is detected from the input document. Next, in the character string detection in step S220, a character string is detected from the document. As a specific example of steps S210 and S220, for example, as in the technique disclosed in Patent Document 2, ruled lines are extracted from a document, intersections and end points of two ruled lines are extracted, and upper right corners corresponding to the four corners of the rectangular frame are extracted. By detecting the upper left corner, lower right corner, and lower left corner, it is possible to use a method for detecting a frame from within a document. In addition, in electronic documents represented by PDF (Portable Document Format), Microsoft Office (registered trademark), and the like, ruled lines, frames, and character string rectangle information may be added to the document in advance. In this case, step S210 is performed. Alternatively, the process of step S220 or both can be omitted.

In the item name dictionary collation processing in step S230, each character string extracted in step S220 is collated with the item name dictionary 304, and whether or not the character string matches a word or a notation grammar registered in the item name dictionary 304. Determine. The item name dictionary 304 defines the character string of the item name to be extracted, and includes a list of words such as “Voltage [V]” and “Power”, “nnn-nnn” (n is a number), and “Type nn "and other grammar lists. At the time of matching with the dictionary, the likelihood indicating how much the character string matches the word or notation grammar in the dictionary is calculated, and if the score of the result of matching with the item name dictionary 304 is high, the character string is an item. If the score is low, it is determined as a data candidate. Specific examples of word collating used in step S230, the example, Non-Patent Document: "Lexical search approach for character string recognition", Proceeding of 3 rd Document Analysis System, pp.237-251, M. Koga, et Like the technology disclosed in .al., the optimal character string path is selected from the state transition network by matching the character network in which individual characters are regarded as nodes and the dictionary expressing the state transition network, and the matching result is obtained. There is a way.

  In the item name frame determination process in step S260, an item name frame is determined among all the frames constituting the table. In the flow of FIG. 2, the frame to which the character string determined as the item name in step S230 belongs is determined as the item name frame. Further, the character string belonging to the item name frame is determined as the item name character string, and the character string belonging to the other frame is determined as the data candidate character string. The method for determining other item name frames will be described later in the description of the processing flow of the invention method.

  In the item name-data relationship determining process in step S270, the arrangement relationship between the character strings is determined as the item name-data relationship for the character string determined as the item name and data candidate in the item name frame determining process in step S260. If it is valid, the character strings are extracted as the item name-data relationship. The determination as to whether the arrangement relationship is valid as the item name-data relationship is made based on the coordinates of the frame detected in step S210 and the coordinates of the character string area detected in step S220, and the frame to which the two character strings belong. This is performed based on the size, the arrangement relationship and the adjacency relationship, and the size and arrangement relationship of the two character string rectangles. For example, the character string in the frame adjacent to the frame to which the character string determined to be the item name belongs or the data candidate character string in the frame adjacent to the right is determined as data and extracted as the item name-data relationship There is.

  Note that the operation of the item name frame determination process in step S260 and the item name-data relationship determination process in step S270 is not limited to the above, and step S260 determines that any frame in the document is an item name frame or a data frame candidate. Have a function to determine whether or not, and step S270 has a function to determine whether the arrangement relationship between any two frames in the document is valid as an item name-data relationship and to extract the item name-data relationship If so, it does not depend on the above.

  Next, a processing flow in one embodiment of the present invention will be described with reference to FIG. The frame detection processing in steps S110 and S210, the character string detection processing in steps S120 and S220, and the item name-data relationship determination processing in steps S170 and S270 in FIGS. 1 and 2 are the same processing. In the basic processing flow of the item name-data relationship extraction method shown in FIG. 2, the present invention replaces the item name dictionary collating means and the item name dictionary 304 in step S230 with the frame feature difference score calculation processing in step S140, step S150. By adding the item name-data boundary detection process, the item name-data relationship can be extracted even if the item name dictionary is not complete for the item name-data relationship to be extracted Is. Hereinafter, the frame feature difference score calculation processing in step S140 and the item name-data boundary detection processing in step S150 will be described in detail.

  In the frame feature difference score calculation process in step S140, the difference between the features of the two frames is determined by referring to the features of the two adjacent frames with respect to the frame extracted in step S110 and the character string extracted in step S120. Calculate the frame feature difference score to represent. The frame feature difference score is a real value representing how different the format and the character string of the two frames are. Features to be used include frame format features (format features) and text features within the frames (character string features). For example, scanned images of input documents or electronic documents created from them. In the case of electronic documents such as Microsoft Office documents, in addition to the features described above, whether to set bold, whether to set italics, and font type , Font size, background color, etc. The calculated frame feature difference score is assigned to a frame tangent between two frames, and is stored in the frame feature difference score tables 1630 and 1640 in FIG. Frame tangents h1 to h13 in the frame feature difference score table 1630 are tangents that constitute the horizontal ruled lines H1 to H4 in the table shown in FIG. Frame tangents v1 to v14 in the frame feature difference score table 1640 are tangents that constitute the vertical ruled lines V1 to V3 in the table shown in FIG. Alternatively, the correspondence between each ruled line and each frame tangent may be extracted and stored in the item name-data boundary score tables 1610 and 1620 shown in FIG. These tables 1610, 1620, 1630 and 1640 are stored in the memory 309 and the external storage device 314.

Next, an example of the frame feature difference score calculation process in step S140 will be described using the flowchart of FIG. 4, the example of FIG. 6, and the example of FIG. The process of FIG. 4 is executed for each pair of adjacent frames among the frames extracted in step S110. In step S401 in FIG. 4, the frame feature difference score is initialized to zero. In this embodiment, the frame feature difference score is a value indicating that the larger the value is, the larger the difference between the format feature and the character string feature of the frame pair is. In step S402, it is determined whether there is a remaining frame feature. If there is no remaining frame feature, the process proceeds to S406. The frame features are, for example, 6101 to 6111 in FIG. 6, format information such as frame decoration and size as shown in the following (1) to (11), and the font of character strings belonging to the frame and the presence / absence of emphasis And regular expressions that match strings.
○ Examples of frame format information (1) Position of the text within the frame (left-justified or centered or right-justified)
(2) Difference in frame background color (3) Ratio of frame width (4) Ratio of frame height ○ Examples of format characteristics of character strings in a frame (5) Difference in font size (6) Font type ( Times, Arial, Century,…)
(7) Presence / absence of bold setting (8) Presence / absence of italic setting (9) Presence / absence of underline ○ Example of semantic feature of character string in frame (10) Edit distance of numeric symbol expression of character string (11) Item name score For example, the frame characteristics of the frame 610 in FIG. 6 are center alignment (6101), background color (R, G, B) = (240, 240, 240) (6102), frame width 324 [dot] (6103) ), Frame height 85 [dot] (6104), font size 18 (6105), font type Arial (6106), bold setting (6107), no italic setting (6108), no underline (6109), numerical symbols The expression is not applicable (6110). Numeric symbol expression is a feature that indicates how a numerical value or symbol (comma, period, hyphen, etc.) is masked on the character string in the frame. "Voltage [V]" Then, “----------” (6110) is obtained, and “1,350” is “n, nnn” (n is a number) (6310). For example, “(10) Editing distance of numerical symbol expression of character string” is 2 for “220” and “1,350”, and 0 for “220” and “240”. (Edit distance is an index that represents the distance between character strings, and is defined by the number of substitutions, insertions, and deletions when converting a character string into a character string. "Nnn" → "n, nnn If "", the insertion distance is 2 because the insertion is 2, and if "nnn-nn" → "nn-nnn", the insertion distance is 1 because the insertion is once deleted.
In step S403, the difference score of the feature is calculated for each of two adjacent frames. The difference score is set so that the larger the feature difference, the larger the value. For example, among the features of (1) to (11) above, those having real values are the absolute value of the difference, and those having no real value are constants when there is the difference, and the difference If there is no, there is a method of setting it to 0. Moreover, you may weight according to each importance for every feature of said (1)-(11).
In step S404, the difference score of the currently referenced frame feature is added to the frame feature difference score, and the next frame feature is referenced in step S405. In step S406, the scores obtained as a result of S401 to S405 are written in the frame feature difference score tables 1630 and 1640.
FIG. 7 is a diagram illustrating a calculation example of the frame feature difference score calculation process of S140. A table 7000 is obtained by calculating the frame feature difference score between the frame 710 and the frame 720 for each frame feature. For example, for the background color (7002), the frame width (7003), and the frame height (7004), the difference score is calculated from the ratio or difference between the feature values of the frames 1 and 2. Also, if the characteristics are different, such as the alignment position (7001), font size (7005), font type (7006), bold setting presence / absence (7007), italic setting presence / absence (7008), underline presence / absence (7009), etc. If the difference score is a constant and the features are equal, the difference score is calculated as 0. The numerical symbol expression (7010) defines (edit distance ÷ character string length) of a masked character string as a difference score. By taking the sum of the difference scores for the respective frame features obtained in this way, the frame feature difference score between the frames 710 and 720 is calculated as 4.591.

  Next, in the item name-data boundary detection processing in step S150, the item name of the item name is calculated based on the frame feature difference scores calculated for every pair of adjacent frames in the frame feature difference score calculation processing in step S140. The item name-data boundary line that is the boundary between the frame and the data frame is detected. The item name-data boundary line is detected by calculating the item name-data boundary score. The item name-data boundary score is calculated for all ruled lines in the table, and the score is increased as the probability that the ruled line becomes the item name-data boundary is larger.

  A processing example of the item name-data boundary detection processing in step S150 of FIG. 1 will be described using the flowchart of FIG. 5 and the example of FIG. The processing flow of FIG. 5 is executed for every ruled line in the table. In the case of the example of FIG. 8, the processing flow of FIG. 5 is executed for each of the vertical ruled lines V1, V2, and V3 and the horizontal ruled lines H1, H2, H3, and H4. Here, as an example, a processing example of item name-data boundary detection for the ruled line H2 will be described. In step S501, the ruled line item name-data boundary score is initialized to zero. In step S502, it is determined whether there is a remaining frame tangent in the table. If there is a remaining frame tangent, the process proceeds to step S503. If there is no remaining frame tangent, the process proceeds to step S506. In step S503, the item name-data boundary line score tables 1610 and 1620 in FIG. 17 are referred to, and it is determined whether or not the currently referenced frame tangent belongs to the ruled line, and if it belongs, the process proceeds to step S504. If not, the process proceeds to step S505. In the example of FIG. 8, frame tangent lines 801, 802, 803, and 804 belong to the ruled line H2. In step S504, the frame feature difference score of the frame tangent is added to the item name-data boundary score of the ruled line. The frame feature difference score is acquired with reference to the frame feature difference score tables 1630 and 1640. In step S505, the next frame tangent is referred to. When the reference of all the frame tangents is completed, the process proceeds to step S506. In step S506, the item name-data boundary score of the ruled line calculated in steps S502, S503, S504, and S505 is normalized by the number of frame tangents. In the case of FIG. 8, the number of frame tangent lines belonging to the ruled line H2 is 4. In step S507, it is determined whether the item name-data boundary score is greater than a predetermined threshold value α. If it is larger, the process proceeds to step S508, where the ruled line is determined as an item name-data boundary line, and if it is smaller, the item name-data boundary line detection process is terminated. In the case of FIG. 8, it is detected that the ruled lines H2 and V1 are item name-data boundary lines. Further, the processing of S507 and S508 may compare the item name-data boundary score of all ruled lines without predetermining a threshold in advance, and use the ruled line with the highest score or a predetermined number of ruled lines with the highest score as the boundary line. .

  In the item name frame determination process in step S160, when the item name dictionary 304 exists, the frame to which the character string determined as the item name in the item name dictionary collation process in step S230 belongs can be determined as the item name frame. In the case where the item name dictionary 304 is not used as in the processing flow of the inventive method of FIG. 1, the item name frame can be determined by a different implementation method. For example, in the item name-data boundary detection processing in step S150, there is a method of determining a frame positioned on the ruled line determined to be the item name-data boundary and a frame positioned on the left as the item name frame. In the case of FIG. 8, since the horizontal ruled line H2 and the vertical ruled line V1 are determined to be item name-data boundary lines, the frames located above the horizontal ruled line H2, that is, the character strings “Electric system” “Input” “Voltage [V ] ”,“ Power [W] ”and“ DC / AC ”belong to the item name frame, and the frame located to the left of the vertical ruled line V1, that is, the character strings“ Minimum ”,“ Normal ”and“ Maximum ”belong to it. The frame is determined to be an item name frame.

  Through the above processing, output results as shown in the lower part of FIGS. 14 and 15 can be displayed on the output device 303 from the table structure shown in the upper part of FIGS. 14 and 15.

As described above, the processing flow of the method of the present invention described with reference to FIG. 1 includes the frame feature difference score calculation processing and the item name-data boundary detection processing, so that the item names can be retrieved from the table without using the item name dictionary. -Data relations can be extracted. As a result, the item name-data relationship can be extracted even when the item name word to be extracted is unknown or the item name has a notation fluctuation. In other words, it can be applied to use cases where the item name-data extraction method could not be applied in the past, such as analyzing unspecified and large numbers of tables.
<Second Embodiment>
Hereinafter, application examples of the method of the present invention will be described.

  1 includes the frame feature difference score calculation process in step S140 and the item name-data boundary line detection item name dictionary in step S150, and the feature difference between the frame format information and the character string information in the frame. The item name frame is determined based on. This makes it possible to extract the item name-data relationship even when the item name dictionary is not complete for the item name-data relationship to be extracted. On the other hand, there is a configuration method using a conventional item name dictionary, and in this case, the item name-data relationship can be extracted more stably. This will be described below with reference to the process flow diagram of FIG.

  The processing flow of FIG. 9 is obtained by adding the item name matching process and the item name dictionary 304 in step S930 to the processing flows S110 to S170 of FIG. Frame detection processing in steps S910 and S110, character string detection processing in steps S920 and S120, item name-data boundary detection processing in steps S950 and S150, item name frame determination processing in steps S960 and S160, step S970 The item name-data relationship determination process in step S170 is the same process. Further, the item name matching process in step S930 and step S230 in FIG. 2 is the same process.

The frame feature difference score calculation processing in step S940 is almost the same processing as step S140, but the item name dictionary score 6111 is added to 6101 to 6110 in FIG. 6 as the frame feature. The item name dictionary score is a score indicating how much the word defined in the item name dictionary seems to be an item name. The item name dictionary score is given when a certain character string matches an item name word in the item name dictionary in the item name dictionary collation process in step S930. FIG. 11 shows an example of the item name dictionary. The item name dictionary stores item name words and item name dictionary scores for the item name words. For example, the item name dictionary score may be defined as a numerical value obtained by normalizing the number of appearances of a certain item name with the sum of the number of appearances of all item names. In the frame 710 in FIG. 7, the character string “Voltage [V]” in the frame 710 matches the item name word “Voltage [V]” in the item name dictionary in FIG. In this case, the item name dictionary score 0.24 linked to the item name word “Voltage [V]” becomes the item name dictionary score in the frame 710. On the other hand, the character string “220” in the frame 720 does not match the item name word in the item name dictionary. Therefore, the item name dictionary score in the frame 720 is 0.00. That is, the difference score between the item name dictionary scores of the frames 710 and 720 is 0.24 (7011). By using the difference score of this item name dictionary score to calculate the frame feature difference score, even in a table in which there is no difference between the item name and the data frame feature, it partially matches the item name dictionary If the word and frame to be found are found, the item name-data boundary line can be detected. Even if the item name dictionary is not complete, the item name-data boundary line can be stably detected by the combined use of the item name-data boundary line detection process in step S950.
<Third embodiment>
Next, another application example of the present invention will be described.

  In the processing flow of FIG. 9, when the item name dictionary is not complete for the item name-data relationship to be extracted, or when the table has a format in which there is no difference between the item name and the data frame characteristics. However, the item name-data relationship can be extracted stably. On the other hand, if a word determined to be an item name does not exist in the item name dictionary when a certain table is analyzed, the item name dictionary is newly added to the item name dictionary. Since automatic learning and automatic maintenance of the dictionary are possible, it is possible to perform table analysis using the maintained item name dictionary when analyzing another table. For example, in the example of determining the item name-data relationship in FIG. 8 and the example of the item name dictionary in FIG. 11, if the frame 810, the frame 820, and the frame 830 are determined to be item name frames, the character string “ “Maximum” and the character string “Minimum” in the frame 830 are registered in the item name dictionary 1101, but the character string “Normal” in the frame 820 is not registered. Here, by newly adding the item name word “Normal” to the item name dictionary 1101, it is possible to extract the item name-data relationship from a table including the word “Normal” in another layout.

  The processing flow of FIG. 10 is obtained by adding the item name dictionary correction processing of step S1080 to the processing flows S910 to S970 of FIG. The processing of FIG. 10 makes it possible to automatically correct the item name dictionary based on the analysis result of the table, and the user does not need to maintain the item name dictionary, and the item name dictionary is automatically adapted to the group of documents to be analyzed. It brings me the advantage of being corrected by. 9 and 10, the frame detection process in steps S910 and S1010, the character string detection process in steps S920 and S1020, the item name dictionary collation process in steps S930 and S1030, and the frame feature difference score in steps S940 and S1040. The calculation process, the item name-data boundary detection process in steps S950 and S1050, the item name frame determination process in steps S960 and S1060, and the item name-data relationship determination process in steps S970 and S1070 are the same process.

  In the item name dictionary correction process in step S1080, a new item name word is added to the item name dictionary based on the item name extracted in the item name-data relationship determination process in step S1070, or each item name word is linked. Correct the item name dictionary score.

A processing flow example of the item name dictionary correction processing is shown in FIG. 13, and an item name dictionary correction example is shown in FIG. For explanation, it is assumed that the item names are extracted as “Voltage [V]”, “Power”, “Maximum”, “Normal”, and “Minimum” as in the example of FIG. In step S1301, the extraction result of the item name-data relationship is referred to, and the extracted item name is referred to. Next, in step S1302, it is checked whether there is a remaining extraction result. If there is, the process proceeds to step S1303, and if not, the process proceeds to step S1307. In step S1303, it is determined whether the extracted item name is in the item name dictionary. If there is, the process proceeds to step S1304, and if not, the process proceeds to step S1305. For example, in the extraction example of FIG. 8 and the item name dictionary 1201 before correction of FIG. 12, the extracted item names “Voltage [V]”, “Maximum”, and “Minimum” are in the item name dictionary, but “Power [W ] ”And“ Normal ”. In step S1304, the number of occurrences of the word in the item name dictionary that matches the extracted item name is incremented by one. For example, in the case of FIG. In step S1305, the extracted item name is newly added to the item name dictionary. In the case of the item name dictionary 1201 before correction and the item name dictionary 1202 after correction in FIG. 12, “Power [W]” and “Normal” are newly added. At this time, the number of appearances may be 1 or a certain constant (10 in the example). In step S1306, the next extraction result is referred to. In step S1307, the number of appearances of the item name in the item name dictionary is normalized. For example, normalization may be performed by dividing by the total number of appearances of all item names, or normalization may be performed by dividing by the number of appearances of the item name having the maximum number of appearances.
The item name dictionary is corrected as needed based on the item name-data extraction result by the item name dictionary correction processing in step 1080 described above. Thus, as the table analysis is accumulated, the user does not perform maintenance, such as automatically adding words that were not in the initial item name dictionary, or automatically setting the item name dictionary score of item names with a high number of appearances. In addition, the item name dictionary can be automatically learned and corrected.

331: Table structure analysis computer 301 ... Document input device 303 ... Output device 309 ... Memory 310 ... CPU
311, 312 ... I / F
313: Network I / F
314... External storage device 315... Network H1 to H4... Horizontal ruled lines V1 to V3 constituting table. Vertical ruled lines h1 to h13 constituting table. Horizontal ruled lines H1 to H4. Frame tangent lines that constitute vertical ruled lines V1 to V3

Claims (15)

A table structure analyzing apparatus having a control unit, a storage unit connected to the control unit, and an input unit for inputting image information,
The controller is
From the image information input from the input unit, detect a frame region constituting a table included in the image,
Calculating a frame feature amount indicating the feature amount of the detected frame region;
Calculating a difference between the frame feature amounts between two adjacent frame regions;
Associating the calculated difference between the frame feature values with a line segment between the two adjacent frame regions;
A ruled line serving as a boundary between a frame area in which the item name exists in the table and a frame area in which data exists, based on the difference in the frame feature amounts associated with the line segments constituting the ruled lines constituting the table As a boundary line,
A table structure analysis characterized in that, based on the boundary line, a correspondence between a frame area where an item name exists in the table or a frame area where an item name exists and a frame area where data exists is determined. apparatus. The table structure analyzing apparatus according to claim 1,
The storage unit stores an item name dictionary in which item names are registered,
The table structure analysis apparatus, wherein the control unit registers a character string in a frame area determined as a frame area in which the item name exists as an item name in the item name dictionary. The table structure analyzing apparatus according to claim 1,
The storage unit stores an item name dictionary registered in association with a character string and likelihood information indicating the likelihood of the item name of the character string.
The control unit changes the likelihood information corresponding to a character string in a frame area determined to be a frame area in which the item name exists among the likelihood information registered in the item name dictionary. Table structure analyzer characterized by The table structure analyzing apparatus according to claim 1,
The storage unit stores an item name dictionary registered in association with a character string and likelihood information indicating the likelihood of the item name of the character string.
The controller is
Detecting a character string in the detected frame region;
From the item name dictionary, the likelihood information corresponding to the detected character string is calculated as the frame feature amount,
A table structure analyzing apparatus that calculates a difference in the likelihood information between two adjacent frame regions as a difference in the frame feature amount between the two adjacent frame regions. The table structure analyzing apparatus according to claim 1,
The controller is
Calculating a frame format feature amount indicating a feature amount of a frame format in the detected frame region as the frame feature amount;
A table structure analysis apparatus that calculates a difference between the frame format feature values between two adjacent frame regions as a difference between the frame feature values between the two adjacent frame regions. The table structure analyzing apparatus according to claim 1,
The controller is
Detecting a character string in the detected frame region;
Calculating a character string feature amount indicating the feature amount of the detected character string as a feature amount of the frame;
A table structure analysis apparatus that calculates a difference between the character string feature amounts between two adjacent frame regions as a difference between the frame feature amounts between the two adjacent frame regions. The table structure analyzing apparatus according to claim 1,
The controller is
Calculating a frame feature difference value indicating a difference in the frame feature amount between two adjacent frame regions;
Associating the calculated frame feature difference value with a line segment between the two adjacent frame regions;
Based on the frame feature difference value associated with the line segment constituting the ruled line, calculating boundary line likelihood information indicating the likelihood of the boundary line of the ruled line,
A table structure analysis apparatus, wherein the ruled line is determined to be the boundary line when the boundary line likelihood information of the ruled line is a predetermined value or more. The table structure analyzing apparatus according to claim 1,
The controller is
Calculating a frame feature difference value indicating a difference in the frame feature amount between two adjacent frame regions;
Associating the calculated frame feature difference value with a line segment between the two adjacent frame regions;
Based on the frame feature difference value associated with the line segment constituting the ruled line, calculating boundary line likelihood information indicating the likelihood of the boundary line of the ruled line,
A table structure analyzing apparatus that determines a predetermined number of ruled lines having a large likelihood information among the plurality of ruled lines constituting the table as the boundary lines. A table structure analyzing apparatus having a control unit and an input unit to which image information is input,
The controller is
From the image information input from the input unit, detect a frame region constituting a table included in the image,
Calculating a frame feature amount indicating the feature amount of the detected frame region;
Calculating a difference between the frame feature values between two adjacent frame regions;
A table structure analyzing apparatus characterized by detecting a frame region where an item name exists based on the calculated difference between the frame feature values. A table structure analysis method for analyzing a structure of a table included in electronic information,
From the input image information, detect the frame area constituting the table included in the image,
Calculating a frame feature amount indicating the feature amount of the detected frame region;
Calculating a difference between the frame feature values between two adjacent frame regions;
Associating the calculated difference between the frame feature values with a line segment between the two adjacent frame regions;
A ruled line serving as a boundary between a frame area in which the item name exists in the table and a frame area in which data exists, based on the difference in the frame feature amounts associated with the line segments constituting the ruled lines constituting the table As a boundary line,
A table structure analysis characterized in that, based on the boundary line, a correspondence between a frame area where an item name exists in the table or a frame area where an item name exists and a frame area where data exists is determined. Method. The table structure analysis method according to claim 10,
A table structure analysis method, wherein a character string in a frame area determined as a frame area in which the item name exists is registered as an item name in an item name dictionary in which the item name is registered. The table structure analysis method according to claim 10,
Of the likelihood information registered in the item name dictionary registered in association with a character string and likelihood information indicating the item name likelihood of the character string, it is determined as a frame region where the item name exists. A table structure analysis method, wherein the likelihood information corresponding to the character string in the frame region is changed. The table structure analysis method according to claim 10,
Detecting a character string in the detected frame region;
From the item name dictionary in which a character string and likelihood information indicating the item name likelihood of the character string are associated and registered, the likelihood information corresponding to the detected character string is used as the frame region feature amount. Extract and
A table structure analysis method, wherein a difference in the likelihood information between two adjacent frame regions is calculated as a difference in the frame feature amount between the two adjacent frame regions. The table structure analysis method according to claim 10,
Extracting the frame format feature amount indicating the format feature of the frame in the detected frame region as the frame feature amount;
A table structure analysis method, wherein a difference in the frame format feature value between two adjacent frame regions is calculated as a difference in the frame feature value between the two adjacent frame regions. The table structure analysis method according to claim 10,
Detecting a character string in the detected frame region;
Calculating a character string feature amount indicating the feature amount of the detected character string as a feature amount of the frame region;
A table structure analysis method, wherein a difference between the character string feature values between two adjacent frame regions is calculated as a difference between the frame feature values between the two adjacent frame regions.

Images