JP2002099876A - Character recognition device, character recognition method, and storage medium with character recognition program stored therein - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition device strong to a break of character stroke. SOLUTION: A profile information extraction part 2 approximates the profile of an input character pattern accumulated in a character pattern input part 11 with a diagonal line, and extracts profile information including a directed segment of line that is this diagonal line. A structure matching part 4 collates the extracted profile information of the input character pattern with the profile information of typical character pattern preliminarily stored in a dictionary 3 for each of character categories with the directed segment of line as unit, and determines and outputs the distance between the input character pattern and the typical character pattern as a distance value. A character recognition result output part 5 determines the character category of the most certain input character pattern from the distance value outputted by the structure matching part 4 for each of the character categories and outputs a character code.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a character recognition device and a character recognition method, and more particularly to a character recognition device and a character recognition method for recognizing a binary character pattern using a structural analysis technique.

[0002]

2. Description of the Related Art In an optical character reading apparatus for automatically reading characters, a character recognition technology for recognizing various character types such as alphanumeric characters and kanji for print characters and handwritten characters has been developed. In such a character recognition technology, there are roughly divided two methods, a pattern matching method and a structure analysis method. The pattern matching method is very excellent in recognizing characters with little deformation, such as print characters. Various reports have been reported that pre-processing for uniforming the character size, that is, so-called normalization processing, is highly effective even for handwritten characters having large deformation. However, in a pattern matching method, a handwritten character having a similar shape may be misread. For example, characters having similar overall shapes, such as katakana characters "shi" and "n", may be misread each other. A structure analysis method is effective for recognizing a handwritten character having such a similar shape. Katakana "shi" and "n"
In the case of (1), the structure analysis method individually registers information of all character strokes and contours in a dictionary, and matches the corresponding character strokes and contours between the dictionary and the input character pattern to enable correct identification.

An example of a conventional character recognition apparatus using a structural analysis technique is disclosed in Japanese Patent Laid-Open Publication No. Hei 7-175895. This conventional character recognition apparatus identifies the number of contours constituting a character, the number of line segments obtained by dividing each contour into concave portions and convex portions of the figure, line segment shape features, inter-contour features, and line segment features. Character recognition is performed by comparing a representative character to be identified stored in the dictionary with an input character pattern. Japanese Patent Application Laid-Open No. 8-77293 discloses another example of a character recognition device. This character recognition device compares the position of a hole extracted from the inclusive relation of a circumscribed rectangle of a contour constituting a character, the position of an end point of a character pattern, the curvature, and the stroke direction with a dictionary as a feature amount. However, these prior arts are liable to be rejected or misread when an input character pattern is interrupted. In Japanese Patent Application Laid-Open No. 8-77293, if it is considered that a part of a hole of a character pattern is cut off due to a distance between end points or the like, a feature obtained by connecting the cut portion is referred to as a second feature. Although the same features are used for matching with the dictionary pattern, not all the problems that may occur in the portions other than the holes are not solved.

[0004]

The above-described conventional character recognition device and character recognition method have a problem that when an input character pattern is interrupted, it is easily rejected or misread. If it is considered that a part of the hole of the character pattern is cut due to the distance between the end points, etc.,
There is public information that uses two types of features, which are the features obtained by connecting the cut portions before the feature, for matching with the dictionary pattern. However, there is a problem that various breaks other than holes are unresolved. is there.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition apparatus and a character recognition method that are robust against breaks in character strokes in order to eliminate such conventional disadvantages.

[0006]

A character recognition device according to the present invention comprises: a character pattern input section for receiving a character pattern sampled in a lattice and quantized into a binary form and storing the character pattern as an input character pattern; A contour information extraction unit for approximating the contour of the input character pattern stored in the input unit and extracting contour information including a directed line segment which is a polygonal line when the contour is approximated; and a representative for each character category. A dictionary in which the contour information including the directed line segment of the character pattern is stored in advance together with a character code corresponding to the character category, and the contour information of the input character pattern extracted by the contour information extracting unit. The outline information of the representative character pattern stored in the dictionary and the directed line included in the outline information for each of the character categories And a structure matching unit that calculates a distance between the input character pattern and the representative character pattern and outputs the distance as a distance value, and the distance value output by the structure matching unit for each of the character categories. And a character recognition result output unit for obtaining the character category of the input character pattern that is likely.

The outline information extraction unit of the character recognition device of the present invention traces the outline of the input character pattern counterclockwise or clockwise with respect to the input character pattern stored in the character pattern input unit, and performs the outline trace of the input character pattern. A contour tracing unit for extracting a coordinate value sequence, and a polygonal line approximation of the contour of the input character pattern from the coordinate value sequence of the contour of the input character pattern extracted by the contour tracing unit. A primitive segment generator that extracts each directed segment as a primitive segment, and the width and height of each of the primitive segments of the input character pattern extracted by the primitive segment generator are set in advance. Each is converted to a fixed size,
The normalization processing unit that outputs the converted outline information including each of the primitive segments to the structure matching unit, and the outline information extraction unit of the character recognition device according to the present invention is configured such that the normalization processing unit A convex segment indicating a set of primitive segments forming a convex portion in the input character pattern from each of the primitive segments of the input character pattern that has been normalized and the input character pattern being a portion other than the convex segment A concave / convex segment generation unit that generates a concave segment indicating a set of the primitive segments that form a concave portion inside, and a corner within a predetermined angle of the concave segment and the convex segment generated by the concave / convex segment generation unit are Check to see if there is A corner division processing unit that divides the segment having the corner at the position and generates the divided segment and the concave segment and the convex segment without the corner as a contour segment; and the contour generated by the corner division processing unit. A new coordinate value sequence is generated by sampling the coordinate value sequence corresponding to the contour segment at predetermined intervals according to the segment and the coordinate value sequence of the contour of the input character pattern extracted by the contour tracking unit. A standard outline segment generating unit for outputting the standard outline segment as a part of the outline information together with a character code corresponding to the input character pattern, and storing the standard outline segment in the dictionary in advance.

In the character recognition device of the present invention, the structure matching unit may include an i-th (where i = 1 to 1) primitive primitive segment included in the input character pattern included in the input character pattern obtained from the contour feature extracting unit. Of the primitive segment Vi) and all the standard contour segments included in any character pattern corresponding to any of the character codes in the dictionary, and a first distance calculation is performed. A character code corresponding to a character pattern that is within the threshold when within a predetermined threshold, and the standard contour segment that is within the threshold included in a character pattern corresponding to the character code; And outputs the primitive segment in any of the character patterns when the calculated value is not within a predetermined threshold value. And a candidate segment detecting unit that notifies that there is no standard contour segment corresponding to Vi, and between each of the standard contour segments and the primitive segments Vi of one or more of the character codes output from the candidate segment detecting unit. Performs the second distance calculation and outputs a distance value, respectively, and the primitive segment V among the distance values output by the inter-segment distance calculation unit
i, the standard contour segment corresponding to the smallest distance value is detected, and a best segment detection unit that outputs the distance value; and a distance value output by the best segment detection unit is aggregated to obtain the input character pattern. And a character pattern distance calculating unit for calculating a distance value between characters with the character pattern in the dictionary.

According to the character recognition method of the present invention, a first step of receiving a character pattern sampled in a lattice and quantized into a binary form and storing it as an input character pattern, and the outline of the stored input character pattern A second step of extracting outline information including directed line segments as respective broken lines when the broken line is approximated, and a representative in each of the character categories stored in advance together with a character code corresponding to the character category. The contour information including the directed line segment of the typical character pattern and the contour information of the input character pattern extracted in the second step are included in the contour information for each of the character categories. A third step of collating in units of line segments to determine a distance between the input character pattern and the representative character pattern;
A fourth step of obtaining the character category of the input character pattern that is most likely than the obtained distance;
Contains.

The second step of the character recognition method according to the present invention is characterized in that the input character pattern accumulated in the first step is traced counterclockwise or clockwise to obtain a contour of the input character pattern. Extracting a coordinate value sequence, and approximating the contour of the input character pattern from the coordinate value sequence of the extracted contour of the input character pattern with a polygonal line. Extracting each as a primitive segment, and converting each of the primitive segments of the extracted input character pattern so that the width and height of the input character pattern are each a predetermined constant size, and Outputting the contour information including the respective primitive segments, Which comprise.

A recording medium on which the character recognition program of the present invention is recorded receives a character pattern sampled in a lattice and quantized into a binary form, and stores it as an input character pattern.
And a second process of extracting outline information including a directed line segment that is a broken line when the outline of the input character pattern thus accumulated is broken and the broken line is approximated.
The outline information including the directed line segment of the representative character pattern in each of the character categories stored in advance together with the character code corresponding to the character category, and the outline of the input character pattern extracted in the second processing And a third process of comparing information with each of the character categories in units of the directed line segment included in the outline information to determine a distance between the input character pattern and the representative character pattern. And a fourth process for obtaining the character category of the input character pattern that is most probable from the distance, and a character recognition program for causing a computer to execute the fourth process.

[0012]

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the character recognition device of the present invention.

The present embodiment shown in FIG. 1 includes a character pattern input unit 1 which receives a character pattern sampled in a lattice and quantized into a binary form and stores it as an input character pattern. A contour information extracting unit 2 for approximating the outline of the accumulated input character pattern and extracting outline information including directed line segments as respective broken lines when the broken line approximation is performed, and representative characters for each character category A dictionary 3 in which contour information including a directed line segment of a pattern is stored in advance together with a character code corresponding to a character category, and contour information of an input character pattern extracted by the contour information extracting unit 2 and a representative stored in the dictionary 3 The contour information of the character pattern is collated with the directed line segments included in the contour information for each of the character categories, and the input character pattern and the representative character pattern are compared. The structure matching unit 4 calculates a distance and outputs the distance as a distance value, and a character recognition result output unit 5 calculates a character category of an input character pattern that is most probable from the distance value output by the structure matching unit 4 for each character category. ing.

FIG. 1 also shows a recording medium 6 for storing a character recognition program used in the character recognition device. The character recognition device reads the character recognition program from the recording medium 6 and executes the character recognition program. To recognize characters.

Next, the operation of the character recognition apparatus according to the present embodiment will be described in detail with reference to FIGS.

FIG. 2 is a diagram showing an example of the operation of the embodiment of the present invention and an example of the character recognition method of the present invention.

FIG. 3 is a block diagram showing one embodiment of the contour information extracting unit.
A contour tracing unit 7 for tracing the input character pattern stored in the character pattern input unit 1 counterclockwise or clockwise to extract a coordinate value sequence of the contour of the input character pattern.
From the coordinate value sequence of the contour of the input character pattern extracted by the contour tracing unit 7, the contour of the input character pattern is approximated by a polygonal line, and each of the directed segments that are the polygonal lines obtained when the polygonal approximation is performed is extracted as a primitive segment. The primitive segment generation unit 8 and the primitive segments of the input character pattern extracted by the primitive segment generation unit 8 are converted into a predetermined size (for example, 40 × 4) in which the width and the height of the input character pattern are respectively predetermined.
0), and a normalization processing unit 9 that outputs contour information including the converted primitive segments to the structure matching unit 4, and a primitive of the input character pattern converted by the normalization processing unit 9. The center position calculating unit 10 determines the midpoint coordinates of the segment and outputs the obtained midpoint coordinates to the structure matching unit 4 as a part of the outline information. The contour information extracting unit 2 also includes a convex segment indicating a set of primitive segments that form a convex portion in the input character pattern from each primitive segment of the input character pattern converted and normalized by the normalization processing unit 9. A concave / convex segment generating unit 11 for generating a concave segment indicating a set of primitive segments forming a concave portion in the input character pattern, which is a part other than the convex segment, a concave segment and a convex generated by the concave / convex segment generating unit 11 It is checked whether or not the segment has a corner within a predetermined angle (for example, 90 degrees). When there is a corner, the segment having the corner is divided at the corner position, and the divided segment and the concave segment having no corner are divided. Corners that generate convex segments as contour segments A predetermined number of coordinate value sequences (corresponding to the contour segments) are obtained from the division processing unit 12 and the outline segment generated by the corner division processing unit 12 and the coordinate value sequence of the outline of the input character pattern extracted by the outline tracking unit 7 ( (For example, 1 to 3)
Every other sampling to generate a new coordinate value sequence and use it as a standard outline segment. This standard outline segment is output as a part of outline information together with a character code corresponding to the input character pattern and stored in the dictionary 3 in advance. And a unit 13.

FIG. 4 is a block diagram showing one embodiment of the structure matching unit.
It corresponds to the i-th primitive segment Vi (where i = 1 to the number of primitive segments included in the input character pattern) included in the input character pattern obtained from the contour feature extraction unit, and an arbitrary character code in the dictionary 3. The first distance (distance between the midpoint of the primitive segment and the end point of all the directed segments constituting the standard contour segment) between all the standard contour segments {RSk} included in the arbitrary character pattern is calculated. When the distance value is within a predetermined threshold value, the character code corresponding to the character pattern within the threshold value and the threshold value included in the character pattern corresponding to the character code are within the threshold value. The standard contour segment RSk is output to the inter-segment distance calculating unit 15 and any character is output when the calculated value is not within a predetermined threshold value. A candidate segment detector 14 that notifies the character pattern distance calculator 17 that there is no standard contour segment RSk corresponding to the primitive segment Vi in the turn; and one or more character codes of the character codes output from the candidate segment detector 14. The second between each standard contour segment RSk and primitive segment Vi
, And outputs the distance value, and the distance value corresponds to the smallest distance value between the primitive segment Vi among the distance values output by the inter-segment distance calculation unit 15. A best segment detecting unit 16 that detects the standard contour segment RSk and outputs its distance value, and a distance value output by the best segment detecting unit 16 is compiled to calculate the distance between the character of the input character pattern and the character pattern in the dictionary 3. A character pattern distance calculating unit 17 for calculating a distance value;

FIG. 5 is a diagram showing an example of an input character pattern.

FIG. 6 is a diagram showing an example of primitive segment extraction.

FIG. 7 is a diagram showing an example of the generation of an uneven segment.

FIG. 8 is a diagram showing an example of standard contour segment generation.

FIG. 9 is a diagram showing an example of contour tracking.

FIG. 10 is a diagram showing an example of primitive segment generation.

FIG. 11 is a diagram showing an example of calculating the distance between a primitive segment and a standard contour segment.

FIG. 12 is a flowchart showing an example of the operation for creating a dictionary.

Referring to FIG. 1, a description will be given with reference to FIG. 2. A character pattern input unit 1 receives a character pattern sampled in a lattice and quantized into binary, and stores it as an input character pattern (S21). The contour information extracting unit 2 approximates the outline of the accumulated input character pattern with a polygonal line, and extracts outline information including directed line segments as respective polygonal lines when the polygonal line is approximated (S22). 4, the outline information including the directed line segment of the representative character pattern in each of the character categories stored in advance together with the character code corresponding to the character category, and the outline of the input character pattern extracted in the second step (S22). Information and the directed character segment included in the contour information for each character category, and the distance between the input character pattern and the representative character pattern is compared. Calculated (S23), the character recognition result output section 5 obtains and outputs a character category the most likely input character pattern from the distance obtained in the third step (S24). Here, the second step (S22)
Performs contour tracing clockwise or clockwise with respect to the input character pattern accumulated in the first step (S21) to extract a coordinate value sequence of the contour of the input character pattern, and The contour of the input character pattern is approximated by a polygonal line from the coordinate value sequence of the outline, and each directed line segment that is a polygonal line when the polygonal line is approximated is extracted as a primitive segment, and each primitive segment of the extracted input character pattern is extracted. The width and height of the input character pattern are respectively converted so as to have a predetermined fixed size, and contour information including each of the converted primitive segments is output.

Referring to FIG. 1 in more detail, the character pattern input section 1 stores an input interface and an input character pattern which receive a character pattern (input character pattern) sampled in a lattice and quantized into binary. Includes memory. Here, the input character pattern is
A partial image that is sampled in a grid pattern obtained by photoelectrically converting a document by an image scanner or the like and is quantized into two or more binary values, and is extracted in units of one character and quantized into two values. It is. In the input character pattern, the pixel value of the background portion is set to 0 and the pixel value of the character portion is set to 1. The pixel having the value 0 is called a white pixel, and the pixel having the value 1 is called a black pixel. In the present embodiment, as described above, a description will be given of a digitized character pattern in which the pixel value of the background portion is 0 and the pixel value of the character portion is 1 as a matter of course. Even in the case of an inverted image in which the pixel value of the character portion is 0, a device similar to that of the present embodiment can be realized by exchanging the determination results for 1 and 0 in the pixel value determination processing unit.

The contour information extraction unit 2 has the configuration shown in the description of FIG. 3, and the contour tracking unit 7 detects, by contour tracing, a contour pixel which is a pixel in contact with the background among the pixels constituting the figure. I do. A pixel on the contour adjacent to the pixel of interest (contour pixel) among the eight neighboring pixels of the pixel of interest on the figure contour
Is detected counterclockwise (or clockwise), and all contour pixels are detected while repeating the process of moving the pixel of interest to the detected pixel. FIG. 9 shows this state. In FIG. 9, a hatched pixel indicates a tracked pixel, an X-marked pixel indicates a target pixel, and an arc arrow indicates an order in which the next contour pixel is found from adjacent pixels. The coordinates of the detected contour pixels are stored in the order of detection. This is the coordinate value sequence of the contour pixels. The details of the contour tracking are also introduced on pages 353 to 360 of the document “Digital Image Processing” (Kindai Kagaku).

The primitive segment generation unit 8 approximates the coordinate value sequence of the outline pixels output from the outline tracking unit 7 by a directed line segment and outputs the result. The direction of each directed line segment is a direction according to the order of the coordinate value sequence of the contour pixels. Various approximation algorithms can be used. In this embodiment, as shown in FIG. 10, a method of repeatedly dividing until the distance dn between the contour and the line segment becomes equal to or smaller than a predetermined threshold value is used. Used. This method is described, for example, in the document “Image Pr
processing, Analysis and Ma
chine Vision "by Milan Sonka,
VaclavHlavac and Roger B
oil (CHAPMAN & HALL), pp. 209-210. This will be described by taking an example. For example, a coordinate value sequence Pi (xi, yi) of contour pixels obtained by a contour tracking method is represented by P1 (1, 1), P2
(2, 1), P3 (3, 1), P4 (4, 2), P5
(5, 3), P6 (6, 4), P7 (5, 4), P8
When (4, 4), P9 (3, 4), and P10 (2, 4) are used, a line segment is approximated in the following steps to generate a primitive segment. However, the threshold value Th is set to 1. Step 1: A directed line segment L (1, 10) connecting the start point and the end point, that is, P1 and P10 is created. P1 to P10
The direction leading to is positive. Step 2: Measure the distances between all the contour pixels from P1 to P10 and the directed line segment L (1, 10) in the coordinate value sequence of the contour pixels, and determine the maximum pixel (point) and the distance at that time. Ask. In this case, the dividing point is P
6, and the distance d1 between the line segment L (1, 10) and P6 at that time is d1 = SQRT ｛L (1, 6) · L (1,
6) −L (1,6) · L (1,10) / (L (1,1
0) · L (1,10))｝ ≒ approximately 4.123. Here, “•” represents an inner product operation symbol, L (1, 6) represents a directed line segment from the pixels P1 to P6, and SQRT ｛｝ represents a square root operation. Step 3: Since d1> Th, division is performed at the pixel P6, and the coordinate value strings P1 to P6 and P6 of the contour pixels are obtained.
Create ~ P10. Step 4: The distance between the directed line segments L (1,6) and L (6,10) is the same as in step 2 for the coordinate value strings P1 to P6 and P6 to P10 of the two divided outline pixels. Ask for. As a result, P1 to P6 and L
Distance d2 to (1,6) d ≒ 1.925 (P3 is the pixel with the maximum distance), P6 to P10 and L (6,10)
Distance d3 = 0. Step 5: Since d2> Th, the coordinate value strings P1 to P6 are divided by P3. Step 6:
Since the maximum distance between all the divided coordinate value strings (segments) and the corresponding coordinate value strings has become 1 or less, the division ends. The generated primitive segment is L (1,
3), L (3, 6) and L (6, 10).

The normalization processing section 9 receives the primitive segment output from the primitive segment generation section 8 as an input, and generates a coordinate value sequence of a contour such that the width and height of the character pattern become predetermined constant values. Scale transform primitive segments. For example, if the width of the character pattern is W and the height is H, and both the predetermined width and the predetermined height are F, the coordinate value sequence of the outline and the x-coordinate values of the start and end points of the primitive segment are all represented by W / F times,
All y coordinate values are multiplied by H / F. This operation is called normalization.

The concavo-convex segment generation section 11 is as described in the description of FIG. 3. For example, from the character patterns shown in FIGS. 5 and 6, S1, S2, and S shown in FIG.
3, four contour segments of S4 are generated.

When there is a corner in the concave segment or the convex segment output from the concave / convex segment generating unit 11, the corner division processing unit 12 divides the concave segment or the convex segment at the corner position. The resulting segment is defined as a contour segment. For example, in FIG. 7, the concave segment S3 is divided at the point P. That is,
A contour segment is obtained by dividing the concave and convex segment in FIG.

The standard outline segment generation unit 13 receives as input the outline segment output from the corner division processing unit 12 and the coordinate value sequence of the outline output from the outline tracking unit 7.
After detecting the outline coordinate value sequence corresponding to the outline segment, the outline coordinate value sequence is normalized, and further, for example, 1 to 3
A new coordinate value sequence is generated by thinning out coordinate values for each distance of about a pixel. The direction of the arrangement order of the coordinate value sequence is given to this coordinate value sequence, and a sequence of directed segments is output as a standard contour segment. For example, from the contour segment of FIG. 7, five standard contour segments as shown in FIG.
1, RS2, RS3, RS4, and RS5 are output.

The center position calculation unit 10 calculates and outputs the center coordinate values of the normalized primitive segments (directed line segments) output from the normalization processing unit 9.

The dictionary 3 stores standard contour segments output from the contour information extracting unit 2 together with character codes for each character category.

The structure matching section 4 has the structure shown in the description of FIG. 4, and the candidate segment detecting section 14 outputs an input character pattern and an arbitrary character code corresponding to an arbitrary character code stored in the dictionary 3. In matching with the pattern, the distance is calculated between the i-th primitive segment Vi included in the input character pattern and all the standard contour segments {RSk} included in the arbitrary character pattern. However, the distance calculation is performed on a one-to-one basis between the midpoint of the primitive segment Vi and the end points of all the directed segments constituting the standard contour segment RSk. When the distance is within a threshold value (for example, depending on the normalized size of the character pattern (the value of F in the normalization processing unit 9), it is about 10 pixels when the size is about 40 × 40 pixels). The inter-segment distance calculation unit 15 is notified of the character code corresponding to the character pattern within the threshold value and the standard contour segment RSk included in the character pattern corresponding to the character code and within the threshold value. When this distance is not within the predetermined threshold value, it notifies the inter-character-pattern distance calculation unit 17 that there is no standard contour segment RSk corresponding to the primitive segment Vi in an arbitrary character pattern. The above operation is performed for all the character codes in the dictionary 3.

The inter-segment distance calculation unit 15 performs a second distance calculation between the primitive segment Vi and each of the standard contour segments RSk of one or more character codes notified from the candidate segment detection unit 14. This calculation will be described with reference to FIG. In FIG. 11, V9 is a primitive segment, RS2 is directed segments u21, u
22, u23, u24, u25, u26, u27, u2
8, a standard contour segment composed of u29.
The distances from the end points (or start points) of all the directed line segments constituting the standard contour segment RS2 to the points (intersection points) perpendicularly intersecting the primitive segment V9 are d21, d22, d23, d24, respectively. d25, d2
6, d27. Here, Primitive Segmet V9
If it intersects a straight line outside the range, the distance calculation is not performed. The distance D (9,2) between the primitive segment V9 and the standard contour segment RS2 at this time is D (9,2)
= (1 / N) ｛d21 + d22 + d23 + d24 + d2
5 + d26 + d27 + α (| Δθ21 | + | Δθ22 |
+ | Δθ23 | + | Δθ24 | + | Δθ25 | + | Δθ
26 | + | Δθ27 |)｝ (Equation 1). Where N
= 7. That is, N is the number of directional segments constituting the standard contour segment RS2 which are perpendicular to the primitive segment V9 and intersect within the range of V9.
Δθ2j is defined by a directed segment u2j and a primitive vector V
9 is the size of the angle. Α is a coupling coefficient obtained experimentally. After the distance calculation, the standard primitive segment RSk writes a collated mark on the directed line segment whose distance from the primitive segment could be obtained and writes it back to the dictionary 3. The distance value D (i, k) between the primitive segment Vi of the input character pattern and the standard contour segment RSk of an arbitrary character code is calculated and output. The above operation is performed for all the character codes in the dictionary 3.

The best-segment detecting unit 16 finishes calculating the distance value between all the standard contour segments RSk notified from the candidate segment detecting unit 14 by the inter-segment distance calculating unit 15 and calculates the distance from the inter-segment distance calculating unit 15. When the value D (i, k) has been received, the received distance value D
Standard contour segment R taking the minimum distance in (i, k)
Skmin is detected, and the distance value at this time is output. After the detection, a standard contour segment RSk (k
(Other than kmin), all the verified marks added by the inter-segment distance calculating unit 15 are cleared.

The inter-character-pattern distance calculating section 17 accumulates the distance values output from the best segment detecting section 16 to obtain the inter-character pattern distance. However, if no candidate is found by the candidate segment detection unit 14, the penalty evaluation value B (here, the penalty evaluation value B is a match between the input character pattern and the standard outline segment of the character code in the dictionary 3) Is a value for evaluating the length of the mismatched segment. If the sum of the lengths of the mismatched segments is equal to or greater than the threshold value, the character recognition result is rejected in the character recognition result output unit 5. There are two types, those included in the input character pattern and those included in the standard outline segment of the character code in the dictionary 3. The sum of the two is used as the penalty evaluation value. In the candidate segment detection unit 14,
A primitive segment that does not correspond to any of the standard contour segments of the character code in the dictionary 3 (has a long distance) is a mismatched primitive segment. Among the standard contour segments of the character codes in the dictionary 3, those that did not correspond to any of the primitive segments Vi of the input character pattern, or those that corresponded only at a very small portion to the mismatched standard contour segments Become. ) Is added to a value proportional to the length of the primitive segment Vi. The value proportional to the length of the primitive segment is, for example, a relative value when the sum of the primitive segment lengths of the input character pattern is normalized to 1. In the standard contour segment of the character pattern in the dictionary 3, G% of a directed segment that does not have a matched mark and that has a matched mark although the matched mark exists (where G is a constant given in advance, For example, even if there is less than 30%), a value proportional to the length of the standard contour segment is added to the penalty evaluation value B. Among the standard contour segments included in the character pattern in the dictionary 3, a value proportional to the length of the standard contour segment having no or almost no matched mark is, for example, a directional value that constitutes all the contour line segments included in the character pattern. The sum of the lengths of the line segments is a relative value when normalized to 1.

The character recognition result output unit 5 outputs a character code in which the distance between character patterns output from the character pattern distance calculation unit 17 of the structure matching unit 4 has a minimum value. However, when the penalty evaluation value output from the character pattern distance calculation unit 17 is equal to or larger than the threshold value, it is determined to be rejected, and a code indicating rejection is output. The threshold value of the penalty evaluation value is set to, for example, about 0.3.

The dictionary 3 is stored in advance as follows. That is, referring to FIG. 12, first, a representative character pattern is input to the character pattern input unit 1. next,
As shown in FIG. 3, the contour tracing section 7 performs contour tracing on the character pattern to generate a coordinate value sequence of contour pixels (S121), and the primitive segment generating section 8 obtains the contour pixels obtained in step 121. Is approximated by an effective line segment and this is defined as a primitive segment (S
122), the coordinates of the primitive segment are normalized by the normalization processing unit 9 (S123). Next, the concave / convex segment generation unit 11 generates a concave / convex segment from the primitive segment normalized in step 123 (S124), and the corner division processing unit 12 divides the concave / convex segment using the corners in the concave / convex segment as division points. The obtained segment is set as a contour segment (S125). Then, the standard contour segment generation unit 13
Then, a standard outline segment is generated from the outline segment, and the generated standard outline segment is stored in the dictionary 3 together with a character code representing a character category (S126).

The operation described above is performed by a computer (not shown) in the character recognition apparatus reading a character recognition program recorded in advance on the recording medium 6 and executing the character recognition program.

[0045]

As described above, according to the character recognition apparatus and the character recognition method of the present invention, the outline of the input character pattern stored in the character pattern input section is approximated by the outline information extracting section, and the broken line is approximated. The contour information including the directed segments, which are the polygonal lines at the time of approximation, is extracted, and the structure matching unit extracts the outline information of the input character pattern extracted by the outline information extraction unit and the representative characters stored in the dictionary in advance. The contour information of the pattern is collated for each character category in units of directed segments included in the contour information,
The distance between the input character pattern and the representative character pattern is obtained and output as a distance value, and the character recognition result output unit determines the character category of the input character pattern that is most probable from the distance value output by the structure matching unit for each of the character categories. Since the unit of character pattern matching is a directed line segment generated by approximating the contour of the character pattern with a broken line, the matching unit is smaller than before, so it is collapsed into characters or parts of characters with broken character strokes. Higher character recognition performance than before can be obtained even for characters with characters.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of a character recognition device of the present invention.

FIG. 2 is a diagram illustrating an example of an operation of the exemplary embodiment of the present invention and an example of a character recognition method of the present invention.

FIG. 3 is a block diagram showing one embodiment of a contour information extraction unit.

FIG. 4 is a block diagram showing one embodiment of a structure matching unit.

FIG. 5 is a diagram illustrating an example of an input character pattern.

FIG. 6 is a diagram illustrating an example of primitive segment extraction.

FIG. 7 is a diagram illustrating an example of uneven segment generation.

FIG. 8 is a diagram illustrating an example of standard contour segment generation.

FIG. 9 is a diagram illustrating an example of contour tracking.

FIG. 10 is a diagram illustrating an example of primitive segment generation.

FIG. 11 is a diagram illustrating an example of calculating a distance between a primitive segment and a standard contour segment.

FIG. 12 is a flowchart illustrating an example of an operation for creating a dictionary.

[Explanation of symbols]

 Reference Signs List 1 character pattern input unit 2 contour information extraction unit 3 dictionary 4 structure matching unit 5 character recognition result output unit 6 recording medium 7 contour tracking unit 8 primitive segment generation unit 9 normalization processing unit 10 center position calculation unit 11 irregularity segment generation unit 12 Corner division processing unit 13 Standard contour segment generation unit 14 Candidate segment detection unit 15 Inter-segment distance calculation unit 16 Best segment detection unit 17 Character pattern distance calculation unit

Claims (8)

[Claims] 1. A character pattern input unit for receiving a character pattern sampled in a grid and quantized into a binary form and storing the character pattern as an input character pattern, and an outline of the input character pattern stored in the character pattern input unit. A contour information extraction unit for extracting contour information including a directed line segment that is a polygonal line when the polygonal line approximation is performed, and the directed line segment of the character pattern representative for each character category. A dictionary that stores the outline information in advance together with a character code corresponding to the character category, and the outline information of the input character pattern extracted by the outline information extraction unit and the representative character pattern stored in the dictionary. The contour information is collated for each of the character categories in units of the directed line segment included in the contour information, and the input character pattern is collated. A structure matching unit that calculates a distance between the character pattern and the representative character pattern and outputs the distance as a distance value; and for each of the character categories, the character of the input character pattern that is most probable from the distance value output by the structure matching unit. And a character recognition result output unit for obtaining a category. 2. The contour information extracting unit performs contour tracking in a clockwise or counterclockwise direction on the input character pattern stored in the character pattern input unit, and calculates a coordinate value sequence of the contour of the input character pattern. A contour tracing unit to be extracted; a line-shaped approximation of the contour of the input character pattern from the coordinate value sequence of the contour of the input character pattern extracted by the contour tracing unit; A primitive segment generator for extracting each of the minutes as a primitive segment, and a constant size in which the width and height of each of the primitive segments of the input character pattern extracted by the primitive segment generator are predetermined. And convert each of the primitives Character recognition apparatus according to claim 1, comprising: the normalization processing unit which outputs the contour information including a segment in the structure matching unit. 3. The primitive segment forming a convex portion in the input character pattern from each primitive segment of the input character pattern converted and normalized by the normalization processing unit. And a concave and convex segment generation unit that generates a convex segment indicating a set of the primitive segments and a concave segment in the input character pattern that is a part other than the convex segment. It is checked whether the generated concave segment and the convex segment have a corner within a predetermined angle.When the corner is present, the segment having this corner is divided at this corner position, and the divided segment and The concave segment and the convex segment without corners A corner segmentation processing unit that generates the contour segment as an outline segment; and the outline segment generated by the corner division processing unit and the coordinate value sequence of the outline of the input character pattern extracted by the outline tracking unit. A new coordinate value sequence is generated by sampling the coordinate value sequence at predetermined intervals, and is used as a standard outline segment. This standard outline segment is output together with a character code corresponding to an input character pattern as a part of the outline information. 3. The character recognition device according to claim 2, further comprising: a standard contour segment generating unit that stores the outline in the dictionary in advance. 4. The method according to claim 1, wherein the structural matching unit includes an i included in the input character pattern obtained from the contour feature extracting unit.
The primitive segment Vi (where i = 1 to the number of primitive segments included in the input character pattern) and all the primitive segments Vi included in an arbitrary character pattern corresponding to the arbitrary character code in the dictionary A first distance calculation is performed between a standard contour segment and a character code corresponding to a character pattern that falls within the threshold when the distance value falls within a predetermined threshold, and the character code corresponding to the character code. Outputting the standard contour segment that is within the threshold included in the character pattern,
A candidate segment detecting unit that notifies that there is no standard contour segment corresponding to the primitive segment Vi in any of the character patterns when the calculated value is not within a predetermined threshold value; An inter-segment distance calculation unit that performs a second distance calculation between each of the standard contour segments of each of the one or more output character codes and the primitive segment Vi and outputs a distance value; A best segment detecting unit that detects the standard contour segment corresponding to the smallest distance value between the primitive segment Vi and each of the distance values output by the unit, and outputs the distance value; Of the input character pattern and the letter Character recognition apparatus according to claim 3, characterized in that and a inter-character pattern distance calculation unit that calculates a distance value between the character and the character pattern in. 5. A first step of receiving a character pattern sampled in a lattice form and quantized into a binary form and storing the character pattern as an input character pattern, and approximating a contour of the stored input character pattern by a polygonal line. A second step of extracting outline information including a directed line segment that is a polygonal line at the time of approximation; and a step of extracting a representative character pattern in each of the character categories stored in advance together with a character code corresponding to the character category. The contour information including a directed segment and the contour information of the input character pattern extracted in the second step are united for the directed segment included in the contour information for each of the character categories. A third step of collating and determining a distance between the input character pattern and the representative character pattern; and A fourth step of obtaining the character category of the input character pattern. 6. The second step extracts a coordinate value sequence of an outline of the input character pattern by performing contour tracing in a clockwise or counterclockwise direction with respect to the input character pattern accumulated in the first step. Performing a polygonal line approximation of the contour of the input character pattern from the coordinate value sequence of the extracted contour of the input character pattern, and extracting, as primitive segments, directed segments that are polygonal lines when the polygonal line approximation is performed. And converting each of the primitive segments of the extracted input character pattern so that the width and height of the input character pattern have a predetermined constant size, respectively, and the converted respective primitives Outputting the contour information including a segment. The character recognition method according to claim 5, wherein 7. A first process for receiving a character pattern sampled in a lattice and quantized into binary values and storing the character pattern as an input character pattern, and approximating a contour of the stored input character pattern with a broken line. A second process of extracting outline information including a directed segment that is a polygonal line at the time of approximation; and a process of extracting a representative character pattern in each of the character categories stored in advance together with a character code corresponding to the character category. The outline information including a directed line segment and the outline information of the input character pattern extracted in the second processing are united for the directed line segment included in the outline information for each of the character categories. A third process of comparing the input character pattern and the representative character pattern to determine a distance between the input character pattern and the representative character pattern; And a fourth process for obtaining the character category of the character recognition program. 8. The second process extracts the coordinate value sequence of the contour of the input character pattern by performing contour tracing in a counterclockwise or clockwise direction with respect to the input character pattern accumulated in the first process. And a line segment approximation of the contour of the input character pattern from the coordinate value sequence of the extracted contour of the input character pattern, and a directed segment that is a polygonal line when the polygonal line is approximated is extracted as a primitive segment. And converting each of the primitive segments of the extracted input character pattern so that the width and height of the input character pattern have a predetermined fixed size, respectively, and the converted respective primitives 8. The character according to claim 7, further comprising: outputting the contour information including a segment. A recording medium that stores a recognition program.