JP2001344569A - Method for recognizing character and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To appropriately segment an adjacent or contact character string and also to improve recognition accuracy while suppressing the increase of processing time. SOLUTION: A cross section series graph is extracted (2) from input image data (1), and an element tag for managing the cross section series graph is generated (3). A virtual boundary dot sequence is generated (4) for the peculiar area of a cross section series. A character candidate is generated by combining element tags, one character us recognized (7) by tilling up the virtual boundary dot sequence necessary to the character candidate, a link is generated between character candidates in which one character is recognized, a path is generated (6) between links, an optimum path is selected and the recognition results of a character string are outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method and a character recognition processing program having a function of recognizing one character and a character string recognition function of appropriately cutting out and recognizing a contact character or a separated character in character recognition. Is applied to, for example, character string recognition in which a plurality of lines are recognized, character string recognition in consideration of ruled line contact, and the like.

[0002]

2. Description of the Related Art As a conventional technique for recognizing a character string, for example, there is a word reading method described in Japanese Patent Application Laid-Open No. 5-233877. In this method, even when characters are extremely separated or in contact with each other, a correct character pattern can be obtained by cutting out characters by a plurality of cutting-out methods.

In the above-described method, when cutting out characters,
The data of the peripheral distribution of the character string pattern and the data of the circumscribed rectangle are used. That is, the rectangle is vertically divided based on the peripheral distribution. Further, the conventional character recognition method employs a configuration in which image data is directly accessed to process a character candidate.

[0004]

There are some adjacent character strings or contact character strings that cannot be separated into rectangular shapes or those that cannot be divided linearly. In the above-described conventional method, recognition processing is also performed on such a proximity character string or a contact character string in units of rectangles, and thus there is a limit in recognition accuracy.

[0005] Further, since it is necessary to access image data, an increase in processing time is inevitable.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object of the present invention is to use a virtual boundary point sequence to appropriately cut out a proximity or contact character string and to increase processing time. An object of the present invention is to provide a character recognition method and a recording medium in which recognition accuracy is improved while suppressing the occurrence.

It is another object of the present invention to provide a character recognition method and a recording medium in which recognition processing is speeded up by managing image characteristics of different hierarchies by using tags.

[0008]

According to the present invention, a contact character can be separated by generating a virtual boundary point sequence for a singular region of a sectional sequence graph. Separated characters are represented by a virtual boundary point sequence and a boundary point sequence and are recognized. The above-described virtual boundary point sequence is formed with a smooth curve, and improves the recognition accuracy of character recognition.

Further, in the present invention, the structural elements of the sectional sequence graph are managed by tags, and the expression form (logical structure) of the tags is shared.

[0010]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be specifically described below with reference to the drawings. First, the principle of the present invention will be described. The basic idea of the present invention is the following three.

The first is a concept of “combining characteristic elements (SS tags, SR tags, and BCC tags described later)”. This is realized by using a sectional sequence graph and a new method of a virtual boundary point sequence. By this method, the recognition accuracy is improved by correctly cutting out the contact character string, and an increase in the processing amount at that time is suppressed.

The second is the idea of "common use of expression forms relating to a plurality of types of character elements (cross-sectional series, singular regions, boundary point sequences) and character candidates". Layers and lower layers) with tags,
This is realized by a new method of standardizing the expression format of tags. Thereby, the processing amount is compressed, and the processing means is particularly simplified.

The third is a concept of "integration of a backtrack type approach and a hypothesis verification type (path selection type) approach utilizing hierarchy". Both are well-known existing methods. To put it simply, the former is an approach of "redoing the process only in the case of doubt." The latter is an approach of "listing more possibilities from the beginning and selecting the best combination among them". These integrations are realized by a new method of reservation / inheritance processing in addition to the introduction of the tag described above. This suppresses an increase in the processing amount while maintaining the recognition accuracy.

That is, according to the present invention, the recognition accuracy and the processing amount are balanced by "adding a virtual boundary point sequence to a sectional sequence graph and managing it by tags".

(Cross Section Series Graph) First, a cross section series graph will be described. The cross-sectional series graph is a method for expressing a line figure proposed by the present inventor (Japanese Patent No.
615247 and JP-A-6-68305).

That is, a cross section substantially orthogonal to the direction of the line segment is extracted from the line graphic image data, a series of the cross sections is used as a cross section series, an area other than the cross section series is used as an unknown area, and a node is used as a cross section series or an unknown area. And a graph structure (cross-section sequence graph) representing the connection relation of nodes with arcs is created, and the main point is to make this cross-section sequence graph a feature of the line figure.

FIG. 12 shows a relationship between line graphic image data (original image data), a sectional series graph, and a skeleton line (core line). In a line segment description using a cross-sectional series graph having such a hierarchical structure, the cross-section that is the structural element has the property of being substantially orthogonal to the direction of the line segment. They are ordered as if they were time series ((a) in FIG. 12).
Part). The unknown region is a region that is not included in the cross-sectional series, and corresponds to a line-shaped end, a bent, branched, or crossed planar region (a part (b) in FIG. 12).

As described above, the cross-sectional series graph is a simple one in which a line figure is clearly divided into a simple line and a part which is not so, and a line figure is described. Rough structural matching is facilitated.

The sectional sequence graph describes image data hierarchically. FIG. 13 shows an example of a character string image. FIG. 14 shows a boundary point and a cross section of the cross-sectional series graph, and FIG. 15 shows a cross-sectional series and a singular region of the cross-sectional series graph.

In FIG. 14, each point on the contour (black pixel side) is called a boundary point. The arrangement of the boundary points corresponding to the outline is called a boundary point sequence. The types of the boundary point sequence include an outer side and an inner side. In the example of FIG.
There are two inner boundary point sequences (“O” and “O” portions inside the character “8” in FIG. 14).

A cross section is a pair of boundary points determined so as to be substantially orthogonal to the direction of the line segment, and is represented by a straight line connecting the boundary points in FIG. A series of cross-sections without any inconsistency are a series of cross-sections, and the rest are singular regions. The specific region is described in the above-mentioned Patent No. 26152.
In No. 47, it corresponds to an unknown area. In the present invention, it is called a unique region. The cross-sectional series represents a normal character line portion, and the unique region represents a connection portion or an end point portion.

The structural elements {boundary point, boundary point sequence, cross section, cross section series, and unique area} are features themselves and also have a role of structuring image data by combining them.

By defining {Boundary point, boundary point sequence, cross section} as a structural element of the lower layer, and {Cross section series, singular region} as a structural element of the upper layer, the upper layer is roughly composed of line figures (characters). The lower layer represents a more detailed structure.

However, in the present invention, a structural element called a cross section will be described as being positioned not in a lower layer but in an upper layer in consideration of error measures in an extraction step described later.

The present invention is based on the use of a cross-sectional series graph.
~ 20. 16 to 20 show the description format and extraction method of each structural element. The extraction method is described in the above-mentioned Patent No. 2615247 and the like, and will not be described.

FIG. 16 shows an expression form of the boundary points. One boundary point is described by one structure. Specifically, as shown on the right side of FIG. 16, the structure is arranged in a continuous area 18 on the memory, and the continuous area 18 has Mb indicating the upper limit thereof and the number of boundary points actually used. Nb 2
Is managed by two variables.

The boundary point structure 10 has a coordinate value (x, y) 1
1. Shading gradient direction (Dx, Dy) 12 (direction from the character background to the character line portion, almost perpendicular to the boundary direction), section sequence number 13, boundary point sequence number 14, front boundary point number 1
5 (the number of the boundary point that appears next when the character line portion is viewed to the left), the number 16 of the rear boundary point (the number of the boundary point that appeared before the character line portion is viewed to the left), and the number of the opposite boundary point 17 (a boundary point q that first appears when a black pixel is searched for from the boundary point p of interest in the direction of shading is referred to as an opposing boundary point to p and is referred to as an extended boundary point in the above-mentioned patent).

FIG. 17 shows an expression format of the boundary point sequence. Similarly to the above-described boundary point structure, one boundary point sequence is described by one structure, and a plurality of boundary point sequences 0 to Mb-1 are arranged in the continuous area 27 on the memory.

The boundary point sequence structure 20 has a status 21
(Outer boundary point sequence, inner boundary point sequence, flag for distinguishing noise), uppermost left boundary point number 22, boundary point number 23 (number of boundary points for one round), upper left coordinate value (Xs , Ys) 24, the lower right circumscribed coordinate value (Xe, Ye) 25 of the circumscribed rectangle of the boundary point sequence, and the work flag 26 (work variables temporarily used in various processes).

FIG. 18 shows an expression form of a cross section. Similarly to the above-described structure, one cross section is described by one structure, and a plurality of cross sections 0 to Ms−1 are connected to the continuous area 3 on the memory.
5.

The cross-sectional structure 30 has one boundary point number (b1) 31, the other boundary point number (b2) 32, b1
No. 33 of the cross section on the front side of No. 3 and number 34 of the cross section on the rear side of b2.

FIG. 19 shows an expression format of a sectional series. The sectional series structure 40 has a status 41 (a flag indicating valid / invalid), a number of sections 42, and a section number 43 on the Head side.
(When a cross section in a cross section series does not have another cross section in the forward direction, the cross section is called a cross section series Head), Tai
Section number 44 on the l side (when a section does not have a section in the backward direction, the section is called a Tail of a section series), He
It is composed of a unique area number 45 on the ad side, a unique area number 46 on the Tail side, an average section length 47 (corresponding to the estimated line width), and other work variables 48 (point (node) numbers in a graph expression, etc.).

FIG. 20 shows the expression format of the unique region. The unique region structure 50 includes a status 51 (a flag indicating valid / invalid) and the number 52 of cross-sectional series to be connected (the effective number of the following two arrays). , Section series number column 53, connection type column 54 (column 53 and column 54 are arrays corresponding to each other,
The type distinguishes the end of the sectional series), circumscribed coordinate values (Xs, Ys) 55, circumscribed coordinate values (Xe, Ye)
56, and other work variables 57 (points (nodes) of graph representation, etc.).

It is possible to access other structures through the numbers of the sectional series of the structures and the numbers of the boundary point sequence.

As described in the above-mentioned patent, in the process of obtaining the cross-sectional series graph, a graph representation of a skeleton is extracted. It is also possible to extract a graph representation of the outline by approximating the boundary point sequence with a polygon. By using the method of analyzing these graph expressions in combination, a unit for recognizing one character can be realized. In addition, a different one-character recognizing unit can be realized by using the method together with a method of statistically analyzing the direction of the boundary point.

In the following description, the above-mentioned structural elements are described as follows. The boundary point is B (abbreviation of Border), the boundary point sequence is C (abbreviation of Contour), and the cross section is S (Sli).
ce), and SS (Slice Seque)
nce), the unique region is designated by SR (Singular R)
(short for egion).

(Regarding Virtual Boundary Point Sequence) One feature of the present invention is that the concept of a virtual boundary point sequence is used.
This will be described with reference to the drawings.

FIG. 21 shows a virtual boundary point sequence segment connected from the boundary point bs to the boundary point be. Each boundary point of the virtual boundary point sequence segment is indicated by a black circle, and these are shown in FIG.
It is described in the same expression format as. The boundary point structure 10 in FIG. 16 includes a front boundary point number 15 and a rear boundary point number 16, and these are used to connect the boundary point structures in a bidirectional list.

Originally, the start boundary point bs and the next boundary point bsR (R is abbreviation of Real) after bs are connected.
By switching this connection, the virtual next boundary point bs between bs and bs
If V (V is an abbreviation for Virtual) is connected, it is as if the point of the train was switched. If the connection is similarly switched on the end boundary point be side and the boundary point is traced, the obtained boundary point sequence (virtual boundary point sequence in a broad sense) is different from the original one.
This virtual boundary point sequence functions effectively when expressing separation of contact characters, as described later.

The virtual boundary point sequence segment and the reconnection information are collectively referred to as a virtual boundary point sequence in a narrow sense, and the virtual boundary point sequence is described in an expression form shown in FIG. The structure management method is the same as the method described with reference to FIGS.

The virtual boundary point sequence structure 60 shown in FIG.
A key 61 described later (shown in FIG. 23), a boundary point number 62, a start boundary point number (bs) 63 described in FIG. 21, a next boundary point number (bsR) 64 for bs, and a virtual next boundary point number (bsV) for bs 65 (these three correspond to the connection change information of the start boundary point), the end boundary point number (be) 66, the front boundary point number (beR) 67 of be, and the virtual front boundary point number (b) of be
eV) 68 (these three correspond to the connection change information at the end boundary point).

The key 61 is used to store the calculated virtual boundary point sequence
Information for later retrieval. The format of the key is shown in FIG.
As shown in the figure, the unique region number 611 and the start boundary point (b
A reference number 612 of the cross-sectional series having the s) and a reference number 613 of the cross-sectional series having the end boundary point (be) represent the connection pattern of the unique region. Note that this key can be omitted when dynamically obtaining a virtual boundary point sequence.

The connection pattern example shown in FIG. 23 targets the unique region (SR5) in FIG. Now, consider a connection pattern when the singular region SR5 and the sectional series SS9 (lower side) and SS6 (left side) are designated. When referring to two cross-sectional series from SR5, first, the unique region structure 50 of the unique region number (SR5) shown in FIG. 20 is referenced, and then the sequence data 53 is referenced. This array data is set in advance so as to have a counterclockwise order.

That is, specifically, S
Suppose that S6 is set and SS9 is set third. In this case, the connection pattern is represented in FIG.
It becomes as shown in.

A problem arises when the cross-sectional series forms a loop and connects to the same unique region. That is, in the example of FIG. 24, when there is the other of SS9 instead of SS6, simply specifying only the cross-section series number indicates which of the four boundary points is the start point and which is the end point. It becomes impossible to specify whether to do it.

The following describes how to specify the start point and end point of the virtual boundary point sequence segment.

In the example of FIG. 24, four cross-sectional series {SS3, SS6, SS7, SS
9｝ are connected.

Information for referring to the cross-sectional series from SR5 corresponds to the array data 53 in FIG. This is SR
5. Notated as SS [i] (i = 0,..., 3). It is assumed that this matches the order when the cross-sectional series are arranged counterclockwise as seen from the left side of SR5, as described above.

That is, in the case of FIG. SS [0] = SS7, SR5. SS [1] =
SS3, SR5. SS [2] = SS6, SR5. S
S [3] = SS9

Now, the designated sectional series is SS9 and S
Since it is S6, SR5. SS [0]: not included, SR5. SS
[1]: Not included, SR5. SS [2]: Include, S
R5. SS [3]: Included.

Assuming that this sequence is used cyclically, the location where the virtual boundary point sequence segment should be generated
5. SS [3] = SS9 to SR5. SS [2] = SS
Since it is 6, it can be understood that it eventually matches the cross-sectional series sandwiching the continuous section of the “excluding” cross-sectional series.

In the above example, SR5. SS [0]-S
R5. SS [1] is a continuous section “not including”, and one of which sandwiches this section is SR5. SS [3] = SS9
And the other is SR5. SS [2] = SS6.

When this is generalized, instead of using the SS numbers as they are, “SS numbers not included when the SSs referred to from the SRs are arranged in order and cyclically used are used.
The SS before the continuous section is the start side, and the SS after
Is the end side. "

Further, “on the start side, the (last) boundary point on the right side toward SR as the start boundary point among the cross sections closest to SR, and on the end side, among the cross sections closest to SR, The left (first) boundary point toward SR is the end boundary point. "

These are rules for determining a connection pattern and specifying a start boundary point and an end boundary point when one SR, an SS on the start side, and an SS on the end side are designated. When the SS on the start side and the SS on the end side are exchanged, the other virtual boundary point sequence segment is selected.

If there are three or more SSs connected to the SR, correct processing can be performed by replacing the pair with a pair and applying the above rules to each pair. In addition, SR5. SS
[3] and SR5. If SS [0] is specified in this order, there is no missing SS in the meantime, so the actual boundary point sequence may be used as it is. In addition, SR5. SS
[I] and SR5. When SS [i] is designated, that is, when exactly the same SS is designated, this designation corresponds to processing as an end point, but can be processed without exception in the above-described processing.

Next, how to generate a virtual boundary point sequence between the two boundary points specified as described above will be described.

The simplest method is to obtain a coordinate value that passes when connecting two boundary points in a straight line, and to generate a virtual boundary point using the coordinate values as the coordinate values. This method loses the curvilinear smoothness and generates unnecessary corners at the connection points, which may reduce recognition accuracy.

Therefore, in the present invention, instead of connecting two boundary points in a straight line, a curve generation method used in fields such as CG and font generation is used. Specifically, 2
Point coordinate values P0, P1 and their tangential directions P0 ', P
Since an estimated value of 1 ′ is obtained, the Ferguson method or the like can be used. The Ferguson method is a parametric method in which the above four vectors are connected by a cubic polynomial, and is known as a method of generating a curve. The present invention is not limited to the Ferguson method, and other curve generation methods may be used.

According to the present invention, since the virtual boundary point sequence can be generated by using the curve generation method as described above, it is effective in improving the recognition accuracy.

When the line width is very small, the length of the virtual boundary point sequence segment becomes short, and the accuracy of interpolation may hardly affect the accuracy of character recognition. The same applies when the boundary points at both ends are extremely close. Therefore, in such a case, the boundary points at both ends may be logically directly connected without intentionally generating a virtual boundary point. Specifically, only the connection information of the boundary point needs to be changed. Thereby, the processing amount is reduced.

As described above, an important point in the virtual boundary point sequence according to the present invention is that the virtual boundary point sequence is not plotted on the image plane, but is merely represented on a feature expression such as a sectional sequence graph. Will be realized. "

(Expression Form of Tag) In the processing of the present invention, a logical structural element called a tag is used. FIG. 25 shows a tag expression format. The tag structure 70 includes a status 71 (valid / invalid, integration permission / non-permission, and a flag indicating the type of tag), the number of children 72 (valid number of the following two arrays), a child number column 73, and a child number column 73. Column 74 (the number column and the type column correspond to each other), circumscribed coordinate values 75 (Xs, Ys), 76 (Xe, Ye), and the number 77 of link destinations (effective number of the following two arrays) Column 78 of the tag number of the link destination, column 79 of the evaluation value of the link destination (the column of the tag number and the column of the evaluation value correspond to each other), the evaluation value 8 of the tag
0, a recognition result number 81, and other work variables 82 (variables such as a variable indicating a connection pattern of a unique region, a corresponding number of boundary points, and a work flag).

As described above, the tag structure has two types of sequence pairs. The former array manages the children, and can thereby uniformly manage other tags or structural elements of the cross-sectional series graph. Conceptually, it represents a vertical connection between the tag and the structural element of the cross-sectional series graph.

The latter array manages link destination tags and conceptually represents a connection in the character string direction. The tag structure is similar to the cross-sectional series graph described above,
It is managed as a continuous area on the memory.

Tags are classified into several types according to the contents of the child tags. An element tag is a tag having a structural element of a black component as a child instead of a white background. In the present invention, the description will be made in combination with the cross-sectional series graph, so the following three types are available, but in general, the present invention is not limited thereto. See FIG. 2 for the following description.

The first of the element tags is an SS tag, which is a tag having a section series as a child. The second of the element tags is an SR tag, which is a tag having a specific region as a child. Third element tag
Is a BCC tag, which is a tag having a boundary point sequence as a child.
BCC stands for Black Connected Component (Black Connected Component)
Component). However, as described later, the black connected component is not actually obtained. When a part of the boundary point sequence is a virtual boundary point sequence, a tag having the virtual boundary point sequence as a child is described as a VC tag. VC
Is a virtual boundary point sequence (Virtual Contour)
Is an abbreviation for

In the tag structure shown in FIG. 25, the types of children can be set individually, so that one tag can have children of different types, but the description is not complicated. Therefore, the above-mentioned names are used without considering such tags.

A VS tag is a tag having no children.
It is used to represent a blank area using the circumscribed coordinate values of FIG. VS is a virtual space (Virtual Spa).
ce). Tags corresponding to the start and end of the character string are denoted as an R tag and an E tag, respectively. Abbreviations of Root and End, respectively.

A VCC tag is an element tag (SS tag, S tag
R tag, BCC tag) or VS tag as a child. VCC is a virtual connected component (Virtual Con
abbreviated term for “connected Component”. VC
The C tag represents one character candidate for the extraction method,
This is a character recognition processing target.

The path tag is a general term for the VCC tag and the VS tag, and is a target for path generation and path selection.

As described above, a link indicates a connection relationship between tags in the character string direction, and is mainly used for a path tag.

A path is a sequence of path tags arranged in sequence via links, and begins with an R tag and ends with an E tag. Represents one interpretation candidate for a character string.

Conceptually, tags also implement a hierarchical description. Element tags and VS tags as lower layers, and path tags (VCC and VS tags) as upper layers
Can be considered. The VS tag straddles both. This is because there is a case where a space is a space between characters and a case where it is a space within a character, and the space corresponds to both.

The element tags can be further divided into two layers. The upper layer is a BCC tag, and the lower layers are an SS tag and an SR tag.

The section sequence graph constitutes a layer lower than the element tag. However, in the cross-sectional series graph, BCC is the lower layer and SS and SR correspond to the upper layer, but as element tags, the SS tag and SR tag are the lower layer and the BCC tag is the upper layer. This is because a plurality of boundary point sequences are collected and the one corresponding to the black connected component is used as the BCC tag, which can be regarded as a coarser expression.

The tag has two roles. One of the functions is to bridge between character elements (structural elements) and character recognition means. This is achieved in that the VCC tag can access the structural element via the element tag. This effectively acts to realize the “combination of characteristic elements”. Also, by introducing element tags, differences in the types of structural elements can be absorbed.

The other is a role as an operation target of character string recognition centering on path selection. A feature of the tag is that it has both of these roles. As described above, since the expression format of the processing target is shared as a tag, various processing can be shared. Therefore,
The implementation means can be simplified.

FIG. 26 shows a tag control structure for controlling the entire operation centering on such a tag. In FIG. 26, the processing mode variable group is a variable group for indicating the character type and function (single character recognition / character string recognition function) designated by the user, and the layer priority. The layer priority is information relating to up to which layer of the sectional sequence graph is to be obtained and which layer of the element tag has priority.

Subsequently, the elements of the sectional sequence graph, the group of pointers to the virtual boundary point array, and the pointers to the tag array are arranged.

FIG. 29 shows a tag sorting result array.
This array shows the result of numbering tags and rearranging the numbers, rather than reordering the tags directly. Although the upper limit of this array is fixed, the effective number of the array changes during the processing.

Further, a variable group such as an array expressing a path, a pointer to a recognition result array, and geometric information of a character string follows. At the end of the tag control structure, there is an array as a group of characteristic tag numbers. FIG. 28 is a diagram illustrating the use of a tag array. Since the tag array is used in order from the beginning,
The effective number will gradually increase in the course of processing. However, as shown in FIG. 28, the number of a characteristic tag can be remembered, and thereby the range of the tag to be processed in each step can be limited.

For example, the element tag is from the beginning to immediately before the first R tag. The VS tag is from the R tag to the E tag. The lower layer tags are from the beginning to the E tag (element tag + VS
(Because it is a tag) The VCC tag is from the beginning of the VCC tag to immediately before the VC tag. The upper layer tag is from the R tag to immediately before the VC tag (VS tag + VCC tag).

If this is used in combination with the sorting result array, it is possible to limit the processing range while avoiding the replacement of tags.

Hereinafter, embodiments of the present invention will be described. FIG.
FIG. 5 is a process flowchart for explaining an outline of an embodiment of the present invention. FIG. 2 is a diagram illustrating a hierarchy of tags according to the present invention. The outline of the processing of the present invention will be described with reference to FIGS.

The image data is inputted (step 101).
Extract a sectional series graph from the image data (Step 1)
02). Next, an element tag for managing the sectional sequence graph is generated (step 103). The BCC tag, which is the upper layer of the element tag, manages the lower layer of the sectional sequence graph, and the SS tag and SR tag, which are the lower layers of the element tag, manage the upper layer of the sectional sequence graph.

Subsequently, a virtual boundary point sequence is generated for the singular region of the sectional series (step 104), and character candidates are generated by combining element tags (step 10).
5). Character candidates are managed by VCC tags.

Next, one character is recognized after supplementing the virtual boundary point sequence required for the character candidate (step 106), and a link is generated between the character candidates recognized for one character (step 1).
07). Following the link, a path is generated, an optimal path is selected (step 108), and a character string recognition result is output (step 109).

FIG. 3 shows the configuration of the embodiment of the present invention. In FIG. 3, reference numeral 1 denotes an image input unit for inputting image data;
Denotes a cross-section sequence processing unit that extracts a cross-section sequence graph from image data, 3 denotes a tag processing unit that generates and manages tags, 4 denotes a virtual boundary point processing unit that generates virtual boundary points, and 5 stores various structures. A memory, 6 is a link / path generation unit that generates links and paths, 7 is a character recognition unit, and 8 is a control unit that controls the whole. The input image data may be binary or multi-valued. In the case of multi-valued image data, the processing is performed while determining the white / black of the image with a predetermined threshold value.

FIGS. 4 and 5 show flowcharts of the entire processing relating to the processing of the present invention. In the preparation processing in step 201, information necessary for the following processing is obtained. That is, a character string image to be processed is input from the image input unit 1.
FIG. 30 shows an example of a character string image to be processed. Also, input user-specified information. This user specification information is information for specifying a character type, a function (information indicating one-character recognition or character string recognition), and the like.

In step 202, the control unit 8 stores the acquired information in a tag control structure (FIG.
The processing mode is set to 6), and the layer priority is set based on the contents.

In step 203, according to the set processing mode, the sectional sequence processing section 2 extracts a sectional sequence graph and stores it in the memory 5 in the format shown in FIGS. FIG. 31 shows the lower layer (sequence of boundary points) of the sectional series graph.
And the upper layer (cross-sectional series SS and unique region SR) of the cross-sectional series graph. FIG. 31 is a diagram modeling the tag structure.

If the BCC tag is prioritized, the upper layer (SS / SR) in the sectional sequence graph does not need to be obtained, so that the processing can be omitted.

In step 204, the tag processing unit 3
Calculates the BCC tag based on the boundary point sequence in the lower layer of the sectional sequence graph. This processing will be described using the boundary point sequence (C0 to C2) in FIG. 31 as an example.

First, for the outer boundary point sequences C0 and C1,
A corresponding tag is created for each, and these are registered in the memory 5 as a BCC tag 0 and a BCC tag 1 (FIG. 2).
5). The corresponding outer boundary point sequence C as a child of the BCC tag
0 and C1 are registered in the memory 5 (FIG. 17). At this time, the outer boundary point sequence in which the number of boundary points is equal to or less than a predetermined threshold is
Since there is a high possibility of noise, it is excluded from the target of BCC tag creation.

Next, a new BCC tag is not created for the inner boundary point sequence C2, but a corresponding one is obtained from the above-mentioned BCC tag and registered as a child of the BCC tag 0. Specifically, on the cross-sectional series graph, the correspondence of the boundary point sequence is obtained, and the result is reflected on the tag. In other words, assuming that the boundary points are extracted in the order of scanning the image, one arbitrary boundary point is extracted from the inner boundary point sequence, the boundary point information is extracted in the reverse order of registration, and the boundary point that appears first is extracted. (However, those belonging to the outer boundary point sequence) may be found. A BCC tag to be obtained has an outer boundary point sequence corresponding to the outer boundary point as a child.

An important point in the above-described BCC tag creation processing is that not only the black connected component itself is obtained but also the boundary point sequence of the black connected component and the structured information for extracting the boundary points as necessary. It is a point that is being created.

In step 205, the tag processing unit 3
Checks the number of BCC tags obtained. If the number of tags is 0, the control unit 8 determines that the input character image is blank, proceeds to step 217 in FIG. 5, sets a blank code, and ends the processing. If the number of tags is not 0,
Proceed to step 206.

In the processing of steps 206 to 216, character string recognition is executed. In step 206, the tag processing unit 3 sets the upper layer (SS
And SR) are determined based on the respective tags. That is, the SR tag 1 corresponding to the unique region SR1 in FIG.
Is created, an SS tag 0 is created corresponding to the cross-sectional series SS0, and so on.
SS tag 2 corresponding to SS2, SR tag corresponding to SR2
The tag 2 is created for the SS tag 1 corresponding to the SS1, and the SR tag 3 is created for the SR3.

If the upper layer of the sectional sequence graph has not been obtained in step 203 according to the layer priority of the processing mode, the number of tags created here is 0, but the element tags are already in the step. B obtained in 204
Since there is a CC tag, the following processing can be continued.

The processing in step 206 is the same as step 204
Is in a correspondence. At this point, as shown in FIG. 31, all necessary element tags are obtained. The structure of FIG. 31 is stored in the structure for controlling the tag of FIG.

However, since the BCC tag, the SS tag, and the SR tag have duplicate character elements represented by them,
By performing selection in step 211 described later, duplication is eliminated.

In step 207, the control unit 8 extracts a feature (for example, the height of a character) relating to the entire character string based on the element tag. The extracted features are set as the contents of a variable group such as geometric information of a character string in the structure of FIG. These are used in a process of evaluating character likeness (step 213) as described later. In step 207, the tag processing unit 3 sorts the element tags. FIG. 32 shows the result of sorting the element tags (tag numbers) in FIG. The sorting of the tag numbers is shown in FIG.
Are rearranged in the arrangement order from the left of the upper layer (SS and SR) in the cross-sectional series graph shown in FIG. Similarly, the numbers (0, 1) of the BCC tags are arranged in the order of arrangement from the left, and inserted between the numbers of the SS and SR tags. The insertion position is determined in consideration of the barycentric position of the circumscribed rectangle of the boundary point sequences C0 and C2 (position in the horizontal direction of the character string) and the barycentric position of the circumscribed rectangle of C1 (position in the horizontal direction of the character string).

The sort result at the time of step 207 is as follows:
As shown in FIG. 29, the arrow indicates the range to be sorted.

In step 208, the tag processing unit 3
Examines the adjacency when converted to a position on the image plane for the combination of BCC tags,
A VS tag corresponding to the gap (character string direction) between C tag pairs is obtained. FIG. 33 shows the arrangement of the tag numbers together with the lower layer (boundary point sequence) of the sectional sequence graph. FIG. 33 shows V
S tag 0 is also shown.

The VS tag is for expressing a (virtual) blank as described above. Since the BCC tag substantially corresponds to the black connected component, there is always a gap between the BCC tags (tag numbers 0 and 1 in FIG. 33) that are adjacent to each other on the image plane. However, there may be a case where there is a completely blank area in the direction perpendicular to the character string direction (the example in FIG. 33), and there may be cases where this is not the case. The case other than the above is a case where there is an oblique blank area or a complicated blank area. Even in such a case, it is assumed that there is a virtual blank area and registered as a VS tag. By examining the coordinate values of the oblique blank area and the intricate blank area described above, it can be seen that the character cutout frames overlap (this is referred to as a negative blank). FIG.
Indicates a negative blank example.

The tag processing unit 3 determines that a VS tag having a sufficient size in the character string direction is not a blank in a character, and sets the status of the tag structure shown in FIG. ". This status is reflected in the processing in step 213 described later, and contributes to suppressing an increase in the number of VCC tags.

Further, the tag processing section 3 determines that there is a virtual blank at both ends of the character string, and sets a blank corresponding to the start end as an R tag and a blank corresponding to the end as an E tag. As described above, both the R tag and the E tag are a kind of the VS tag.
An R tag, a VS tag, and an E tag are added to the element tags in FIG.

In step 208, VS tags are added to element tags for sorting. The sorting result is shown in FIG.
Shown in Further, in the sorting result at this point, as shown in FIG. 29, the arrow up to the E tag indicates the range to be sorted.

With the above processing, the element tags (BCC tag, SS tag, SR tag) and the VS tag are obtained. However, since the element tag is not a path tag, it is not directly used as a component of the path (that is, what is considered to correspond to one character). V that is a component of the path
The CC tag represents various combinations of element tags with one tag, and they are obtained in step 213.

In step 209 (FIG. 5), the tag processing section 3 performs a reservation process for the element tags. This reservation process has the following four roles.

First, characters that should be processed as exceptions in the normal processing after step 210, specifically, characters having a special size such as punctuation, are recognized in advance at this point. The reason is that special characters such as punctuation are determined according to the arrangement of surrounding characters,
This is because it is difficult to determine with high accuracy only by one-character recognition based on a shape or the like without depending on the arrangement. By performing the processing appropriately at this point, it is not necessary to perform special processing such as evaluating the size in the post-processing.

Second, a backtrack type approach,
It consists in fusing a path selection (or hypothesis testing) approach. The combination of the reservation processing and “selection of element tags” in step 211 results in a backtrack type, and the combination of step 215 and subsequent processing results in a hypothesis verification type (or path selection type). More specifically, a reservation is made for the upper layer (BCC tag), and the corresponding lower layer (SS
The combination of the remaining element tags is created in step 213.

Third, a special character recognition method is effectively introduced. At this point, if high-quality characters can be determined early, the number of VCC tags and the number of passes can be reduced.

The fourth is appropriate noise countermeasures. If noise removal is applied uniformly to the entire character string image, small character elements such as katakana “Ushisotshomin” will also be lost, so too strong noise removal cannot be applied. Therefore, a reservation process is used to realize noise removal in consideration of the adjacent relationship.
At this point, a reservation that does not satisfy a certain size is made as "possible noise". This is reflected in the status. The result is inherited by the VCC tag in step 213, and reflected in the evaluation as a path in steps 214 and 215, thereby realizing noise removal in consideration of the adjacent relationship.

FIG. 6 shows a detailed processing flowchart of the element tag reservation processing. In step 301, the processing mode is interpreted and basic conditions used in the reservation processing are set. That is, the reservation process is controlled based on the character type condition.
For example, it is information such as whether or not a character type representing a minute symbol is specified, and a recognition method to be used among a plurality of character recognition methods. If there is no need to make a reservation (YES in step 302), this process ends. If a reservation is required (NO in step 302), the following processing is performed.

Steps 303 and 304 are repetitive processes for sequentially extracting element tags. The processing range is
In FIG. 28, it is from the top to immediately before the R tag.

In step 305, it is checked whether or not the extracted element tag is to be processed. Here, it is checked whether the tag is a BCC tag. If the tag is a BCC tag, the process proceeds to step 306; otherwise, the process returns to step 303.

In step 306, a temporary VCC tag is prepared and initialized, and in step 307, the extracted BC tag is
Register the C tag as a child of the temporary VCC tag.

In step 308, one-character recognition described later is performed on the temporary VCC tag.

At step 309, the character recognition result is
Reflect on C tag. Originally, the character recognition result is reflected on the VCC tag, but since a common format is used as the tag structure, it can be reflected on the BCC tag.

According to the result of the character recognition described above, the processing content is one of the following three types. (A) When the recognition result is discarded and is not reflected on the BCC tag. (B) When the recognition result is left and reflected on the BCC tag. (C) When the recognition result is corrected and reflected on the BCC tag.

FIG. 36A shows a case where the recognition result is discarded and the BCC
FIG. 36B shows an example in which the recognition result ("2") is not reflected on the tag 0 and is reflected on the BCC tag 1.

The case where the recognition result is rejected or similar (including determination based on the degree of certainty) corresponds to (a). In this case, the status of the recognition result structure in FIG. Specified. In addition to this, in many cases, the position and size are special characters, and the first candidate in the recognition result is not a special character.

On the other hand, the position information of the tag is
If it is determined that the character is a special character by referring to the condition defined in the above, the content of the first candidate in the character recognition result is examined. The case where the first candidate is a special character that matches the estimation result based on the position information corresponds to (b).

If the character code is not a special character but is a predetermined specific character code, the character code of an appropriate special symbol is inserted into the first candidate of the recognition result, and the BCC
Reflect on tags. This corresponds to (c). Thereby, for example, for a small tag determined to be “NO” at a low position, a process of inserting “,” as the first candidate can be realized.

[0127] Such a candidate insertion rule is predetermined in accordance with the characteristics of the one-character recognizing means. Also,
As other rules, depending on what combination of methods is used to recognize characters in the one-character recognizing means, a case where the result is ignored and a case where the result is reflected while remaining are predetermined.

In step 309, processing is performed according to the result of character recognition as described above. In other words, for tags,
The arrangement conditions including the position and size, the character type conditions, and the character recognition result (including the type of recognition) are checked, and the above-described processes (a) to (c) are selected and executed. In the cases of (b) to (c), the presence / absence of permission for integration of tags is also set, and these are set in the status 71 of the tag structure 70 in FIG.

That is, when the tag is not reserved (a)
And (b) and (c), which are related to controlling whether or not to separate contact characters.

On the other hand, when a reservation is made, whether or not to separate the separated characters is controlled according to the presence or absence of the integration permission.

At this point, the information reflected on the BCC tag is transmitted through step 213 in FIG. 5 to step 51 in FIG.
2 to the VCC tag.

If the reservation processing for all element tags has been completed (YES in step 304), step 31
At 0, the special symbol is excluded from the character type condition of the processing mode.

Returning to FIG. 5, in step 210, the virtual boundary point processing section 4 analyzes the SR (singular area) and generates a virtual boundary point sequence. FIG. 7 is a detailed flowchart of a virtual boundary point sequence generation process based on SR (singular region) analysis.

In step 401, the layer priority set in the processing mode of the tag control structure shown in FIG. 26 is checked, and if the BCC tag has priority, the following processing is skipped.

Steps 402 to 404 are loop processing for extracting a valid SR tag from element tags. The processing target at this time is immediately before the R tag in FIG.

At step 405, the SR tag
(Specific region) is taken out. In step 406, all the SS pairs connected to the SR are obtained. In FIG.
Four SSs are connected to the SR, in this case SS
The number of pairs is 16, which is the square of the number.

Steps 407 to 408 are loop processing for sequentially taking out SS pairs. In step 409, a connection pattern of the unique region is obtained as shown in FIG.
In step 410, the necessity of a narrowly defined virtual boundary point sequence is determined according to the connection pattern. At this time, an object that can be substituted by an actual boundary point is not required.

In FIG. 24, the connection pattern is, for example, SS3 as a part of a character, and SS6, SS7 and SS9.
Is a connection pattern that cuts vertically between SS3 and SS6, 7 so that is a part of a character. If the character string direction is horizontal, SS3 is a part of the right character, SS6 and SS7 and SS9 cannot be part of the character on the left, and therefore, a narrowly defined virtual boundary point sequence is not generated for the connection pattern as described above. In this way, no virtual boundary point sequence is generated between series of cross sections that are unnatural in the character string direction.

In step 411, a virtual boundary point sequence segment is obtained based on the connection pattern obtained in step 409. Here, an appropriate virtual boundary point sequence segment is obtained by using the method described above.

In step 412, a virtual boundary point sequence is registered in the form of FIG. 22 with the virtual boundary point sequence segment and connection information.

Returning to FIG. 5, in step 211, the tag processing section 3 selects element tags. The selection is necessary only when the BCC tag and the SS / SR tag are correctly obtained and the SS / SR tag priority is set in the processing mode. In this case, there is an overlap between the BCC tag and the SS / SR tag, and it is necessary to eliminate the overlap.

First, the layer priority is determined. The simplest determination method is a method in which a user specifies a layer priority. This is a processing request such as “prioritize recognition accuracy” or “prioritize processing speed” according to the user's use environment and purpose.

Another determination method is a method in which a layer priority is defined in advance according to a combination of character types, and the layer priority is automatically selected according to character type information designated by the user. This method can reduce the burden on the user.

However, in the above two methods, only one of the layers is used, so that the processing is simple, but it is difficult to balance the processing accuracy and the processing amount.

Therefore, in the present invention, a method of automating the layer priority in order to balance the processing accuracy and the processing amount is considered. As one of the methods, a method using the result of the reservation processing in step 209 will be described below.

Based on the result of the reservation processing in step 209, the element tags are sequentially extracted, and if they are SS tags or SR tags, the BCC tags overlapping with the element tags are extracted. If the BCC tag has been reserved in step 209, the status of the SS / SR tag is set to the removal state while leaving the BCC tag. Otherwise, the status of the BCC tag is set to the removal state, and the SS / SR tag is left as it is. By such processing, when step 211 is completed, the remaining element tags have no duplication.

FIGS. 37A and 37B are diagrams for explaining selection of element tags. In (a), since the BCC tag has been discarded in step 209, the SS tag 0, the SS tag 2, the SR tag 0, and the SR tag 1 remain. In (b), since the BCC tag has been reserved in step 209, the SS tag 1, the SR tag 2, and the SR tag 3 are discarded. FIG.
8 shows the result of selecting the element tags.

At step 212, initial integration of the remaining element tags is performed. This is a process of integrating two element tags of the same type if the overlap in the character string direction is large. For example, when the character string direction is the horizontal direction, deviations, sidelines, and the like are easily integrated. Integration is achieved by copying one child tag to the other child tag, updating the number of children, and removing the former status.

FIG. 39 shows an example of initial integration of element tags. Since the boundary point sequences C5 and C6 overlap in the character string direction, the respective BCC tags having the boundary point sequences C5 and C6 as children are integrated.

However, as a result of the reservation processing in step 209, an element tag for which “integration is not permitted” is not performed in step 212. In step 212, the element tags and the VS tags are sorted again along the character string direction (FIG. 29).

In step 213, the tag processing unit 3
Generates a VCC tag by combining element tags. V
The CC tag means “a candidate that can be one character” as viewed from the character segmentation process, and is represented by one or more combinations of element tags. That is, the element tag and the VS tag are children of the VCC tag.

FIGS. 8 and 9 are flowcharts showing the details of step 213. 8 and 9 assume that the character string is in the horizontal direction (from left to right).
The same processing can be performed in the vertical direction. FIG.
It is a figure showing the generation process of a tag.

In step 501, a VCC tag to be used temporarily in the following processing is prepared. Step 502
Step 504 is a loop process in which tags to be processed (element tags or VS tags) are sequentially taken out and set to the left end. Similarly, steps 506 to 508 take out tags to be processed in order and set the right end This is a loop process. Also, the loop processing for extracting the right end starts from the same element tag as the left end.

By the above-described processing, all combinations of tags that are continuously arranged at the time of sorting in step 212 are created. As described above, since the tags are combined on the condition of continuity, the number of VCC tags is reduced as compared with the case where all combinations of tags are obtained.

The VS tag is not used at either end of the element tag, but is used only when it is sandwiched between the element tags (VCC tag 9 and VCC tag 11 in FIG. 40). Also, in step 508, the retrieved tag is a VS tag, and its status is "Does not allow integration".
If so, the process returns to step 502.

At the time of the processing in step 509, the rightmost element tag extracted in the above loop processing is temporarily
Register as a child tag of the C tag. In step 510,
With reference to predetermined conditions (eg, geometric conditions of characters),
It is determined whether to leave the VCC tag. If a predetermined condition is met (for example, the right end is a VS tag), it is determined in step 510 that it is not a target, and step 50 is performed.
Return to 6. Note that the size in the character string direction and the like can be added as the above-described predetermined condition.

In step 511, one formal VCC tag is taken out and the contents of the temporary VCC tag are copied (registered). That is, as shown in FIG. 40, a VCC tag is added to an element tag.

In step 512, it is checked whether or not the currently added child tag, that is, the element tag at the right end has been reserved. If the child tag has been reserved, the contents relating to the one-character recognition result are inherited. That is, BC which is a child of the VCC tag
When the C tag holds the recognition result (reserved), for example, as shown in FIG.
Give it to C. Further, when the status is “unacceptable” as a result of the reservation, the right end cannot be further advanced, and the process returns to step 502 to take out the next left end. If "integration is permitted" or if the reservation has not been made, the process returns to step 506, and the right end is advanced.

When the extraction of all element tags has been completed for both the left end and the right end (YES in step 503), the flow advances to step 514 in FIG.

Steps 514 and 515 correspond to VC
This is a loop process for sequentially extracting C tags. Step 51
In step 6, one-character recognition described below is performed. In step 517, the result is reflected on the VCC tag. That is, VC
The recognition result is set to the recognition result number 81 of C (FIG. 25). When one-character recognition is completed for all VCC tags, the process proceeds to step 214 in FIG.

Returning to FIG. 5, in step 214, the link generation / path generation unit 6 generates a link for the path tag (VCC tag and VS tag). This is to determine a local adjacency for generating a path. FIG. 41 is a diagram illustrating the generated link.

In step 215, a path (for example, R → V0) is used as shown in FIG.
→ V4 → V7 → V11 → E) (in FIG. 42, V
CC tag 0 is represented as V0 and VCC tag 1 is represented as V1), and the most optimal path is selected. FIG. 43 shows the selected path. Since the above paths and links form a tree structure, an optimal path is selected by applying, for example, DP matching.

At step 216, a specific character re-recognition process is executed for the optimum path selected at step 215. The processing content corresponds to the one-character recognition processing extended to a character string, and most of the processing is shared.

FIG. 10 is a detailed processing flowchart of step 216. Step 601 to Step 603
Is a loop process for sequentially extracting the VCC tags from the optimal path.

In step 603, the recognition result is taken out from the VCC tag, and it is checked whether or not the recognition result is to be subjected to the following processing. In the present invention, when the necessity of re-recognition is indicated in the status of the recognition result structure (FIG. 27), the processing of step 604 and subsequent steps is executed.

In step 604, only necessary images are reconstructed from the VCC tags. Specifically, it can be realized by a method often used for font display and the like. An image can be reconstructed by extracting boundary points from the VCC tag via the element tag, plotting these on the initialized image data, and filling black pixels between the plotted black pixels.

In step 605, adaptive filtering is performed on the character image. When the line is thin, the number of black pixels is increased near the boundary point, and when the line is thick, a filter that removes the black pixel near the boundary point is selected and applied.

In step 606, a sectional sequence graph is extracted and described again from the filtered image. This is the same process as step 203 in FIG.

In step 607, a BCC tag is created based on the newly obtained sectional series graph. This process is the same as step 204 in FIG.

In step 608, one temporary VC
A C tag is prepared, and the newly obtained BCC tag is registered as a child. In step 609, one-character recognition, which will be described later, is performed on a VCC tag having a BCC tag as a child.

In step 610, the presence or absence of an error in the processing from step 605 to step 609 is checked.
If there is no error, in step 611, the recognition result of the VCC tag is replaced, and the re-recognition of the specific character ends.

In steps 604 to 606, access to image data occurs.
Therefore, the importance of managing the upper layer and the lower layer of the cross-sectional series graph with tags having the same expression format, which is the central idea of the present invention, is not lost. Conversely, in the case of the conventional method that does not use the concept of the combination of characteristic elements as in the present invention, for all VCC tags, for example, step 604 is performed between step 710 and step 711 in one-character recognition.
It is necessary to repeat the same processing as. In this case, since the ratio of access to the image data becomes 100%, a considerable increase in the processing amount is suppressed by the present invention.

In step 217, the recognition result obtained by the above processing is set. This is the end of the entire process.

Finally, the processing of one-character recognition in the character recognition section 7 will be described. FIG. 11 illustrates step 308 of FIG.
11 is a detailed processing flowchart of one-character recognition in step 516 of FIG. 9 and step 609 of FIG.

A feature of this processing is that one-character recognition is actually performed while absorbing the difference between the types of element tags. The basic idea is to unify the recognition result into a boundary point sequence expression. For this purpose, if necessary, the boundary point sequence is temporarily reconnected using a narrowly defined virtual boundary point sequence. Furthermore, a plurality of character recognition methods are used together as a recognition method, but the features used in performing the recognition are not obtained from the image, but are obtained from the cross-sectional series graph and the virtual boundary point via a tag. This is obtained from the columns, and thereby realizes the concept of “combination of characteristic elements” of the present invention.

In step 701, a temporary VCC tag (empty VCC tag) is prepared and initialized.

Steps 702 and 703 are repetitive processes, in which all element tags are sequentially extracted from the VCC tag to be processed. Step 704 checks whether the extracted element tag is a BCC tag, and if so, registers it as a child in the temporary VCC tag in step 707. FIG. 44 shows an example where the child is a BCC tag. FIG. 45 is a diagram in which a BCC tag is registered as a child of a temporary VCC tag.

In the example of FIG. 45, since there is one BCC tag, the process proceeds from step 703 to step 710. BC
Since the child of the C tag is the boundary point sequence C1, step 71
No processing is performed in steps 0, 711, and 712, and step 713
Proceed to.

In step 713, the character recognition of the boundary point sequence C1 is performed according to the processing mode, and the result is stored in the BCC tag. Here, multiple character recognition methods (for example,
A matching method described in Japanese Patent No. 2719202) is prepared, character recognition is performed by combining predetermined recognition methods according to the processing mode, and the results are integrated to recognize the recognition result in this step. And The details are not directly related to the present invention, and are therefore omitted. The important point is that the features used in these character recognition methods, such as the features based on the gradient gradient direction of the boundary points, the graph representation of the boundary point sequence, and the skeleton graph representation, are consistent with the cross-sectional series graph. The point is that these features are not directly obtained from the image plane, but are obtained from the sectional series graph via the tag.

In the example of FIG. 45, the processing ends without any processing in step 714. A case where the element tags extracted in step 702 are SS and SR tags will be described with reference to FIG. 46 as an example. In the example of FIG. 46, three element tags (SR
Tag 1, SS tag 0, SR tag 0)
The R tag 1 is taken out (step 706). Since the SR tag 1 is an end point, nothing is performed and the next SS tag 0 is taken out.
Since the tag is an SS tag (step 705), the four boundary points at both ends are stored (step 708). That is, in FIG. 47, two boundary points of the cross section on the SR1 side and two boundary points of the cross section on the SR0 side are stored. The four boundary points are information necessary for temporarily changing the boundary point sequence (step 710).

Finally, the SR0 tag is taken out (step 7)
02). In step 706, the retrieved element tag is S
Check whether the tag is an R tag, and if so, step 7
At 09, a virtual boundary point sequence in a narrow sense is acquired. That is, first, a unique region is obtained via the SR tag, and a connection pattern (FIG. 23) of the unique region is obtained by combining with the cross-sectional series obtained via the SS tag. FIG. 48 shows an example of a connection pattern of the unique region SR0 by SR analysis. Here, it is indicated that both the section number having the start boundary point and the section number having the end boundary point are 2.

Next, in step 210 in FIG. 4, a narrowly defined virtual boundary point sequence obtained in advance in the format of FIG. 22 is collated with the connection pattern of FIG. 23 as a key to obtain a narrowly defined virtual boundary point sequence. (FIG. 47).

At step 706, the answer to NO is VS
This is the case for tags. In step 710, step 70
Only when there are boundary points obtained in step 8 and temporarily stored, the boundary point structures are taken out in the order of connection using these as starting points. Perform a global replacement. That is, the section on the SR0 side of SS0 is replaced with the virtual boundary point sequence (dotted line in FIG. 48).

In step 711, the four boundary points and the temporarily replaced virtual boundary point sequence are traced to create a broadly defined virtual boundary point sequence (FIG. 47). A BCC tag (ie, a VC tag) to be a child is created (FIG. 49). As described above, the cross-sectional series and the singular region in FIG. 46 are separated by the singular region and converted into a boundary point sequence as shown in FIG.
The recognition processing is performed in 3.

In step 712, only when there is a VC tag created in step 711, the
Register as a child of the CC tag.

In step 713, character recognition is performed according to the processing mode, and the result is stored in a VC tag. As the character recognition method, a plurality of recognition methods are prepared. Character recognition is performed by combining predetermined recognition methods according to the processing mode, and the results are integrated to obtain a recognition result of this step.

In step 714, if there is a virtual boundary point sequence temporarily reconnected in step 710,
Return it to the state before connecting it. With the above processing, one-character recognition is performed.

FIG. 50 shows the result when the above processing is applied to the contact character string. Here, it can be seen that the touching character (a) is appropriately extracted from each other via the virtual boundary point sequence (c) generated based on the sectional sequence graph (b). The processing result of (d) is obtained by plotting the boundary points in order to visualize what is expressed by the feature level in an image.

(Reduction of the Number of VCC Tags) The process of generating a VCC tag from an element tag has been described in detail with reference to FIGS. However, in the above-described process of generating a VCC tag, there is a problem that the number of VCC tags increases.

This problem will be described with reference to FIG. In the example shown in FIG. 40, five element tags (two SS tags,
Two SR tags and one BCC tag) and one VS
Thirteen VCC tags have been generated from the tags. As shown in this example, the number of VCC tags increases in proportion to the square of the number of element tags. Therefore, the number of generated VCC tags is suppressed by the method described below, and the processing amount is further reduced.

One point of focus for suppressing the number of VCC tags is the handling of SR tags. In the method described above (the processing of step 213, the processing of steps 501 to 517 in FIGS. 8 and 9), the SR tag is
Since the tags and the BCC tags are handled in the same row, processing waste is likely to occur.

In FIG. 40, the VCC tag numbers 0 and 7 are
Although each is generated from one SR tag, it is extremely unlikely that a single SR tag becomes one character,
In other words, the singular region originally expresses an end point, a branch point, or an intersection. Since such a special point is one character and cannot be considered to be in contact with an adjacent character, the above-described VCC is used. Tag numbers 0 and 7 are VCCs that can be suppressed.

In the above-described method, the unrecognizable VC
Easy to generate C tag. Specifically, this is a case where at least one of the SR tags that should be at both ends when viewed from the SS tag is not included. In such a case, the method of generating the virtual boundary point sequence is not specified, and is therefore processed as unrecognizable, so that the VCC tag is generated but is not substantially used. In FIG. 40, the VCC tag numbers 1, 4, 5, 6, 10, and 11 correspond to the corresponding VC tags.
It is a C tag.

Therefore, the number of VCC tags is reduced by treating the SR tag not as the same tag as the SS tag or the BCC tag but as an auxiliary to the SS tag.

Hereinafter, description will be made with reference to FIGS. In step 213, the tag processing unit 3 generates a VCC tag by combining the element tags. As described above, VC
The C tag means a “candidate that can be one character” as viewed from the character segmentation process, and is represented by one or more combinations of element tags. Here, the SR tag is attached to the SS tag. Treat as something. That is, the tags that are directly targets for generating the VCC tag are the SS tag, the BCC tag, and the VS tag, and only the selected SS tag is supplemented with the SR tag connected thereto.

FIGS. 8 and 9 are detailed flowcharts of the process in step 213, and FIG. 51 is a diagram showing a process of generating a VCC tag based on the above-described new rule.

In step 501, a VCC tag to be used temporarily in the following processing is prepared. Step 502
-Tags to be processed in step 504 (SS tag, B
(CC tag, VS tag) are taken out in order and set to the left end.
Reference numeral 08 denotes a loop process in which tags to be processed are sequentially extracted and set to the right end. Also, the loop processing for extracting the right end starts from the same element tag as the left end.

By the above-described processing, a combination of all tags except for the SR tag among the tags arranged continuously at the time of sorting in step 212 is created. The SR tag is handled as a tag attached to the SS tag as described above. As described above, since the tags are combined on the condition of continuity, V is compared with the case where all combinations of tags are obtained.
The number of CC tags is reduced. Also, compared to the case where the SR tag is handled in the same row as the SS tag and the BCC tag (FIG. 40),
The number of VCC tags is greatly reduced. By the above processing, the number of VCC tags becomes five as shown in FIG. Note that VCC tag number 2 (VCC tag number 8 in FIG. 40) is a combination of SR tag 0 (tag number 5) and SS tag 2 (tag number 4), which is the number shown in FIG. It corresponds to the loop part of “0”, and it is considered that there are SR tags at both ends of the SS tag, so that the VCC tag number 2 is generated.

Since the number of VCC tags is reduced as described above, the number of links that are inevitably generated is also reduced, as shown in FIG. Further, the number of generated paths is greatly reduced as shown in FIG.
C tag 0 is represented as V0, and VCC tag 1 is represented as V1). Also,
The path selection result is as shown in FIG.

The present invention can be realized by software. FIG. 55 shows a system configuration example when the present invention is realized by software. The image to be recognized is input from a scanner or the like, and is taken into the system. Alternatively, image data stored in a hard disk or image data taken into a system via a network is to be recognized.

In the system configuration, the CPU performs a character recognition process by executing the processing steps and processing functions of the above-described embodiment on an input image, and displays a recognition result on a display device.

The program for executing the above processing is C
A program recorded on a recording medium such as a D-ROM is read from a CD-ROM device,
It is executed by installing it in the system, and realizes the processing functions described in the above embodiments. In addition, the above-described program is provided by being downloaded from a server or the like via a communication device or a network in addition to a medium.

[0203]

As described above, according to the present invention, since a sequence of virtual boundary points is generated for a singular area of a sectional sequence graph, a contact character string is appropriately cut out and recognized with high accuracy. can do.

Also, since a virtual boundary point sequence is generated in a unique region where the character string is in contact, the contact character can be separated.

Further, since the virtual boundary point sequence is generated by a predetermined curve generation method, the character curve is smoothed and the recognition accuracy of character recognition is improved.

In addition, since image features having different hierarchies are managed by tags and character candidates are generated by combining the tags, access to image data is reduced.
The speed of the recognition process is increased.

In addition, when tags are combined, the duplication of tags over a plurality of layers is eliminated, so that the number of character candidates can be reduced and the processing amount can be reduced.

Also, since the expression form (logical structure) of the tags is the same, the processing can be shared, thereby simplifying the device configuration.

Further, when generating the VCC tag,
Tags, BCC tags, and VS tags are the main processing targets, and when they are combined, the method of generating the VCC tag by supplementing the SR tag required for the SS tag is adopted, so that the cutout accuracy of characters is impaired. Therefore, generation of a VCC tag can be suppressed.

[Brief description of the drawings]

FIG. 1 is a processing flowchart illustrating an outline of an embodiment of the present invention.

FIG. 2 is a diagram illustrating a hierarchy of tags according to the present invention.

FIG. 3 is a diagram showing a configuration of an example of the present invention.

FIG. 4 is a flowchart of the entire processing according to the processing of the present invention.

FIG. 5 is a processing flowchart continued from FIG. 4;

FIG. 6 is a detailed processing flowchart of element tag reservation processing.

FIG. 7 is a detailed flowchart of virtual boundary point sequence generation processing by SR (singular region) analysis.

FIG. 8 is a detailed processing flowchart of generating a VCC tag.

FIG. 9 is a processing flowchart continued from FIG. 8;

FIG. 10 is a detailed processing flowchart of specific character re-recognition.

FIG. 11 is a detailed processing flowchart of one-character recognition.

FIG. 12 is a diagram showing a relationship between line graphic image data (original image data), a sectional series graph, and a skeleton line (core line).

FIG. 13 is a diagram illustrating an example of a character string image.

FIG. 14 is a diagram showing a boundary point and a cross section of a cross-sectional series graph for the image of FIG.

15 is a diagram showing a cross-sectional sequence and a unique region of a cross-sectional sequence graph for the image of FIG.

FIG. 16 is a diagram showing an expression format of a boundary point.

FIG. 17 is a diagram showing an expression format of a boundary point sequence.

FIG. 18 is a diagram showing an expression form of a cross section.

FIG. 19 is a diagram showing an expression format of a sectional series.

FIG. 20 is a diagram showing an expression form of a unique region.

FIG. 21 is a diagram illustrating a virtual boundary point sequence segment.

FIG. 22 is a diagram showing an expression format of a virtual boundary point sequence.

FIG. 23 is a diagram illustrating a connection pattern of a unique region.

FIG. 24 is a diagram illustrating a state of connection of unique regions.

FIG. 25 is a diagram showing an expression format of a tag.

FIG. 26 is a diagram showing a tag control structure.

FIG. 27 is a diagram showing an expression format of a recognition result.

FIG. 28 is a diagram showing how a tag array is used.

FIG. 29 is a diagram showing a state of a tag array and a sort result array.

FIG. 30 is a diagram illustrating an example of a character string image to be processed.

FIG. 31 is a diagram modeling a tag structure.

FIG. 32 is a diagram showing a result of sorting tag numbers.

FIG. 33 is a diagram showing a lower layer of the sectional sequence graph and a VS tag to be added.

FIG. 34 is a diagram illustrating an example of a negative blank.

FIG. 35 is a diagram showing a result of sorting by adding a VS tag to an element tag.

FIG. 36 is a diagram illustrating a reservation process of an element tag.

FIG. 37 is a diagram illustrating selection of an element tag.

FIG. 38 is a diagram showing the result of selecting element tags.

FIG. 39 is a diagram illustrating an example of initial integration of element tags.

FIG. 40 is a diagram showing a process of generating a VCC tag.

FIG. 41 is a diagram showing a generated link.

FIG. 42 is a diagram illustrating a generated path.

FIG. 43 is a diagram showing a selected path.

FIG. 44 is a diagram illustrating an example in which a child of a VCC tag is a BCC tag.

FIG. 45 is a diagram in which a BCC tag is registered as a child of a temporary VCC tag.

FIG. 46 is a diagram illustrating an example in which a child of a VCC tag is an SS / SR tag.

FIG. 47 is a diagram for explaining processing of SS and SR tags.

FIG. 48 is a diagram showing an example of a connection pattern of a unique region SR0.

FIG. 49 is a diagram illustrating generation of a VC tag having a virtual boundary point sequence in a broad sense as a child.

FIGS. 50A to 50D are diagrams showing processing results according to the present invention.

FIG. 51 is a diagram illustrating another generation process of the VCC tag.

FIG. 52 is a diagram showing a link generated from FIG. 51.

FIG. 53 is a diagram showing a generated path.

FIG. 54 is a diagram showing a path selection result.

FIG. 55 is a diagram illustrating a configuration example when the present invention is executed by software.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input part 2 Section series processing part 3 Tag processing part 4 Virtual boundary point processing part 5 Memory 6 Link generation / path generation part 7 Character recognition part 8 Control part 10 Boundary point structure 20 Boundary point sequence structure 30 Cross-sectional structure 40 Cross-section sequence structure 50 Singular region structure 60 Virtual boundary point sequence structure 70 Tag structure

Claims (19)

[Claims] 1. A character recognition method for recognizing a character using a cross-sectional sequence graph as a feature of a character image, comprising: extracting a cross-sectional sequence graph from a character string image; analyzing a singular region of the cross-sectional sequence graph; A virtual boundary point sequence is generated in the singular region based on the result, a character candidate is generated by combining structural elements of the cross-sectional series graph, and a necessary virtual boundary point sequence is supplemented for the generated character candidate. A character recognition method characterized by recognizing a character as a single character and recognizing a character string based on an adjacency relationship between the recognized character candidates. 2. The character recognition method according to claim 1, wherein the unique region in which the virtual boundary point sequence is generated includes a region where a character string is close to or in contact with the character string. 3. The generation position of the virtual boundary point sequence is determined on the basis of the connection order and position of the singular region and a series of cross-sections connected to the singular region.
The character recognition method described. 4. The method according to claim 1, wherein the direction of the character string is referred to when the virtual boundary point sequence is generated, and the virtual boundary point sequence is not generated between the cross-sectional series that does not correspond to the character string direction. The character recognition method according to 2, 3 or 4. 5. The character recognition method according to claim 1, wherein the virtual boundary point sequence is generated using a predetermined curve generation method. 6. The character recognition method according to claim 1, wherein the structural element is divided into a first layer and a second layer, and each is managed by a tag. 7. A first layer of the structural element is a section sequence and a singular region, and the section sequence is managed by a first tag,
The character recognition method according to claim 1, wherein the unique area is managed by a second tag. 8. The character recognition method according to claim 1, wherein the second layer of the structural element is a sequence of boundary points, and the sequence of boundary points is managed by a third tag. 9. The character recognition method according to claim 1, wherein a blank area between the character strings is managed by a fourth tag. 10. A character candidate is generated by combining the first, second, third, and fourth tags, and the generated character candidate is managed by a fifth tag. Item 7. The character recognition method according to Item 7, 8 or 9. 11. When combining the first, second, and third tags, for the same structural element, either the tag of the first layer or the tag of the second layer is used, and duplication of the tags is performed. The character recognition method according to claim 10, wherein the character recognition is canceled. 12. A character candidate is generated by combining the first, third, and fourth tags and supplementing the second tag required for the first tag, and generating the character candidate. 8. The management by a fifth tag.
8. The character recognition method according to 8 or 9. 13. Creating a link between the fifth tags,
13. The character recognition method according to claim 10, wherein a path is generated between the generated links, and a character string is recognized and output by selecting an optimum path from the generated paths. . 14. The first section graph of a cross-sectional series graph from a character string image
A function of extracting a cross-sectional series and a singular region, which is a layer, and a function of extracting a boundary point sequence, which is a second layer of a cross-sectional sequence graph, and generating a character candidate by combining the cross-sectional sequence with a singular region and a boundary point sequence Computer-readable program recording a program for causing a computer to implement a function of recognizing the generated character candidate as one character, and a function of recognizing a character string based on the adjacent relationship between the recognized character candidates. Possible recording medium. 15. A function of analyzing the singular region and generating a virtual boundary point sequence in the singular region based on the analysis result, and a character candidate consisting of a sectional sequence and a singular region when recognizing the one character. 15. The computer-readable recording medium according to claim 14, which stores a program for causing a computer to realize a function of supplementing with a necessary virtual boundary point sequence and converting the boundary point sequence. 16. The method according to claim 16, further comprising:
For realizing a function of generating a tag, a second tag for managing the singular region, a third tag for managing the boundary point sequence, and a fourth tag for managing a space between the character strings. 16. The recording medium according to claim 14, wherein the program is recorded and is readable by a computer. 17. A program for causing a computer to realize a function of generating a fifth tag for managing a character candidate generated by combining the first, second, third, and fourth tags. 17. The recording medium according to claim 16, which is readable by a computer. 18. A fifth tag for managing a character candidate generated by combining the first, third, and fourth tags and supplementing the second tag required for the first tag is generated. 2. A computer-readable recording program for causing a computer to realize the function of performing the function.
6. The recording medium according to 6. 19. A computer readable recording program for causing a computer to realize a function of managing the first, second, third, and fifth tags with the same logical structure. Recording medium.