JPH0887528A - Document filing system - Google Patents

Abstract

PURPOSE: To store the text data for each document structure by automatically analyzing ana recognizing the document structure in a document filing system which contains a full text searching function. CONSTITUTION: The pattern component which is extracted by an image processing means 322 is analyzed to refer to layout rules of the document structures stored in a document knowledge file 327. Then the character patterns constructing the characters are segmented for each document structure, and the characters are recognized by a character recognition means 331 so that the character strings are obtained. Thus a document recognition means of such a constitution is available. At the same time, a storage means stores the character strings obtained by the document recognition means in response to each document structure.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a document filing system for filing a document as an image, and more particularly to a document filing system characterized by being capable of full text search (text search).

[0002]

2. Description of the Related Art Conventional information retrieval systems have mainly provided retrieval means in accordance with keywords and classification codes. Literature information and patent information have been stored in databases so far using the above-mentioned method. Here, the bibliographic information including the abstracts is mainly stored in the database, and only a part of the functions can be provided for the true information retrieval needs.
In other words, even if a document or patent that seems to be related is found, it is necessary to search through a large number of bookshelves in order to obtain the text.

On the other hand, an optical disk capable of storing a large amount of data has appeared, and it has been highlighted as a social need to store a text in a database and provide a so-called original document information service. The paperless plan at the JPO follows this trend. In these systems, a large amount of documents are stored in the form of image data on an optical disc, and a conventional keyword-based information retrieval technique is applied.

However, in the above-mentioned conventional information retrieval technology, it is possible to narrow down only several tens to several hundreds of orders, and there is a demand for a method of narrowing down related documents to about 1/10. One method is to call an original document (text) stored as image data on a terminal and read it by a searcher. This method is reliable in principle, but reading a maximum of several hundred documents in the form of image data requires a large amount of data, and it is not efficient to read each document visually, which makes it practical. Is a problem.

On the other hand, the conventional method using a keyword or a classification code has an essential problem because the classification system itself needs to be constantly updated because it changes with time.
For example, it is practically impossible to change a large amount of documents that have already been classified, because it becomes necessary to change the classification system later. Documents and patents that record the progress of science and technology are new and valuable only if the concept that does not fit into the conventional classification system is presented. In this sense, it is impossible to define in advance a keyword or a classification system that originally represents a concept, and it is an essential subject as an information retrieval system.

For the above reasons, there is a demand for a method of directly referring to the text of a document to retrieve the content. According to the method of referring to the text, it is possible to search for a concept that was not considered important when registering a document in a database, and using a vocabulary recognized as a new concept at the time of searching. Become. Alternatively, it becomes possible to directly search for an important document without using the "filter" called an indexer (a dedicated person who gives an index) when registering.

In order to satisfy such a demand, it is necessary to extract a character pattern from a document as image data and replace the body with a character code. For this purpose, a character recognition technique may be applied. . However, even if the document to be subjected to document filing is a printed document, it is not possible to expect complete character recognition by the conventional character recognition technology because of the print quality and the variety of type (font) types. difficult. In a conventional character reading device, an incomplete recognition such as erroneous recognition or unrecognizable (rejection) is checked and corrected by an operator (for example, "Introduction to Character Recognition" by Hashimoto, Ohmsha, 1982, pp.15
3-154). Therefore, even if the recognition accuracy is extremely high, if the amount of documents is enormous, it is not realistic for a human to check the result of recognizing the text, and image-based document filing that enables text search The system has not been realized to date.

[0008]

SUMMARY OF THE INVENTION An object of the present invention is to provide a document filing system having a full-text search function for searching by directly referring to the text of a document by solving the above problems. To do.

[0009]

In order to achieve the above object, the present invention provides an image file for storing document images, a document knowledge file for storing layout rules of a document structure for each document type, and a document. An image processing unit for extracting a pattern component from an image and a character recognition unit for recognizing a character pattern cut out from a document image are provided, and the layout rule of the document structure stored in the document knowledge file is referred to. A document recognition means for analyzing a pattern component extracted by the image processing means to cut out a character pattern forming a character for each document structure, and recognizing the cut out character pattern by a character recognition means to obtain a character string, and a document A storage means for storing the character string obtained by the recognition means in association with the document structure, and a search for the storage means in response to a search request And having search means for identifying documents that satisfy the determined, and output means for outputting the document image of the document identified by the search means from an image file.

[0010]

That is, the document filing system according to the present invention improves the disadvantages of simultaneously treating a document or the like as an image without deteriorating the advantage of treating a document or the like as an image. That is, in the filing system that handles images, conventionally, it has been the main to search mainly based on keywords and bibliographical items added separately, but according to the present invention, it is possible to refer to the sentences written therein. You can search.

For example, by inputting "Honbunkensaku" from the search terminal, for example, "... text search by character recognition ..." Is written in the text of a document in the document group to be searched. If there is a document, the document can be identified and extracted, and the document can be displayed as an image on the terminal.

By displaying as an image, it is possible to avoid loss of information due to character recognition. Generally, in character recognition, secondary information such as the position, size, and font of each character is discarded in the process of normalization. Therefore, it is not known after recognition whether the size is Gothic or Mincho, and it is meaningless to print in Gothic or a large font to show the importance. Speaking of voice,
If you do voice recognition, you will not be able to understand who spoke and the emotions at that time.
Even in the case of a document, such secondary information is important to the reading person, and simply recognizing the character is not a good idea.

As described above, the first principle of the system of the present invention is that the document is stored as an image while the character portion is stored as a character code in an overlapping manner.

Now, in order to extract a character portion from an image and replace it with a character code, it is necessary to perform character cutting and character recognition. Although a conventional technique can be used for this, a 100% recognition rate cannot be expected.

The second principle of the system of the present invention is that with respect to a character that cannot be determined as a result of character recognition, the character categories remaining in the higher order are treated as a set and are left as they are in the recognition result character string. is there.

For example, when "... body search by character recognition ..." is recognized, the recognition result of this system is "... sentence [character study] [honki] sentence search by [text] recognition [knowledge] ...". Represents Here, the characters enclosed in [] are the recognition results for a certain character pattern, and “[knowledge]” means that “knowledge” is either “weave”. Conventionally, a character that cannot be determined is always replaced by a correct character code due to operator intervention and used as a character recognition result (OCR output). Here, the symbols "[" and "]" are special symbols, and generally, a code that does not appear in the text is assigned. For the sake of simplicity, the symbols [,] shall be used for easy understanding.

In the system using the present invention, the document 10 is converted into a symbolic expression as shown by 20 as shown in FIG. The symbol string follows a notation called an S-expression used in the LISP language and the like. Document (image) 10 symbolic expression 20
The process of converting to is called document understanding or document recognition.
The symbolic expression has the following meanings. That is, it is document # 99, its class is “paper”, VOL = 5, NO = 7, and title is “sentence [literary].
The author's name is "Yamada [tadai] rou", the text is "... full text by automatic character reading [lili], [...] -..." Where [riri] is hiragana and katakana, [tou] is katakana and kanji,
[-One-] means katakana long sound, Chinese numeral 1, and minus sign]. Among the ambiguous characters in character recognition, there are many character patterns that cannot be usually dealt with, as in the above example.

In the search, the user inputs "honbunkensaku" in Roman letters or katakana. The system converts this into Kana-Kanji. In general, there are homonyms, in which case "honbun" is either "text" or "main text" and "kensaku" is either "search" or "dedication." This method can handle such ambiguity automatically.

Similarly, when "Mojiyomitori" is entered, there is ambiguity (more than one possibility) in the sending kana.
There are "character reading", "character reading", and "character reading", and it is not possible to know what kind of futuristic kana is attached to the unknown text, so it is necessary to handle all possibilities in principle.

Furthermore, when "Mojininshinki" is entered, "character recognition" is uniquely obtained by kana-kanji conversion, but "character recognition" is sometimes referred to as "character reading". It is also desired to automatically select "reading characters" as a synonym as a search key. In this case, similar to the above example, the possibilities of plural sending kana are also listed. Here, the asymmetry that "character recognition" does not occur in the synonym of "character reading" even if "character reading" comes up in the synonym of "character recognition", is also required in this method. Is full.

After all, a plurality of partial character strings to be found in the sentence to be searched are expressed as a finite state automaton as shown in FIG. On the other hand, the character string of the searched sentence shown in the example of FIG. 1 is similarly expressed by the automaton of FIG. The present invention provides a text search function in the case where there is ambiguity (a plurality of possibilities; a situation in which there is an element that cannot be uniquely determined) in both the search key (substring) and the searched sentence. It is the third principle.

As a method for finding a plurality of substrings from unambiguous text by using their finite state automata, there is a method described in [A. V. Aho, et al. “Efficient
String Matching: An Aid to Bibliographic Se
arch, "Communications of the ACM, Vol.18,
No. 6, 1975] is known.

[0023]

EXAMPLES The present invention will be described below based on examples.

FIG. 4 is a block diagram of a document filing system which is an embodiment of the present invention. The system includes a control subsystem 100 that provides control of the entire system and a database function, an input subsystem 200 that inputs documents and registers them in a file, a document recognition device 300 that recognizes documents, and a high-speed It is composed of a text search subsystem 400 for performing a text search and a terminal subsystem 800 for performing a search.

The configuration and operation flow of each subsystem will be described in detail below.

The input subsystem 200 has a CPU (central processing unit) 201 for controlling the subsystem, a main memory 202, a system file 251, and a terminal 203 as basic parts. The subsystem is controlled by an operation from the terminal 203, and the image of each page of the document 220 is scanned by the scanner 22.
First, the image data optically read by 1 and digitized is stored in the video memory 224 via the bus 210. Next, the image data is subjected to redundancy compression by an image processing device (IP) 223 to obtain MH (Modified Huffman).
It is converted into an n) code or an MR (Modified Read) code and is returned to another area of the video memory 224 again.

The input document image is displayed on the terminal 203 for confirmation, and at the same time, the operator can input bibliographic items while looking at the displayed image.
As will be described later, the bibliographic items of the fixed form document can be automatically read by understanding the document, but the bibliographic items of the irregular form document and information not entered on the paper must be input by a human. For example, it is natural that the operator has to rely on the classification code of the document content defined by the user and the input of the keyword not on the paper. Further, the value and position of each document must be uniquely attached by the user of the document, and these can also be input from the terminal 203. The input data such as the bibliographic item is stored in the main memory 202 in association with the image data (compressed data) in the video memory 224.

Here, a unique number (document ID) is attached to each document, and the image data and bibliographic items are stored in the memory so that the document unique number can be used as a key. The document unique number is, for example, the subsystem ID
It can be represented by a concatenation of a character string representing "date / time" (such as "INSYS 01"). For example INSYS0
1.850501.132437 is May 1, 1985
Input subsystem INSYS at 13:24:37
Indicates that the document is input from 01. The input time may be important depending on the application of the system, and it also functions as a time stamp.

Now, a predetermined amount of documents is stored in the subsystem 20.
When a certain amount is accumulated in 0 or a predetermined command is issued from the terminal 203, an interrupt signal is sent to the bus adapter 171.

The control subsystem 100 senses the interrupt signal to read a predetermined address in the memory 202 of the input subsystem 200. This makes it possible to judge the content of the request from the input subsystem.

In the case of a request for registration of the input document in the database, the operation is as follows.

The central processing unit (CPU) 101 knows the unique numbers of the documents (plurality) temporarily stored in the input subsystem according to a predetermined program in the main memory 102, and further bibliographical data (bibliographic items) related to them. ) And the storage address of the image data.

The control subsystem 100 is a database file 1 for storing and managing symbol data such as bibliographic data.
51 and an image file 152 for storing and managing image data.

The bibliographic data read out from the input subsystem 200 is written as a new record in the tabular database (stored in the file 151) shown in FIG. The above table has the name MAIN-DIR (main directory) and has the following columns (data columns).

DOC #: serial number for a registered document in this system ID #: document unique number given by the input subsystem NP: number of pages constituting the document TITLE: title (character string) AUTHOR : Author name (repetition, that is, multiple data is allowed.)-CLASS: Code indicating the classification, type, etc. of the document-PUBL #: In-system registration number of the publication (details are managed in the table shown in FIG. 7)- VOL, NO, PP: Volume, Issue, Page ・ KWD: Multiple keywords ・ ABS: Text unique number of abstract represented as character code string (text data) ・ TXT: Unique number of text as character code string IMG: Unique number of image data. Since each image data is managed for each page, a plurality of image unique numbers are recorded.

When registering bibliographic data, among the above columns,
Only part of the data related to the bibliographic data is newly written.

Next, the images of the pages forming each document are transferred from the predetermined storage area of the input subsystem to the control subsystem 10.
It is read out to 0 and sequentially stored in the empty area of the image file 152. At the same time, an image unique number (IMGID) is assigned to each image (page unit). In addition, the volume number (VOLSER) of the file storing the image data, the file device number (UNIT), the storage physical address (PHYSA) of the file, the length of the storage area occupied by the file (SLENG), etc. are shown in FIG. Fill in the table as shown in (b) and FIG. The image unique number IMGID newly assigned is the IM of the table MAIN-DIR (FIG. 5).
It is also recorded in the G column.

Here, the table IMG-LO shown in FIG.
C is particularly effective when the image file 152 is composed of a plurality of drive devices or a plurality of volumes, and manages the location of each image. Naturally, it is updated every time the operator unmounts or mounts the volume.

FIG. 8 shows a directory provided for each volume of the image file 152 and has the following columns.

IMGID: image unique number PN: page serial number in document (1 to n) PHYSA: physical address in volume SLENG: recording length (for example, number of sectors) CODE: image compression code name SIZE: Image size (number of pixels) DOC #: Document serial number. Further, in the figure, the data in the column PHYSA of the record 157 indicates the start address of the image data 158 in the image data area 156 in the image file.

When the above operation is completed, the present system searches the terminal group 800 for bibliographic items and keywords.
You can start from.

The search condition input from the search terminal is passed through the gateway 175 to the C of the control subsystem 100.
It is transferred to the PU 101. The table MAIN-DIR 153 (FIG. 5) in the database file 151 is searched according to a predetermined search processing program in the memory 102. It goes without saying that indexing (means for speeding up search such as hashing and reverse file) is applied to the main columns of Table 153.

As a result of the search processing, table 153 (FIG. 5)
To DOC # list and image unique number IMGID list are created and stored in a predetermined area of the memory 102.
When a display request is issued from the search terminal, the table IMG-LOC1
54 (Fig. 6 (b)) and Table IMG-DIR155 (Fig. 8).
Is used to identify the position in the image file, and the image data is sequentially read out onto the memory 102. At the same time, the read image data is transferred to the search terminal in order,
It is displayed on the screen according to the instructions on the terminal.

Next, a method of managing texts used for body text search will be described.

As described in the main directory MAIN-DIR (FIG. 5), each document stores and manages not only image data but also text represented by a character code string. In the case of the present embodiment, the abstract and the text are stored as text in the text files 451, 452, 453, respectively.
Managed. A unique text number is assigned to each text (character string), and the ABS column and the TXT column of the table 153 (FIG. 5), the TXTID column of the TXT-LOC table shown in FIG. 6A, and the TEXT-DIR shown in FIG. Recorded in the TXTID column of the table.

FIG. 9 shows the text files 451, 452.
A method of text storage and management at 453 is shown. In the figure, the text body is one-dimensionally stored in the file storage area 466. A unique number TXTID is assigned to each text (one character string), the directory table, TEX
It is managed by T-DIR465. Table 465 has the following columns.

-TXTID: Text unique number-NCH: Total number of characters forming the text-PHYSA: Physical address at which the text is recorded-SLENG: Length of recording of the text on the storage medium-CCLASS : Class of characters expressing the text (Japanese characters mixed with kanji, English, Roman characters, kana characters, etc.) The record 467 of the table 465 indicates that the text represented by the record is 468 part in the storage area in the same file. It means that

On the other hand, as shown in FIG. 4, it is possible to record text in a plurality of volumes, and the text directory manages the text in each volume. When mounting multiple volumes, it is necessary to know in which volume a certain text exists, but the TXT-LOC table shown in FIG. 6A manages the location of each text. Volume serial number VOLSER in which the text with the text unique number TXTID is recorded
And the file device number UNIT on which the same volume is mounted is managed. Of course, when the physical volume is unmounted or newly mounted by the operator, TXT-LOC is automatically updated.

As a large flow of operations, when the document image input, the bibliographical item input, and the document registration are completed, the document recognition apparatus 300 performs the body recognition (document understanding) of the registered document. The input of the recognition device is the document image 10 in the image file 152 as shown in FIG. 1, and the recognition result output is the symbolic expression 20 as shown in FIG. The abstract and the text portion of the text in the symbolic expression 20 are newly stored and managed in the text files 451 to 453 as described above. Document recognition will be described with reference to the detailed block diagram of the document recognition apparatus shown in FIG.

The recognition device 300 includes the control subsystem 10
It is connected to the 0 bus 110 via the bus adapter 371 and is controlled by the CPU 301. The memory 302 stores data such as programs and parameters for controlling the operation of the device.

The image data to be recognized is transferred from the image file 152 to the memory 321. The image data is compression-encoded, decoded into a bit representation image by the image processing circuit IP322, and stored again in the memory 321. Then, the IP 322 extracts the contour of the pattern from the image converted into the bit representation, and stores the extraction result in the memory 321 again.

The extracted contour data is expressed as follows.

[0053]

(1) (i Ci x _max , _i x _min , _i y _max , _i y _min , _i x _si y _si (Θ _1i L _1i ) ... (θ _ni L _ni )) (1) Here, i is the unique number (1, 2, 3, ...) Of the contour, and Ci represents the class of the contour. Ci = 0 represents the outer contour (solid line 1001 in FIG. 11), and Ci = 1 represents the inner contour (broken line 1002 in FIG. 11). x _max , x _min , y
_As shown in FIG. 11, _max and y _min represent the coordinates of the vertices of the outline quadrangle. (X _s , y _s ) is the coordinate of one point Ps having the contour length (for example, the point first found in the contour search). The contour data itself is based on the point Ps as the base point in FIG.
As shown in FIG. 2, it is represented by a set of columns of a quantized direction code θ and the number of pixels L in which the same direction continues.

Next, the inclination correction circuit 323 detects the inclination angle generated at the time of document input from the contour data represented by the equation (1), corrects the contour data, and writes it back to the memory 321. As the inclination correction algorithm, for example, the method disclosed in Japanese Patent Application No. 60-152210 can be used.

Of the contour data which has been subjected to the inclination correction, particularly the data portion (x _max , x _min , y _max ,
y _min ), row cutting and column cutting are then performed by the bottom-up segmenter (BSG) 324.

Bottom-up segmenter BSG is number (1)
The data represented by the formula (2) is input, the pattern list represented by the equation (2) is generated and stored in the memory 321.

[0057]

(2) (j x _max , _j x _min , _j y _max , _j y _min , _j ) (2) Here, j is a pattern unique number, and the patterns are defined as rectangular areas that do not overlap each other. 2) further defines the vertex coordinates of the rectangular area. For example, in FIG.
Although rectangular areas 1008 and 1009 indicated by broken lines are inputs to the BSG, a rectangle 1010 is obtained as a result of the BSG. Rectangles 1008 and 1009 are components (elements) each formed from one contour, and rectangle 1010 is a pattern forming one character. The component forming the pattern j can be obtained by searching the rectangle included in the rectangular area defined by the equation (2) from the contour data of the equation (1). Alternatively, it may be separately obtained and stored as data. FIG. 14 schematically shows the result of the row cutout process, and FIG. 15 schematically shows the result of the column cutout process.

The character slicing unit (CSG) 325 refers to the document knowledge that summarizes the rules such as the document format,
Patterns that form characters are extracted from the pattern list. The document knowledge is stored in the document knowledge file (DKF) 327 as shown in FIG.

In the document knowledge file, for each type of document,
The title, author name, author affiliation, abstract, structural rules of layout (arrangement) such as text are stored together with parametric knowledge such as font size. These knowledges are described by the format description language. As the format description language, the language disclosed in Japanese Patent Application No. 60-122425 can be used.

In the character slicing section CSG, although one character is originally formed, it is divided into two or more patterns, but conversely, two or more characters are fused into one pattern by contact. It also performs a process of forcibly separating what has been lost.

The character slicing section CSG, as a processing result, for each item such as a title, an abstract, or a text,
The numbers of the patterns that make up each character are output as a list. For example,

[0062]

(Abstract (j ₁ j ₂ j ₃ ... (j _n j _{n + 1} j _{n + 2} ) ... j _N )) (3) is composed of a character string whose abstract is represented by the pattern number j _k It means to be done. Here, (j _n j _{n + 1} j _{n + 2} ) indicates that the character is composed of the three patterns j _n , j _{n + 1} and j _{n + 2} .

The character recognition unit (CRG) 331 uses the pattern list (for example, the number (3)) and the contour data (represented by the number (1)) on the memory 321 to extract the contour data forming each character pattern. The data is extracted as described above and converted into a data structure capable of feature extraction.

Since a known technique can be used as the character recognition method, detailed description thereof will be omitted. However, after the feature extraction from the contour data, the standard pattern file 333 is extracted.
Each character can be recognized by performing pattern matching with the standard pattern in. In FIG. 10, the memory STP
M334 is for storing a standard pattern having a high reference frequency, and is intended for high-speed processing.

As described above, the result of character recognition is shown in FIG.
The symbol 20 as shown in FIG. When the similarity obtained as a result of the pattern matching satisfies the expression (4) in the final determination process in the character recognition, the character category (character code) ω _k that gives the similarity is output.

[0066]

(4) ρ _k ≧ ρ _l min (ρ _k −ρ _l ) ≧ ε (where k ≠ l) for l = 1, 2, ..., K (4) where ρ _k is the similarity to the character category k, K is the total number of categories, ε is a relative threshold.

If the number (4) is not satisfied, the set of character categories satisfying the number (5) {ω _k | k = k ₁ ,
k ₂ , ...} is sandwiched between two special character codes and output. For example, the character (code) string ω _s ω _k1 ω _k2 ... ω _e is output. Here, ω _s represents “[” and ω _e represents “]”.

[0068]

(5) ρ _k ≧ ρ _l for l = 1, 2, ..., K ρ _k −ρ _ki ≦ ε ₁ k _i {1, 2, 3, ..., K} (5) By the above processing, the similar character exists and the number (4)
If is not satisfied, for example, a recognition result of "full text search" is obtained for the input pattern of "full text search". The recognition result is buffered in the memory 321, and then transferred to the memory 102 (FIG. 4) collectively.

In the control subsystem 100, the table TXT-
The maximum text unique number is detected with reference to LOC (FIG. 6), and the value obtained by adding the value 1 is used as a new text unique number, and the character code string (text) of the recognition result is registered.
The registration process is performed in the main directory 153, table TXT-L.
This is performed for the OC and the table 465 (FIG. 9), and the text data itself is stored in any of the text files 451 to 453.

As described above, it is possible to search the document to which the text data is given by using the text searcher subsystem 400.

Next, the text search subsystem 400 for searching the contents of the text and its operation will be described in detail.

A request for body content search issued from the terminal 800, for example, "ABS = * modin *" is first transferred to the control subsystem 100. Subsystem 100
Then, if the document to be searched has already been narrowed down by a keyword search or the like, the unique number of the text attached to the document is selected from the main directory 153, and further, by referring to the table TXT-LOC, the text For each file, the list number (6) of the searched text unique number is created.

[0073]

(U _i v _i (t _i1 t _i2 ... t _in )) i = 1, 2, ..., M (6) where u _i is the i-th file device number and v _i is the volume. The serial number, t _ik, is the text unique number of the kth text to be searched on the volume. M is the maximum number of text file devices.

On the other hand, if the searched document is the entire document,
Special symbols (eg number (7)) are sent for all text files.

[0075]

(U _i v _i *) i = 1, 2, ..., M (7) The list number (6) or the number (7) and the partial character string (for example, “modininski”) are control subsystems. 100 is transferred to the memory 402 in the text search subsystem 400 via the bus adapter 172.

The subsystem 400 (FIG. 4) performs kana-kanji conversion, different notation generation processing, synonym processing, etc. of the transferred partial character string according to a predetermined program in the memory 402. The kana-kanji conversion dictionary, different notation generation rule, and synonym dictionary are stored in the file 403.

By converting the kana into kanji, "character recognition" can be obtained from "modininshiki". Further "character reading" can be obtained by referring to the synonym dictionary. Applying the variant notation rule to these results would result in "character reading".
From this, different notations "character reading" and "character reading" with different sending kana can be obtained. Known techniques can be used for kana-kanji conversion and synonym generation.

The different notation generation rule is for handling the variety of old fonts such as sending kana and personal names, and is expressed by the following rewriting rule.

[0079]

[Equation 8] (R1) X only Y → XY only | XY (R2) X only Y → XY (XY) (R3) X only Y → XY R | XY (R4) X only Y → XY only XY :: (R101) X-Yru → XY-R (R102) X-Yu → XY-R (R103) X-Y-YM → XY-Mu: (R201) X becomes → X-Ru: (R501) Fujisawa → Fujisawa : (8) Here, X and Y are arbitrary kanji, and “|” means juxtaposition. Further, regarding the occurrence of different notation, see, for example, JP-A-6
The display method can be taken from 0-150176.

The different notation generation process is performed by inputting the number (8) in the input character string.
It is determined whether or not the left side of the rule is applicable, and if there is, the right side of the rule is generated. However, the fitted Chinese characters are inserted in the variables X and Y.

By the above processing, a set of character strings (character recognition, character reading, character reading, character reading) is finally obtained for "Mojininshinki". This will be represented by the number (9).

[0082]

(A ₁ ... A _i ... A _n ) = ((a ₁₁ a ₁₂ ... a _1m1 ): (a _i1 a _i2 ... a _imi ): (a _n1 a _n2 ... a _nmn ) (9) where Where n is the number of character strings, m _i is the length of the i-th character string, and a _ij is the j-th character code from the beginning of the i-th character string A _i .

The subsystem 400 further converts the number of character string sets (9) into the number of state transition lists (10) representing the finite state automaton described with reference to FIG. 2 by a predetermined program.

[0084]

Alist = ((S _j1 C _k1 S _l1 ): (S _ji C _ki S _li ): (S _jm C _km S _lm )) (10) Here, each element of the list alist number (10) Is the state S _ji
In, when the character C _ki is input (matches),
The state means that it can transit to S _li . In the same equation, {S _j1 , ..., S _ji , ..., S _jm } include the same ones.

Further, the number of output lists (11) is generated.

[0086]

Σlist = ((S _j1 A _j1 ): (S _jp A _ip ): (S _jn A _in )) (11) where (S _jp A _ip ) is the time point when the state S _jp is reached. Means that the character string A _ip has been found. It generally corresponds to what is called an output function in an automaton. FIG.
6, the number of character string sets (11) to the number of state transition lists (1
0) and a PAD diagram (Program Analysis Diagram) of an algorithm for deriving the number of output lists (11).

Next, the failure transition list number (12) is created from the state transition list number (10).

[0088]

Flist = ((S ₀ S _j0 ) ... (M _m S _jm )) (12) The element (S _m S _jm ) of flist corresponds to the character C _k input in the state S _m. The state to be transited is the number of alist (1
If it is not specified in 0), it refers to flist and specifies to transit to the state S _jm . It is generally called a failure function.

The purpose of providing flist is that in the partial character string matching, if the matching is successful halfway through a character string but the next character does not match, that is, if a predetermined state transition destination is not found, the initial state S ₀ This is to deal with the fact that returning the state to is generally incorrect. For example, assume searching for two substrings {character recognition, optical character reader}. Now
If the sentence "... Optical character recognition ..." is input, the part up to "optical character" matches the second partial character string, but the next character "recognition" does not match. Here, if the state is reset to S ₀ and reset, the automaton will use the sentence after “recognition ...” as the input character, and eventually miss the partial character string “character recognition”. It will be. Therefore, when the matching fails, the state to be transited is not S ₀ , but the state up to “letter” in the transition path of “letter recognition” needs to be in the state of being matched.

Next, the subsystem 400 causes the state transition list alist and the output list σ created as described above.
The list and the failure transition list flist are transferred to the lower flexible string matching circuits FSM501 to 503.

Flexible string matching circuit 5
A more detailed block diagram of 01 is shown in FIG. (FMS
The same applies to 502 and 503. The three types of lists alist, σlist, and flist are stored in a predetermined area of the memory 513 via the bus adapter 571. The microprocessor 511 generates an extended finite state automaton shown in FIG. 18B in the form of a state transition matrix based on the above information by a predetermined microprogram.

The finite state automaton that the lists alist and flist directly mean has a simple form shown in FIG. 18 (a). This figure is in alist

[0093]

[Equation 13] (S _j C _k1 S _l1 )} (13) Two transitions of (S _j C _k2 S _l2 ) are illustrated.

The microprocessor 511 is shown in FIG.
The finite state automaton shown by is extended and transformed as shown in FIG.
It This conversion is a conversion that is uniquely determined. With this conversion
A certain part of the searched text with ambiguity
It is possible to find a character string. Here, in the figure
And f (S_j) Is created from the failure transition list flist
Failure function, state S_jWhen matching fails with
Represents the state of the transition destination of. Also, state W_jIs state S_jOne
It has a one-to-one correspondence and is an ambiguous character string (symbol []
(Character string enclosed in) is being scanned. Change
Also, state T _j1, T_j2Is state S_jCorresponding to the transition from
State W_jIt is a state derived from
The character being searched for (in the case of the figure, C_K1Or C_K2)To look at the
It is in a state of being put out.

In reality, the microprocessor 511 has two
The state transition table shown in FIG. 19A can be directly generated from the two lists alist and flist. The column (vertical) of the state transition table represents the current state, and the row (horizontal) corresponds to the character (code) input in the same state. In the table, the state to be changed next is described. Since the algorithm for generating the same state transition table can be easily inferred from the description with reference to FIG. 18, the description will be omitted.

The microprocessor 511 further converts the output list σlist into the form of the output table shown in FIG. 19B and records it in a predetermined area of the memory 513 together with the above state transition table.

The string search algorithm using the finite state automaton will be described below.

[0098] Here, the function (c, S) is a function for obtaining the next state from the state transition table shown in FIG. 19A based on the character c and the current state S. The function out (S) is shown in FIG.
It is a function that determines whether or not there is an output in the state S by referring to the output table shown in FIG.

In the above description, the state is assigned to the unit of 1-character code, but if the 1-character code has 2 bytes, as in Japanese, divide it into 1-byte units.
The above method can be applied.

Next, the text search subsystem 400
Accepts the searched text unique number lists (6) and (7) sent from the upper layer, and transfers them to the corresponding FSMs as the text unique number lists to be searched by each FSM. Therefore, if a search target exists in the corresponding text file, each FSM has its unique number (t _i1 t _i2
t _i3 ...... t _in ) is obtained. The text unique number list is stored in the memory 513 (FIG. 17). The microprocessor MPU 511 first detects the physical address of each text according to a predetermined program (see FIG. 20) in the micro program memory 512. The text unique number and physical address are managed by the TEXT-DIR described in FIG. 9, and the table can be read from the file 451 and detected.

The microprocessor 511 next reads each text data from the file 451. The file control unit 531 sequentially inputs the read text data (character string) to the FIFO (First-in-first-out) circuit 532. The microprocessor MPU511 is a FIFO532
Each character is read out from and the presence or absence of a predetermined partial character string is verified according to the finite state automaton (FIG. 18B) defined in the memory 513. The string matching result blist (see FIG. 20) is returned to the memory 402 of the upper processor.

According to a predetermined program, the CPU 401 compiles the text unique number lists returned from the plurality of lower FSMs and having the matching search conditions, and transfers the list to the memory 102 in the upper control subsystem. From the text unique number, the main directory 153 (Fig. 5)
By referring to, the unique number DOC # of the document in which the partial character string is matched or the unique number IMGID of the document image
Alternatively, the title TITLE or the like can be identified.

These search results are returned to the terminal 800. The user can call and display an image of a desired document on the CRT while viewing the title and the like on the CRT.

Next, a second embodiment will be described. In this embodiment, the flexible string matching circuit 501
Only the configuration method is different. FIG. 21 is a configuration diagram of the flexible string matching circuit FSM in the second embodiment.

In the figure, the secondary storage device (text file) 461 has a plurality of heads capable of reading signals simultaneously. In this embodiment, data can be read from four heads at the same time. is there. The data is transferred to each of the four FIFO circuits 551 to 554 via the file control unit FCU541.

On the other hand, the search condition sent from the upper subsystem 400 is translated by the microprocessor 511 and then transferred to the microprocessor units MPU ₁ to MPU ₄ 561 to 564 which include the data memory.

The text data read from the text file 461 is read to the microprocessor units 561 to 564 via the FIFO circuits 551 to 554, respectively. In parallel, the microprocessor unit searches for a predetermined partial character string from the four character strings (text data), and returns the result to the microprocessor 511 via the data bus 521.

Since the other parts are the same as those in the first embodiment, the description thereof will be omitted.

Next, the third embodiment will be described. In this embodiment, the hardware configuration is the same as in the first or second embodiment, but the text search processing is different.

Considering the case where the search target documents are narrowed down by using the keyword or the classification code by using the hierarchical search method, the search target documents screened in the same process are unevenly distributed in the volume of a certain text file. Generally there is.

In the system of this embodiment, text data is stored in duplicate in a plurality of text file volumes in order to enable multiplicity. According to a predetermined program, the CPU 401 (see FIG. 4) selects a volume to be accessed so that the number of times of access to the plurality of volumes is equalized for the texts stored in the plurality of volumes in a multiplexed manner. By using this method, all flexible string matching circuits operate efficiently, and high-speed search is possible as a whole.

In the above embodiments, the flexible string search circuit has a multiplicity of 3 to 4, but the multiplicity of the present invention is not limited.

Although the text search is performed uniformly over the entire document, information about page boundaries is recorded in the text with special symbols so that the page numbers for which string matching has succeeded can be matched. It can be expanded to output as a result, and the same method is also included in the present invention.

Furthermore, although the explanation has been given for the Japanese text, it can be applied to other languages such as English exactly.

Further, although the text data is extracted by the character recognition in the above-mentioned embodiment, the text content search method can be applied to the text data which is obviously input manually and is included in the present invention. .

Further, the system form has been described with reference to the form shown in FIG. 4, but the essentials of the small system and the stand-alone system are not changed, and the present invention includes them. In particular, it is conceivable that a text file and an image file prepared by another system may be loaded into a small-scale search station, which is included in the present invention.

Further, it goes without saying that the search conditions can be combined by a logical operator and can be expanded so that a partial character string satisfying a certain relative positional relationship can be searched for. In particular, by outputting where each of the plurality of sub-character strings exists, post-processing enables high-speed combinatorial retrieval.

[0118]

As described above, according to the system of the present invention, it is possible to search for a desired document at high speed by referring to the contents of the document or the like, and even from the concept that was unthinkable when the document was registered. It is possible to search efficiently.
In particular, when registering, there is no need to worry too much about what is appropriate as a classification code or keyword. As a result, it is possible to improve the search accuracy and at the same time keep the noise generation rate low.

Furthermore, by parallelizing the inside of the text search subsystem, high-speed text search can be performed.
In particular, the speedup is achieved by adding the string matching circuit to each read head.

In the case of a search for a large-scale document file, the text contents can be searched after reducing the number of documents to be searched by using keywords or bibliographic items, and the entire search can be performed efficiently.

Further, in the prior art, in order to obtain text data from a document image, it was necessary for a human to sequentially inspect the document recognition result and correct errors, but according to the present invention, human intervention can be eliminated. Is. Conventionally, the text content search is not substantially realized for the above reason, and the present invention enables effective text content search.

[Brief description of drawings]

FIG. 1 is a diagram showing a document image and a result of document understanding.

FIG. 2 is a state transition diagram of a homonym synonym and a synonym character string generated from a partial character string.

FIG. 3 is a state transition diagram of a character string resulting from character recognition including ambiguity.

FIG. 4 is a system configuration diagram of the first embodiment.

FIG. 5 is a diagram illustrating a method of storing and managing documents, images, and texts.

FIG. 6 is a diagram illustrating a method of storing and managing documents, images, and text.

FIG. 7 is a diagram illustrating a method of storing and managing documents, images, and texts.

FIG. 8 is a diagram illustrating a method of storing and managing documents, images, and texts.

FIG. 9 is a diagram illustrating a method of storing and managing documents, images and texts.

FIG. 10 is a block diagram of a document recognition device.

FIG. 11 is an explanatory diagram of a rectangular area surrounding a character pattern.

FIG. 12 is a diagram illustrating a method of expressing a contour shape that describes a pattern.

FIG. 13 is a diagram illustrating a relationship between a pattern component and a character pattern.

FIG. 14 is a diagram showing a result of line segmentation by a bottom-up segmenter.

FIG. 15 is a diagram showing a result of row segmentation by a bottom-up segmenter.

FIG. 16 is an explanatory diagram of an algorithm for obtaining a state transition list from a character string set.

FIG. 17: Flexible string matching circuit (F
(SM circuit) block diagram.

FIG. 18 is an extended finite state automaton that allows ambiguous strings.

FIG. 19 is a state transition table of an extended finite state automaton.

FIG. 20 is a diagram illustrating a program of an FSM circuit.

FIG. 21 is a configuration diagram of an FSM circuit according to a second embodiment.

[Explanation of symbols]

100 ... Control subsystem, 200 ... Input subsystem, 300 ... Document recognition device, 400 ... Text search subsystem, 800 ... Search terminal subsystem, 501
... flexible string matching circuit, 151 ... database file, 152 ... image file, 45
1 ... Text file.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G06K 9/20 J (72) Inventor Junichi Higashino 1-280, Higashi Koikeku, Kokubunji, Tokyo Hitachi, Ltd. Central Research Laboratory (72) Inventor Toshihiro Hananoi 2880 Kozu, Odawara City, Kanagawa Stock Company Hitachi Ltd. Odawara Factory

Claims (3)

[Claims] 1. An image file for accumulating a document image, a document knowledge file for storing a layout rule of a document structure for each type of document, an image processing unit for extracting a pattern component from the document image, and an image processing unit for extracting the pattern component from the document image. A character recognition means for recognizing the cut-out character pattern, and referring to the layout rule of the document structure stored in the document knowledge file, analyzing the pattern components extracted by the image processing means. A document recognizing unit that cuts out a character pattern that constitutes a character for each document structure, and character-recognizes the cut-out character pattern by the character recognizing unit to obtain a character string, and a character string that is obtained by the document recognizing unit. And the storage means that stores the data corresponding to the Document filing system, comprising search means for, and output means for outputting the document image of the document identified by the search means from said image file. 2. The storage means has a database file for storing bibliographical items in the document structure and a text file for storing text including the text of the document, and the document recognition means has a document structure name. 2. The document filing system according to claim 1, further comprising: outputting the corresponding character string as a set and storing the corresponding character string in the database file or the text file according to the name of the document structure. 3. The document knowledge file in the document recognition means includes a document title, author name, author affiliation, abstract,
2. The document filing system according to claim 1, wherein parametric knowledge such as font size is also stored in addition to the layout rule of the document structure including the text.