JPH1166230A - Device, method, and medium for document recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device, method, and medium for document recognition which can recognize a document more efficiently. SOLUTION: A document area initializing device generates a document area object, a black pixel circumscribed rectangle extracting device 108 extracts the circumscribed rectangle of connecting black pixel components, and a blank- extracting device extracts a zone of white pixels in the area object as a blank zone. A document area dividing device 109 discriminates and divides a document area, a paragraph area dividing device 110 divides a paragraph as a set of characters, and a character string area dividing device 111 divides a character string as a set of characters. Then a character area dividing device initializes the attribute of a character area object, a character-recognizing device 105 recognizes the characters in the character area, and a closed area dividing device 112 discriminates and divides a closed area which cannot be divided by the blank zone. The areas divided by the respective dividing devices are generated as area objects and given contiguity or inclusion relation as attributes to generate an area division tree, and the dividing process is repeated until all the area objects cannot be divided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a document recognition device, a document recognition method, and a medium used for analyzing a structure of a document image and digitizing data in the document.

[0002]

2. Description of the Related Art When trying to analyze the structure of a document image, a black pixel region is extracted and the image is divided into regions based on a separator (blank region or ruled line) existing between the extracted black pixel regions. . It is necessary to successfully extract separators whose contents vary greatly.

When analyzing the structure of a document image in a conventional apparatus, a character string, vertical and horizontal ruled lines, and other black pixel areas are extracted from the image. Subsequent processing is performed based on the extracted rectangular data.

First, all white regions and long ruled lines having a long width serving as a separator are extracted from the coordinates of a rectangle. Next, the character area is roughly divided using the separator extracted after removing the graphic area. Further, a break (sub-separator) of a component is obtained from a change in line pitch or character size in the character area, and the area is further subdivided according to the sub-separator. By analyzing the image in this manner, the area structure data is obtained as a tree structure. Here, as a method of extracting a black pixel area, a low-pass filter is applied to a white run to combine adjacent black pixels into one to use a bottom-up method. When dividing an area, the area is divided alternately vertically and horizontally.

[0005]

However, in the conventional apparatus, when the document is surrounded by a frame or when an area such as a table exists in the document, the ruled lines are not isolated (the vertical and horizontal ruled lines are in contact with each other). Therefore, it was not possible to proceed with the area division into the frame or table. In addition, since a white area having a size equal to or larger than a certain threshold is roughly divided, it is difficult to set a threshold for a document in which characters of various sizes are mixed. That is, the threshold value set with reference to the large character area cannot be used for the small character area.
As a result, the tree structure, which is the finally obtained region structure data, greatly deviates from the region structure of the document, and there is a problem in that the tree structure must be corrected later.

On the other hand, the inventor of the present application makes it possible to express the region structure well without modifying the obtained tree structure by properly selecting the separator for each division target region. Finally, the blank area and the ruled line in the area to be divided are extracted, and after the extraction, each is considered as the same separator without distinction, and the processing is performed by giving priority to the separator when dividing instead. Simplified. Also, when a blank area is not found when dividing into each division target area, a ruled line is searched and set as a separator, so that even if there is a document surrounded by a frame or a table in the document, the area can be divided well. Like that.

Further, according to the conventional method, since the contents of the area cannot be entered, if a document of a similar format is input, it is not possible to further narrow down the document, so that document recognition cannot be performed efficiently. There was an issue.
If the result of character recognition is used from the structure of the format, the format can be easily limited. Also, not all areas are handled one-dimensionally, but the format identified by a tree structure can further limit the identified format.

An object of the present invention is to provide a document recognizing device, a document recognizing method and a medium capable of performing document recognizing more efficiently in consideration of such problems of the conventional device.

[0009]

According to the present invention, there is provided a document area initializing means for generating a document area object, a black pixel circumscribed rectangle extracting means for extracting a circumscribed rectangle of a connected black pixel component, A blank band extracting unit for extracting a white pixel band in the object as a blank band, a document region dividing unit for identifying and dividing a document region, a paragraph region dividing unit for dividing a paragraph which is a set of character strings, and a set of characters A character string region dividing unit that divides a character string, a character region dividing unit that initializes an attribute of a character region object, and a closed region dividing unit that identifies and separates a closed region that cannot be divided by the blank band. A region divided tree is generated by generating a region divided by each dividing means as a region object and having adjacent or inclusive relation as an attribute, and an entire region object is generated. Once divided up can not divide, the region partition tree is a document recognition apparatus according to the format information of the document area object.

According to a ninth aspect of the present invention, there is provided a document recognition method for recognizing an area of document data from an input document image,
A ruled line is extracted from the input document image and / or a blank band where no document data exists is extracted, and the image data is divided into a predetermined area using the extracted ruled line and / or blank band. A ruled line and / or a blank band is extracted from a predetermined region, the predetermined region is further divided using the extracted ruled line and / or a blank band, and the division is repeated. Generated as
This is a document recognition method for outputting as format information of the document data.

Thus, for example, the document region initializing device generates and initializes a document region object, the black pixel circumscribed rectangle extracting device extracts the circumscribed rectangle of the connected black pixel component, and the blank band extracting device outputs the circumscribed rectangle in the region object. The band of white pixels is extracted as a blank band, and the document area dividing device identifies the document area.
Dividing, the paragraph region dividing device divides a paragraph which is a set of character strings, the character string region dividing device divides a character string which is a set of characters, the character region dividing device initializes attributes of the character region object, The character recognition device performs character recognition of the character region, the closed region dividing device identifies and divides a closed region that cannot be divided by a blank band, generates the divided region in each of the dividing devices as a region object, and determines the adjacent or inclusive relationship. A region division tree is generated by giving it as an attribute, division is performed until all the region objects cannot be divided, and the region division tree is obtained as format information.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the document recognition apparatus according to the present invention will be described below.

In implementing the present invention, object orientation is introduced. The advantage of object orientation is that it can organize problems and encapsulate program details. Before starting the description of the embodiments, the classes used here will be described. In this embodiment, an area object is considered as an object. The region object includes position information, class information indicating the type of the region, format information of the region, separator information dividing the inside of the region, adjacency and inclusion information constituting the region division tree, a black pixel circumscribed rectangle included in the region It has information, structure information indicating the geometric structure of the area, an estimated value of the character width in the area, and the like as attribute values. The class of the basic area of this area object is called the SegCell class, and S
Based on the egCell class as a base class, area objects of the following classes derived from this are generated.

DocCell class (document area: whole document or embedded document) ParaCell class (paragraph area: set of character strings) TextCell class (character string area: set of characters) CharCell class (character area) -OtherCell class (closed area: charts, ruled lines,
TableCell class (table area) LineCell class (ruled line area) FigCell class (figure area) PictCell class (photograph area) First, document recognition of the present invention will be described with reference to FIG. The configuration and operation of an embodiment of the apparatus will be described, and an embodiment of the document recognition method of the present invention will be described at the same time.

A document image is input from an image input device, a document region initialization device generates and initializes a document region object (DocCell class) by regarding the entire input document image as one region. Area identification and division of the document area object, generation and initialization of an area object according to the identification result of the divided area, and if a paragraph area is generated, use a paragraph area division device to divide it into a character string and Generates and initializes a line area object, and if a character string area is generated, generates a character area object by dividing it into characters using a character string area dividing device.
If a closed region is generated, the region is identified and divided using a closed region dividing device, and a region object is generated and initialized according to the class of the region divided by the closed region dividing device. Performs processing according to the class of the area. For example, in the case of a table area, the image is further divided based on ruled lines, and in the case of a figure or photograph area, the image in the area is compressed. The area is divided hierarchically until division is no longer possible by such a method. The attribute value of the character region object generated by the character string region dividing device is set by the character region dividing device. FIG. 2 is a diagram showing how a document area class (DocCell class) is divided in a hierarchical structure. Characteristic are the document area class (DocCell class) and the closed area class (OtherCe class).
ll class), both of which have the possibility of identifying the area in the corresponding area dividing device and dividing the area to generate a document area class (DocCell class).
For example, in the case of a two-column document, the case where the document area object indicating the entire area is divided and two new left and right area objects are generated is the same.

Next, a blank band extracting apparatus for extracting a blank band which is one of the separators will be described with reference to FIG. First, vertical and horizontal separators (Equation 1) 302 and (Equation 2)
301 is obtained. Here, i is a subscript of S _V,
j is a subscript of S _H.

[0017]

(Equation 1)

[0018]

(Equation 2)

Each separator is represented by a section of coordinates orthogonal to the direction. That is,

[0020]

(Equation 3)

## EQU1 ## Here, n _V and n _H are vertical, respectively.
This is the number of blank bands in the horizontal direction, both of which are 2 or more.

Here, the method of obtaining the separator is as follows. If the blank band is a separator, the projection distribution in the target area is obtained in the vertical or horizontal direction, and the threshold (eg, 1) in the projection distribution is obtained. The smaller section is used as a separator. On the other hand, when the ruled line is a separator, the section including the ruled line extracted by the ruled line extraction algorithm is set as the separator. When calculating the projection distribution, it is possible to perform high-speed processing by using the black pixel circumscribed rectangle set in the area object instead of counting the black pixels of the actual image. It is an area and can be processed easily. Further, by using the black pixel circumscribed rectangle, the inclination of the image can be corrected at high speed. That is,
Applying the affine transformation to the entire document image takes a very long processing time, but rotating only the circumscribed rectangle does not take much processing time. For example, the speed can be increased by rotating the center of gravity of the circumscribed rectangle and correcting the width and height of the circumscribed rectangle according to the rotation angle to obtain the rotated width and height of the circumscribed rectangle. Alternatively, a circle (center and radius) inscribed in the original circumscribed rectangle may be rotated. As described above, a circumscribed rectangle may be used in a portion where an image need not be handled, and an image surrounded by the original circumscribed rectangle may be used when an original image such as character recognition is required.

Next, the definitions of symbols used in the following description will be described.

Between the separators S _Vi , S _{Vi + 1} , and S _Hj ,
_Assuming that the section between S _{Hj + 1} is G _Vi and G _Hj , these G _Vi and G _Hj can be expressed as _Equations 4 and 5. In FIG. 3, G _Vi is _denoted by reference numeral 304, and G _Hj is _denoted by reference numeral 303. Here, i, i + 1, j , j + 1 is the index of the S _V, S _H, and G _V, G _H.

[0025]

(Equation 4)

[0026]

(Equation 5)

If the function for determining the size (width) of the separator S _Vi or the interval between the separators is w (x), the size of the separator S _Vi can be expressed as w (S _Vi ).

In addition, the average of variables is μ (x), the standard deviation is σ (x), the maximum value is max (x), and the mode value is f (x). The mode value f (w) of the variable w is obtained by the following equation (Equation 6).

[0029]

(Equation 6)

Here, D (w) is a frequency function, and (Equation 7) is a function obtained by smoothing the frequency function.

[0031]

(Equation 7)

Is as follows. For example, n = 5.

The area is divided on the basis of the separator obtained in this manner. In particular, the area identification processing (classification) is performed when the document area class and the closed area class are divided. The document region dividing device and the closed region dividing device perform special operations. First, the operation of the document area dividing apparatus will be described with reference to FIG. When an isolated ruled line is extracted by the isolated ruled line extracting device 402, the area is divided using the isolated ruled line as a separator. If an isolated ruled line does not exist, it is checked whether the target area is a paragraph area class (ParaCell class) and a character string area class (TextCell class) using the character string determination device 403 and the paragraph determination device 404. If not, a separator (blank band) to be used for division is selected using the document area structure identification device 405, and the area is divided based on the selected separator. In any case, if the separator is selected, the divided area object is generated and initialized by the re-divided area generating apparatus 406, and the area dividing apparatus according to the class of the area is called. Further, when there is no separator to be divided, a region object of a closed region class (OtherCell class) is generated and initialized, and the closed region dividing device 112 is called.

Next, the closed region dividing device 112 will be described with reference to FIG. A ruled line plays an important role in identifying a closed region. In the closed region dividing device 112, a ruled line is extracted by a ruled line extracting device 501, and if there are a plurality of ruled lines, the ruled line is identified as a table (or a form or a TableCell class) in a broad sense. Is generated by the subdivision region generation device 502.
Initialize, otherwise, if there is only one ruled line, a ruled line region object (LineCell class), if the size of the region is close to the size of the character, a character region class (CharCell class), and a photo region class based on the density of black pixels ( PictCell class) or diagram area class (FigCell class). Then, an area object corresponding to each area class is generated and initialized by the subdivision area generation device 502,
Also, a processing device corresponding to the class of the area object is called. For example, if the object is a document area object or a character area object, the above-described document area dividing apparatus 109 and character area dividing apparatus 113 are called. For other area objects, processing such as compression is performed in the case of a photograph area or a figure area, and processing such as vectorization is performed in the case of a ruled area object.

Here, the re-divided area generating devices 406 and 502
Will be described. The overlapping part W _{i, j} of the selected two types of vertical and horizontal separators is obtained by (Equation 8).

[0036]

(Equation 8)

Here, R (l, t, r, b) indicates an area surrounded by two points (l, t) and (r, b).

Then, two overlapping portions W _{i, j} and W _{i + 1, j + 1}
Is a divided region, and this region is subdivided in a similar manner.

The information used for identifying the area is the separator and the estimated value of the character width described above. Here, a character width estimation device will be described. Area containing characters (DocCell, ParaC
ell, TextCell class), a method deemed appropriate in each case is used for estimating the character width. DocCell / Par
In the case of aCell class, character width is estimated from the circumscribed rectangle of the black connected component. A rectangle whose circumscribed rectangle is close to a square region is obtained, and the mode of the width of this circumscribed rectangle is set as an estimated value of the character width of this region. On the other hand, in the case of the TextCell class, the character width is the height of the area. (In the case of vertical writing, it is the width of the area.) Here, the character width estimated in each class is w
Expressed as _C.

Next, the operation of the character string determination device will be described (see FIG. 6). If the following conditions are satisfied, it is determined that the character string is written horizontally. In other words, the fact that the width of the character is similar to the height of the character string is used. Note that the same condition applies to the case of vertical writing.

[0041]

(Equation 9)

[0042]

(Equation 10)

The paragraph judging device performs the following operation (see FIG. 7). The regularity of a character string is used to determine a paragraph. Further, the separator is divided in the direction in which the maximum value of the width of the separator is larger. That is, it is assumed that if (Equation 11), vertical writing is used, and if not, horizontal writing is used.

[0044]

[Equation 11]

This utilizes the fact that the space between character strings is generally larger than the space between characters. Here, assuming horizontal writing, if the following equations (Equation 12) and (Equation 13) are satisfied,

[0046]

(Equation 12)

[0047]

(Equation 13)

The target area can be determined as a set of character strings, that is, a horizontally written paragraph.

A region which has not been determined to be any of the closed region, the character string region, and the paragraph region is identified by using the document region structure identifying device 405 (see FIG. 8). Areas other than the three areas include a column, a chapter, and a title page such as a front page of a dissertation. In the first place, in order to separate a region having a single function or meaning from the other region in a document, a separator having high separability is used so as to be easily understood by humans. For example, where chapters change, new chapters are arranged using blank bands that are wider than the character spacing. In addition, the property that the size of the text in the title and the text in the main text are greatly different like the title page can be used. The document area structure identification device 405 is configured based on such a concept. First, using the blank band width change point extracting device 801 and the projecting width changing point extracting device 802, a blank band in which the contents of the area greatly changes, and at the same time, an average value of the blank band width and an average value of the projected width are obtained. Using the information, the effective blank band selecting device 803 determines how to divide the area.

Here, the blank band width change point extracting device 801
Finds the width of two adjacent blank bands, and if the ratio of the smaller projected width to the larger projected width is equal to or less than a certain value (eg, 0.5 or less), the larger blank band is used as a separator candidate, and , The projection width change point extraction device 802
When two projection widths are obtained and the ratio of the smaller projection width to the larger projection width is equal to or less than a certain value (for example, 0.8 or less), a blank band between the two projections is a candidate for a separator. .

Next, the operation of the character string area dividing device will be described. It is important to divide a character string region not only into a small region but also to cut out the character string region as a character which is a minimum component of a region including a character. For example, it is preferable that the character "matsu" is not divided into two regions, "tree" and "public", but is divided as "pine". The character string region dividing device also involves the identification processing as in the region division described above. However, in the case of a character area, the result of character recognition is an identification process. In addition, it is considered appropriate to use the character recognition result for extracting characters, since humans can read characters correctly because characters can be read.

In this character string region dividing apparatus, a value defined below (hereinafter referred to as a cutout score) is used. Normally, in character recognition, a character candidate area is normalized to a certain size before processing, and thus information such as a character size and an aspect ratio is lost. In other words, in the case of the above-mentioned “pine”, it is divided into “tree” and “public”, and there is a possibility that a high recognition result is obtained.

A cutout score is defined so as to compensate for such inconvenience. Let w _{C be} the estimated value of the character width described above,
When the width of the character candidate region _{RC in} the character string direction and the score of the character recognition i-th candidate are respectively (Equation 14),

[0054]

[Equation 14]

The cutout score d _RC (Equation 15) is

[0056]

(Equation 15)

Is defined. Here, the difference between the scores of the first candidate and the third candidate is obtained because there is no difference in the score when the first candidate and the second candidate are very similar characters. For example, "to" and "to". Note that the difference between the first candidate and the second candidate may be used. Using such a character extraction score, two character extraction methods will be described below. The first is a method similar to the chart method of natural language processing (see FIG. 9). First, take a look strings from the top all divided at a point that can split its and vertically, n pieces determine the small area 901 (the number 16) of the index 902 (the number before and after each of the small areas r _i 17 ).

[0058]

(Equation 16)

[0059]

[Equation 17]

Then, continuous small areas are integrated, and a cutout score of the integrated area is obtained. However, the width of the integrated area to be obtained is all the widths whose width is smaller than 1.2 times the estimated character width. This integrated area can be considered as a character area candidate.

A character lattice corresponding to each of the m integrated areas thus obtained is obtained. Here, the character lattice is defined as a set of four elements of indices v _s and v _e of the corresponding start point and end point of the integrated area, a cutout score d, and a set of small areas constituting the integrated area (expression 18). And represented by the symbol l _j .

[0062]

(Equation 18)

A set of character lattices l _j (Equation 19)

[0064]

[Equation 19]

Can be formed. However, the number of integrated regions (character region candidates) constituting the character lattice l _i is n _i
And

From the lattice set (initial lattice set) obtained in this way, two lattices l _i and l _j connectable by the following connection rule are connected to generate a new lattice l ′, and a set ( Add to equation (19). Here, the lattice connection rule is expressed by the following equation (Equation 20).

[0067]

(Equation 20)

For example, in the case of FIG.
And a character lattice 905 by connecting
, The score of the character lattice 905 is (1615 + 1710) / 2 = 166 according to the lattice connection rule.
It becomes 2.

As described above, the lattices are connected by a method similar to the chart method of the natural language processing, and the lattice having the highest score among the lattices represented by (Equation 21) in the set (Equation 19) is subjected to character segmentation of the target area. Result.

[0070]

(Equation 21)

This method has two advantages. One is to select an arbitrary character-like lattice in a character string and connect adjacent lattices sequentially. Therefore, an indefinite pitch character (alphanumeric) that cannot be extracted by the method of sequentially extracting from the front (described later)
And the other is that the character lattice is held in the smallest unit, so that it is easy to provide feedback in post-processing.

In the second method of extracting characters, small areas are sequentially integrated from the beginning of a character string and characters are extracted (FIG. 10).
reference). Assuming that the cutout of characters has been completed up to a certain small area, adjacent small areas are sequentially integrated starting from the next small area to obtain a cutout score. Then, a character is cut out as an integrated area where the cutout score is maximized.

For example, in the case of FIG. 10, if the division is completed up to r ₁ , first, if r ₂ is a character,
The character "Park" is obtained as a recognition result in the cutout score 176. Next, when a character obtained by integrating r ₂ and r ₃ is a character, a character “dead” is obtained as a recognition result in the cutout score 1710. As in the case of using the character lattice, since the search width of the character search is up to 1.2 times the estimated character width, no more small areas are integrated, and as a result, a character starting from r ₂ is cut out. The "dead" with the highest score is selected. Thereafter, the same processing as the starting point a small region r ₄ is continued until the last small region of the line.

This method is easier and faster than the first method, but has the drawback that if the cutout is incorrect, it is difficult to correct it. Either of these methods may be used in the character extracting device.

Next, a format identification device will be described as a format identification function with reference to FIG. The format identification device performs an operation like a parser (syntax analysis) in natural language processing in combination with the format database. The grammar of natural language processing is
In the area division, format information in a format database, that is, the above-described area division tree, which corresponds to a parser, is a format identification device. FIG.
Shows an image of the format information in the database. FIG. 11A shows a format in which a character string area 1101 is provided and a table area 1102 is provided below the character string area 1101. FIG. 11B shows a format in which a photographic area 1103 is provided and a paragraph area 1104 is provided therebelow. Is shown. As described above, since information on which direction the region is divided by the separator is added to each region, it is possible to know the positional relationship with the adjacent region. Specifically, a method of dividing an area using format information will be described.
It is assumed that an image is input to the document area dividing device, and an area dividing tree, that is, format information is finally obtained. Then, the format is compared with the format information in the format database using the format identification device. Here, instead of directly operating the format information in the format database, the format information used for comparison is temporarily stored in a memory. In the initial state, it is assumed that all the format information in the format database is stored in the memory. Then, following the root of the format information obtained from the input document, the area object corresponding to each node is compared with the area object at the same position in the format information in the memory. For example, if the area located at the top of the format information of the input document image is a character string area, the area object () in the memory at the same position in the format information of FIGS. The character area and the photograph area are compared, and the format information in FIG. 11B that does not match is deleted. As described above, the format information (area division tree) is searched, and the format information on the memory that does not match is deleted, and the last remaining format information is a candidate for the format information that matches the format of the input document. And Note that tracing down to a fine node such as a character area may cause erroneous identification. Therefore, a position of a node to be searched can be determined in advance based on format information. Here, when there is a large number of matching format information, the format information candidates can be further narrowed down by the short distance of the center of gravity of the corresponding area. It is also possible to compare the characters in the corresponding character areas and select a matching character. As a result, if the format does not match any of the formats registered in the format database, the input document is determined to be in an unknown format and presented to the user to prompt registration. At this time, the user looks at the new format displayed on the display device while watching the keyboard 114 and the pointing device 115.
To make corrections, additions and deletions.

According to the first and second aspects of the present invention, it is possible to hierarchically divide an area and generate an area division tree at the same time, and obtain a finally obtained area division tree as a document format.

According to the third aspect of the present invention, accurate format identification can be performed by comparing not only the shape of the area but also the contents in the area.

According to the fourth aspect of the present invention, since the image itself is not processed, the processing time can be reduced, and only the position of the connected component of the black pixel (black pixel circumscribed rectangle) in the area is handled, so that the area of any shape However, a blank band can be easily extracted.

According to the fifth aspect of the present invention, since the rotation processing of the image itself is not performed, the processing time can be reduced.

According to the sixth aspect of the present invention, by identifying a document area, it is possible to limit the processing method of the area.

According to the seventh aspect of the present invention, an area can be divided based on the degree of separation of a blank band, and an area division tree can be generated so as to represent a format.

According to the eighth aspect of the present invention, by identifying a closed area, it is possible to limit the processing method of the area, and it is also possible to divide a complicatedly structured table area.

The same effects as described above can be obtained by using a medium in which a program for causing a computer to execute the functions of all or a part of each means (each apparatus) described in the above embodiment is used. To demonstrate.

The processing operation of each means of the above-described embodiment may be realized as software by a program using a computer, or the processing operation may be performed without using a computer. The circuit configuration may be implemented as hardware.

[0085]

As is apparent from the above description, the present invention has an advantage that document recognition can be performed more efficiently.

[Brief description of the drawings]

FIG. 1 is an exemplary view for explaining the overall configuration of a document recognition system according to an embodiment;

FIG. 2 is a diagram illustrating an area class and hierarchical division according to the embodiment.

FIG. 3 is a diagram illustrating a blank band extracting device according to the embodiment.

FIG. 4 is an exemplary view for explaining a document area dividing apparatus according to the embodiment;

FIG. 5 is a diagram illustrating a closed region dividing device according to an embodiment.

FIG. 6 is a diagram illustrating a character string determination device according to an embodiment.

FIG. 7 is a diagram illustrating a paragraph determination device according to an embodiment.

FIG. 8 is an exemplary view for explaining a document area structure identification device according to the embodiment;

FIG. 9 is an exemplary view for explaining a method of cutting out characters using a character lattice in the character string region dividing apparatus according to the embodiment;

FIG. 10 is an exemplary view for explaining a method of sequentially cutting out characters in the character string region dividing apparatus according to the embodiment.

FIGS. 11A and 11B are diagrams for explaining the operation of the format identification device according to the embodiment;

[Explanation of symbols]

 103 Image Input Device 104 Inclination Detection Device 105 Character Recognition Device 106 Format Recognition Device 107 Format Database 108 Black Pixel Bounding Rectangle Extraction Device 109 Document Region Decomposition Device 110 Paragraph Region Decomposition Device 111 Character String Region Decomposition Device 112 Closed Region Decomposition Device 113 Character Region Dividing device 401 Character width estimating device 402 Isolated ruled line extracting device 403 Character string judging device 404 Paragraph judging device 405 Document region structure identifying device 406 Subdivision region generating device 501 Ruled line extracting device 502 Redivision region generating device 801 Blank band change point extracting device 802 Projection width change point extraction device 803 Effective blank band selection device

Claims (11)

[Claims] 1. A document area initializing means for generating a document area object, a black pixel circumscribing rectangle extracting means for extracting a circumscribed rectangle of a connected black pixel component, and a white pixel band in the area object is extracted as a blank band. Blank band extracting means, document area dividing means for identifying and dividing a document area, paragraph area dividing means for dividing a paragraph which is a set of character strings, and character string area dividing means for dividing a character string which is a set of characters. A character region dividing unit for initializing the attribute of the character region object; and a closed region dividing unit for identifying and dividing a closed region that cannot be divided by the blank band, and generating the region divided by each of the dividing units as a region object. When an area division tree is generated by giving adjacent or inclusive relations as attributes, and divided until all area objects can no longer be divided In document recognition apparatus, wherein the area division tree showing the format information of the document area object. 2. The area object includes position information of the area, class information indicating the type of the area, format information of the area, separator information that divides the inside of the area, and adjacent and adjacent areas constituting the area division tree. The character area object includes inclusive information, black pixel circumscribed rectangle information included in the area, structural information indicating a geometric structure of the area, and an estimated value of a character width in the area. 2. The document recognition apparatus according to claim 1, further comprising a character recognition result obtained by a character recognition unit in addition to the information held by the object. And a format database for storing the format information; a display means; a keyboard and a pointing device, wherein the format information obtained from an input document image and A corresponding area object of the format information present in the format database is compared, and when it is determined that the format information does not exist in the format database, based on the image displayed on the display means and the corresponding format information. 2. The document recognition apparatus according to claim 1, wherein the keyboard and / or the pointing device register the new format information in the format database. 4. The document recognition method according to claim 1, wherein said blank band extracting means extracts a blank band by projecting a black pixel circumscribed rectangle in said area object in a horizontal and / or vertical direction. apparatus. 5. A black pixel circumscribing rectangle which is a part of an input document image, comprising: detecting means for detecting the inclination of a black pixel circumscribing rectangle, rotating the black pixel circumscribing rectangle to correct the inclination, and A document recognition device for performing area division using a pixel circumscribed rectangle. 6. The document area dividing means includes a character width estimating means, an isolated ruled line extracting means, a character string judging means, a paragraph judging means, a document area structure identifying means, and a re-divided area generating means. An isolated ruled line is searched for by the isolated ruled line extracting means, and if the isolated ruled line exists, it is selected as a separator. If the isolated ruled line does not exist, a character string determining means determines whether the target area is a character string or a paragraph. A determination is made using a paragraph determination means, and if it is a character string or a paragraph, a character string area object or a paragraph area object is generated in the re-divided area generation means, respectively. The blank band to be used for division is selected as a separator, and the selected isolated ruled line or blank band is selected at least by the re-divided region generating means based on the separator. Individual document area objects are generated, and if neither a character string nor a paragraph is present and no separator is present, a closed area object is generated by the subdivision area generating means, and the classes of the generated objects, that is, a document area class and a paragraph area class , String area class,
2. The document recognition apparatus according to claim 1, wherein the dividing means corresponding to each of the first class and the closed region class is called, that is, a document region dividing unit, a paragraph region dividing unit, a character string region dividing unit, and a closed region dividing unit. 7. A blank band width change point extracting unit for extracting a blank band width change point at which a blank band width greatly changes in each of a vertical direction and a horizontal direction; Projection width change point extraction means for extracting a projection width change point at which the width of the projection of the sandwiched portion changes greatly, and a blank band used for area division based on the blank band width change point and information on the projection width change point 7. The document recognition apparatus according to claim 6, further comprising an effective blank band selecting means for determining the area, and dividing the area based on the strength of separation of the blank band. 8. The closed area dividing means includes: a ruled line extracting means;
And a ruled line extracting means for extracting ruled lines in the horizontal and vertical directions. If the number of the extracted ruled lines is one, a ruled line area object is generated. An object is generated, and at least one document area object is generated by the re-divided area generating means using the ruled line as a separator. If the ruled line does not exist, the character determining means determines whether the size is close to the size of the character, If it is close, a character area object is generated. If the area is not determined as a character, an area larger than the character generates a figure area object and a photograph area object based on the density of black pixels in the area. Generates an area and creates a document area class, figure area class, and photo area Document area division means, figure area processing means, photograph area processing means, character area division means, ruled area processing means, and noise area processing corresponding to each of the class, the character area class, the ruled line area class, and the noise area class 2. The method according to claim 1, further comprising: calling means.
Document recognition device as described. 9. A document recognition method for recognizing a region of document data from an input document image, comprising: extracting a ruled line from the input document image and / or extracting a blank band where no document data exists; And / or dividing the image data into predetermined regions using blank bands, extracting ruled lines and / or blank bands from the divided predetermined regions, and using the extracted ruled lines and / or blank bands. Further dividing the predetermined area, repeatedly performing the division, generating the inclusion relation of the division as an area division tree,
A document recognition method, wherein the document data is output as format information of the document data. 10. A medium storing a program for causing a computer to execute the functions of all or a part of each of the means described in any one of claims 1 to 8. 11. A medium on which a program for causing a computer to execute all or a part of each step according to claim 9 is recorded.