JPH04589A - Method for extracting feature - Google Patents

Abstract

PURPOSE:To stabilize extracted features even in the case of an inclined character pattern, to make it unnecessary to use a dictionary and to shorten a collating time by detecting the inclination angle of a character pattern, scanning the character pattern and extracting features such as the number of intersections between scanning lines and strokes. CONSTITUTION:A subpattern extracting part 9 scans a pattern register 4 by vertical, horizontal, right oblique, and left oblique subpattern extracting parts 9a to 9d based on an inclination angle obtained by a character inclination detecting part 7 and extracts respective subpatterns. A character frame detecting part 5 detects coordinates regulating the circumcribed frame of the character pattern, and when a division point detecting part 8 detects the coordinates of an objective dividing point of the pattern, a feature matrix extracting part 10 extracts feature values expressing a character line variable from the pattern of each subpattern to form a feature matrix. Thereby, the extracted feature can be stabilized in the case of an inclined character, the preparation of a dictionary corresponding to the deformation of character inclination in an identi fication part 11 is made unnecessary, the dictionary collating time can be shortened and a processing speed can be increased.

Description

Detailed Description of the Invention (Industrial Application Field) The present invention provides a feature extraction method that scans a character pattern obtained by photoelectrically converting characters on a medium in multiple directions to extract the characteristics of the character pattern. , especially regarding a fast and stable feature extraction method.

(Prior art) Conventionally, this type of feature extraction method was disclosed in Japanese Patent Application Laid-open No. 57-2.
Publication No. 3185 (Document 1) and Special Publication No. 58-5555
There was one described in Publication No. 1 (Reference 2).

In the technique of Document 1, subpatterns representing line element components (stroke components) in each direction are extracted from a character pattern obtained by photoelectrically converting characters on a medium. Then, the sub-pattern is divided into multiple regions, and for each divided region,
A feature matrix is generated by extracting features representing the amount of line elements in the sub-pattern. Then, the feature matrix and
The characters were recognized by comparing them with a dictionary prepared in advance.

Further, in the technique of Document 2, the number of intersections between a scanning line and a stroke is extracted as a feature by scanning a character pattern.

(Problems to be Solved by the Invention) However, the above feature extraction method has the following problems.

(i) FIG. 2 is a diagram showing an example of printed characters in italics. A diagram illustrating the number “O” in this figure is shown in the third figure.
Figure 3(a) is an example of a non-italic character pattern "O', and Figure 3(b) is an example of an italic character pattern 101.
An example of the same figure (C) is the same figure (a
), and (d) in the same figure is a vertical sub-pattern extracted from the character pattern in (b) using the method of Document 1.

When the vertical sub-pattern of the character pattern shown in FIG. 3(a) is extracted using the feature extraction method of Document 1, the vertical component can be faithfully extracted as shown in FIG. 3(C).

However, when a vertical sub-pattern is extracted for the italic letters in FIG. 3(b), the line element components are tilted as shown in FIG. 3(d). Moreover, since the line element component shown by the broken line in FIG. 3(d) is to be extracted, the line element is scanned perpendicularly to the inclined line element, so both ends of the line element are missing. As a result, there was a problem that the extracted features were different (in other words, the extracted features were unstable) even though they had the same character shape "O゛".In addition, handwritten characters are generally transformed into upward-to-the-right characters. Since there are many elements and the horizontal line elements are tilted, there is a problem that the extracted features become unstable, as in the above case.

Therefore, when character recognition is performed by comparing characters with dictionaries after feature extraction, it is necessary to prepare a large number of dictionaries corresponding to the deformation of the character slope. As a result, the dictionary capacity increases, the processing speed decreases due to longer verification times, and the hardware scale (device scale) increases, which are disadvantageous.

(ii) In the technique of Document 2, the number of intersections between a scanning line and a line element is used as a feature, so if a line element has an inclination as shown in FIG. 3(b), when the scanning direction is fixed, the extraction The number of intersections (between features) that occur varies greatly, causing the same problem as in (i) above.

The present invention provides a feature extraction method that solves the problems of the prior art, such as instability of extracted features, reduction in processing speed, and increase in device size.

(Means for Solving the Problems) In order to solve the above-mentioned problems, a first invention is to scan a character pattern obtained by photoelectrically converting characters on a medium in a plurality of directions, and determine the characteristics of the character pattern. In this feature extraction method, the following measures are taken.

That is, in a plane defined by the X-Y coordinate system including the character pattern, αx + βy (where x, y are X-Y Find the coordinates that give the maximum and minimum values (coordinate values of a coordinate system), detect the inclination angle of the character pattern based on the coordinates, and extract the features by scanning the character pattern based on the inclination angle. ing.

In the second invention, a character pattern obtained by photoelectrically converting characters on a medium is scanned in multiple directions, and line elements in the scanning direction are determined based on the relationship between the number of consecutive black pixels in the scanning direction and the line width. In a feature extraction method for extracting a subpattern representing a component, dividing the subpattern into a plurality of regions, and extracting a feature representing the amount of line elements of the subpattern for each divided region, the following means are used. This is what I learned.

That is, in the plane defined by the X-Y coordinate system including the character pattern, αX + βy regarding the black pixel of the character stamp for at least two sets of specific real numbers α, β (where x, y are in the X-Y coordinate system The coordinates giving the maximum and minimum values of the coordinate values> are determined, the inclination angle of the character stamp is detected based on the coordinates, and the sub-pattern is extracted according to the inclination angle.

(Function) According to the first and second inventions, since the feature extraction method is configured as described above, when a character slam is input, the maximum value and the minimum value of αx and βy regarding the black pixels of the character slam are input. The coordinates that give the value are detected, and the inclination angle of the character button is detected based on the coordinates.

In the first invention, the character stamp is scanned according to the detected inclination angle, and features such as the number of intersections between a scanning line and a stroke are extracted.

Further, in the second invention, a sub-pattern is extracted according to the detected tilt angle, and then the sub-pattern is divided into a plurality of regions, and the characteristics of the sub-pattern are extracted for each of the divided regions.

This makes it possible to stabilize the extracted features even for character stamps with slopes, and to solve the above-mentioned problem.

(Embodiment) FIG. 1 is a functional block diagram of a character recognition device using a feature extraction method showing an embodiment of the present invention.

This character recognition device has a photoelectric conversion unit 1 that converts an optical signal IN of a character image on a medium such as a form into a quantized electrical signal (digital signal), and on the output side thereof, a line buffer 2 is provided. is connected. Row buffer 2 has a width of 20, for example.
It has a size of 48 x 128 bits in height and is configured to store the digital signal of one line of character image, and the output side is connected to a button register 4 via a character cutting section 3.
is connected.

The character cutting section 3 has a memory, and receives a digital signal for one character from the output of the line buffer 2 (this is called a "character slam").
) is stored in the button register 4. The button register 4 has a storage capacity of, for example, 64 x 64 bits, and has a character frame detection section 5, a line width measurement section 6, a character slope detection section 7, a dividing point detection section 8, and a subpattern extraction section 9 on its output side. is connected. On the output side of the line width measurement section 6, division point detection section 8, and subpattern extraction section 9,
A feature matrix extraction section 10 is connected, and an identification section 11 is connected roughly to its output side.

The character frame detection unit 5 has a function of scanning the character slams in the button register 4 to detect a circumscribing frame, that is, a character frame, and providing the detection result to the division point detection unit 8. The line width measurement section 6 has a function of measuring the line width of the output of the button register 4 and providing the measurement results to the sub-pattern extraction section 9 and the feature matrix extraction section 10.

The character slant detection unit 7 has a function of extracting the slant of the character slams stored in the button register 4 and outputting the extraction result to the sub-pattern extraction unit 9. The division point detection section 8 has a function of detecting division point coordinates for dividing the inside of the circumscribed frame into a plurality of partial regions, and outputting the detection result to the feature matrix extraction section 10.

The sub-pattern extraction section 9 scans the button register 4 in a plurality of directions based on the output slope of the character slant detection section 7 to extract vertical, horizontal, right diagonal, and left diagonal sub-patterns.
Vertical sub-pattern extraction section 9a.

Horizontal sub-pattern extraction section 9b, right diagonal sub-pattern extraction section 9
C1 and a left diagonal sub-pattern extraction section 9d. Each of the extraction units 9a to 9d has a memory for storing a button.

The feature matrix extraction unit 10 extracts features from each of the vertical, horizontal, right diagonal, and left diagonal subpatterns output from the subpattern extraction unit 9 to create a feature matrix, and provides it to the identification unit 11. It has a function. The identification unit 11 has a dictionary memory that stores a standard character feature matrix (standard character mask) G(k) and the character name of the standard character having this feature matrix G(k). Then, by comparing the feature matrix F (k) extracted by the feature matrix extraction unit 10 with the feature matrix G (k) in the dictionary memory, the region within the circumscribed frame from which the feature matrix F (k) is obtained is obtained. It has the function of recognizing characters and figures and outputting character names OUT.

FIG. 4 is a functional block diagram showing a configuration example of the character slant detection section 7 of FIG. 1.

This character inclination detection section 7 includes an X coordinate generating means 21 for assigning an X coordinate to a pixel of the image data M inputted from the button register 4, and a Y coordinate generating means 21 for assigning a Y coordinate to a pixel of the image data M. generating means 22, and coordinate detecting means 30, 40 are connected to their output sides. The coordinate detection means 30 and 40 detect the maximum and minimum total sieve values αx +βy regarding the pixels of the extracted button using the X, Y coordinates and at least two sets of specific α values and data values, and X, Y of the pixel of the extracted button that gives the minimum calculated value
The coordinates are output as feature point coordinates,
Connected to its output side is a slope calculation means 50 that calculates the slope of the character slam based on the feature point coordinates.

The coordinate detecting means 30 includes an X+Y calculating means 31 for calculating the calculated value αX+βy, a maximum value detecting means 32 for detecting the maximum calculated value, and a means for storing the X and Y coordinates of the pixel giving the maximum calculated value. A maximum value coordinate storage means 33, a minimum value detection means 34 for detecting the minimum calculated value, and a minimum value coordinate storage means 35 for storing the X and Y coordinates of the pixel giving the minimum calculated value. has been done. Similarly, the coordinate detection means 40 includes an X-Y calculation means 41 and a maximum value detection means 42.
, maximum value coordinate storage means 43, minimum value detection means 44, and minimum value coordinate storage means 45.

In this character inclination detection section 7, in order to detect feature point coordinates,
For example, since two sets of α and β values, α=B=1, α−1, and β=−1, are used, two coordinate detection means 30 and 40 are provided. One of the coordinate detection means 30 calculates the calculated value X+Y.
The maximum and minimum calculated values for the other coordinate detection means 4
0 has a function of detecting the coordinates of pixels that respectively give the maximum and minimum calculated values regarding the calculated value X-Y.

For example, the XY calculation means 31 is an addition circuit, the XY calculation means 41 is a subtraction circuit, and the detection means 32°34.42.44
are a comparator and a register, respectively, and coordinate storage means 33.
35, 43, and 45 are each composed of registers and the like. Moreover, the X coordinate generation means 21 and the Y coordinate generation means 22
are configured to be used in common with the coordinate detection means 30.40, and therefore the coordinate detection means 30.40 receives the X and Y coordinates output from the X coordinate generation means 21 and the Y coordinate generation means 22h). It has a function to input each input and detect feature point coordinates.

Next, processes (I) to (VI) of each functional block will be described regarding a feature extraction method using the character recognition device configured as described above and a method of performing character recognition from the extraction results.

In this embodiment, as shown in FIG. 2, feature extraction and recognition processing for a character string including characters whose vertical sides are inclined, such as italic characters in printed characters, will be described below.

(1) When the optical signal IN of the character image written on the character stamp generation processing form is input to the photoelectric conversion unit 1, the photoelectric conversion unit 1 converts the optical signal ]N into 2
The digital signal of the value, that is, the character line part is converted to "1'" and the background part to "O". The digital signal of one line of character images converted by the photoelectric conversion unit 1 is stored in the line buffer 2.

The character cutting section 3 reads out a digital signal for one character (character slam) from the digital signal of the character image stored in the row bang 72, and stores it in the bang register 4. In this embodiment, the form format is specified in advance, and the address indicating the character position within the line buffer 2 is stored in the memory of the character cutting section 3. Therefore, the character extraction operation is executed by reading the contents of the row buffer 12 specified by the address.

(ff) Character frame detection/line width measurement processing character frame detection section 5
Now, scan the button in the button register 4 and find the left end coordinates of the button X, I! , the right end coordinate Xr, the upper end coordinate Yt, and the lower end coordinate Yb are detected. The circumscribing frame, that is, the character frame is (X
J! , Yt), (XJI, Yb), (Xr, Yt), (
A rectangular frame connects the four points (Xr, Yb).

In addition, after character frame detection, in order to normalize between features,
Calculate the size of the required character frame. In other words, the size of the character frame in the parallel direction (horizontal direction) to the X-axis of the button register 4 is WPh = Xr - XJ + 1, and the size of the character frame in the perpendicular direction (vertical direction) is WPV. W
Pv=Yt-Yb+1 are respectively excluded. moreover,
The size of the character frame in the diagonal right and 45° left diagonal directions is WPr.
and WPN.

On the other hand, the line width measurement unit 6 inputs the digital signal from the button register 4, and counts the number Q of states in which all points in a 2×2 window are black bits, and the number A of all black bits, for example. The line width WL is calculated according to the conventionally known formula (1).

WL=A/ (A-Q) ・・・・・・(
1) (III) Character slant detection processing The character slant detection section 7 extracts the slant degree of the character slams stored in the button register 4. This extraction method
Figure 4, Figure 5 (a), (b) and Figure 6 (a), (b)
).

FIGS. 5(a) and 5(b) are diagrams showing an example of detecting feature points to explain the slope detection method.On the plane expressed in the X-Y coordinate system, there is a button M2 to be extracted. Characters with no slope (FIG. 5(a)) and italic characters (FIG. 5(b)) are shown. BL, BR, TL, and TR are feature points.

Figure 6(a).

(b) explains the character slant detection method;
FIG. 11A is a flowchart of maximum value processing, and FIG. 10B is a flowchart of minimum value processing.

This character inclination detection process will be explained below by dividing into (A) detection process of feature point TL and BR coordinates, (B) detection process of feature point BL and TR coordinates, and (C) inclination calculation process.

(A> Detection process of feature points TL and BR coordinates In this process,
The following (1) is detected by the coordinate detection means 30 shown in FIG.
This is executed in the processing steps of ~(5).

(1) Steps 81 to S3 First, scanning of the slam register 4 is started, and the image data M output from the slam register 4 is input to the character slant detection section 7 (step 31).

This image data M is input to the maximum value detection means 32 and the minimum value detection means 34 in the coordinate detection means 30 for each pixel.

Along with this, two X coordinate generating means and two Y coordinate generating means are provided.
2 generates X and Y coordinates associated with each pixel of the image data M in synchronization with the output of the image data M. As a result, these generating means 21 and 22 assign X and Y coordinates to the image data M. and the output X
, Y coordinates are determined by the X+Y calculation means 31 in the coordinate detection means 30.
, maximum value coordinate storage means 33, and minimum value coordinate storage means 3
5 (step 82>).

When the X+Y metering means 31 inputs the X and Y coordinates, it calculates the calculated value X + Y from these X and Y coordinates, and outputs the calculated value to the maximum value detection means 32 and the minimum value detection means 34 (step 33). Steps 82 to S3 [Thus, the maximum value detection means 32 receives the image data M and the calculated value, the minimum value detection means 34 receives the image data M and the calculated value, and the maximum value coordinate storage means 33 stores the X and Y coordinates, and the minimum value Coordinate storage means 35
Inputs the X and Y coordinates for each pixel. Then, the coordinate detection means 30 performs steps S4, S5a, which will be described later.
Repeating the judgment in S5b or S7 for each pixel,
Processing is performed according to the determination result.

(2) Step S4 The maximum value detection means 32 and the minimum value detection means 34 detect that the image of the input image data M has a predetermined pixel value (for example, "1").
').

(2) When the pixel does not have a predetermined pixel value, the maximum value detection means 32 and the minimum value detection means 34 do not rewrite the stored comparison values, and at the same time, the maximum value coordinate storage means 33 and the minimum value coordinate storage means 35 also Do not rewrite the stored X and Y coordinates. The detection means 32, 34 performs the determination in step S7 after step S4.

■ When the pixel value is a predetermined value, the maximum value detection means 32 compares the comparison value and the calculated value (step 55a) after step S4, and the minimum value detection means 34
After step S4, the comparison value and the calculated value are compared (step 55b).

(3) Step S5a In this step, it is determined whether the calculated value is greater than the comparison value, and the following processing is executed depending on the determination result.

■ When the calculated value>comparison value, the maximum value detection means 32 stores a calculated value larger than the comparison value as a new comparison value instead of the previously stored comparison value (rewriting the comparison value), and at the same time The set pulse is output to the maximum value coordinate storage means 33. The maximum value coordinate storage means 33 into which the set pulse is input stores the stored X,
Instead of the Y coordinate, the X and Y coordinates of the pixel that gives a calculated value larger than the comparison value are newly stored ("Rewrite rX, Y coordinates" step 36).

■ When the calculated value≦comparison value, the maximum value detection means 32 stores the previously stored comparison value without rewriting it, and the maximum value coordinate storage means 33 also rewrites the stored X and Y coordinates. do not have.

The maximum value detecting means 32 performs the determination in step S7 after step S5a in both cases (1) and (2).

Incidentally, as the initial value of the comparison value stored in the maximum value detection means 32, for example, a value smaller than the value that can be taken as the calculated value α may be used. For example, α=β=1, and the image data M is divided into 1 row and m columns of pixels (therefore, O≦X≦m
−1 and O≦Y≦1−1), −1 can be set as the initial value of the comparison value. Alternatively, the calculated value αX+βy inputted first to the maximum value detection means 32 may be used as the initial value of the comparison value.

Further, the maximum value detection means 32 rewrites the comparison value and the X, Y coordinates when the calculated value αx + βy is larger than the comparison value, and when the calculated value is equal to the comparison value, and also performs the rewriting of the comparison value and the X, Y coordinates. When the value is smaller than the comparison value, the comparison value and the X, Y coordinates may not be rewritten. Furthermore, any suitable numerical values may be used as the X and Y coordinates of the maximum value coordinate storage means 33.

(4) Step S5b In step S5b of FIG. 6(b), it is determined whether the calculated value is smaller than the comparison value, and the following processing is executed based on the determination result.

(i) When the calculated value is less than the comparison value, the minimum value detection means 34 stores a new calculated value smaller than the comparison value as a comparison value instead of the previously stored comparison value (rewrites the comparison value), and The minimum value coordinate storage means 35 outputs a set pulse. When the set pulse is input, the minimum value coordinate storage means 35 stores the stored
, Y coordinates are rewritten to the X, Y coordinates of the pixel that gives a calculated value smaller than the comparison value (rX, Y coordinate rewriting step 36).

■ When the calculated value≧comparison value, the minimum value detection means 34 stores the previously stored comparison value without rewriting it, and the minimum value coordinate storage means 35 also rewrites the stored X and Y coordinates. do not have.

In both cases (1) and (2), the minimum value detection means 34 performs the determination in step S7 after step S5b.

Note that, as the initial value of the comparison value stored in the minimum value detection means 34, a value larger than the possible value of the calculated value αx + βy may be used, for example. For example, α=β=1,
Image data M is divided into 1 row and m columns of pixels (therefore, 0
≦X≦m-1 and O≦Y≦1-1),
m+n-1 can be set as the initial value of the comparison value. Alternatively, the calculated value αx+βy inputted first to the minimum value detection means 34 may be used as the initial value of the comparison value.

Further, the minimum value detection means 34 rewrites the comparison value and the X, Y coordinates when the calculated value is smaller than the comparison value or when the calculated value is equal to the comparison value, and at the same time, the calculated value is smaller than the comparison value. When the value is large, the comparison value and the X, Y coordinates may not be rewritten. In general, any suitable numerical values may be used as the X and Y coordinates of the minimum value coordinate storage means 35.

(5) Step S7 In step S7, it is determined whether or not scanning of the image data M has been completed, and the following processing is executed depending on the determination result.

■ When the scanning of the image data M is finished, the maximum value detection means 32 and the minimum value detection means 34 detect the
, Y coordinates are outputted to the maximum value coordinate storage means 33 and the minimum value coordinate storage means 35. Then, the coordinate storage means 33 and 35 output the stored X and Y coordinates as feature point coordinates. At the same time, the detection means 32, 34 initialize the comparison values (step 38).

When processing is completed for all pixels, the X stored in the maximum value coordinate storage means 33 and the minimum value coordinate storage means 35
, Y coordinates of the pixel that gives the maximum and minimum values
coordinates, that is, feature point coordinates.

In the coordinate detection means 30, α=β−1, so at the end of all processing, for example, the coordinates of the feature point BR of the extracted pattern M2 shown in FIG. 5 are stored in the maximum value coordinate storage means 33. , and the downward coordinates of the feature points are stored in the minimum value coordinate storage means 35.

■ When the scanning of the image data M is not completed If the scanning of the image data M is not completed and the processing has not been completed for all pixels of the data M, the coordinate detection means 30 detects the remaining pixels of the image data M. Step 34. S5a, S5
b or S7 is made, and processing is performed according to the result of the judgment.

(B) Detection processing of feature point BL and TR coordinates This processing is performed by the coordinate detection means 40 and the like. That is, the coordinate detection means 40 performs almost the same processing as the coordinate detection means 30 in parallel with the operation of the coordinate detection means 30 described above.

In this coordinate detection means 40, α=1 and β=-1, so when the processing is completed for all pixels, for example, the coordinates of the feature point TR of the extracted pattern M2 shown in FIG. 5 will be the maximum value. It is stored in the coordinate storage means 43, and the feature point B
The coordinates of L are stored in the minimum value coordinate storage means 45.

(C) Slope calculation process In the processing (A>, (B)), the feature points TR and TL.

Coordinates of the four points SR and BL (TRx, TRy>.

(TLX, TLV), (BRx, BRy), (BLX,
When Blj') is detected, the inclination calculation means 50 performs the following steps.
From the coordinates of the four points, a value (degree of inclination) TH corresponding to the inclination angle of the input character pattern is calculated according to the following formula.

That is, the inclination calculation means 50 uses equation (2) to calculate the difference between the X coordinate values of the upper and lower ends of the right side of the quadrilateral obtained by connecting the four feature points, and the X coordinate values of the upper and lower ends of the left side. A value obtained by dividing the average difference in coordinate values by the average character height is outputted as the slope TH and given to the sub-pattern extracting section 9.

(IV) Division Point Detection Process The division point detection unit 8 determines division point coordinates on the X-axis and Y-axis for dividing the area within the circumscribing frame into NXXNY partial areas for each circumscribing frame. However, NX is the number of divisions in the X-axis direction, and NY is the number of divisions in the Y-axis direction.

The numbers of divisions NX and NY are preferably set to arbitrary suitable values depending on the complexity of the characters. For example, when recognizing characters with a small number of strokes such as kanji and katakana, (2X
2) Divide the area within the circumscribing frame into a small number of partial areas of about 3×3.

When kanji are to be recognized, (4X4) to (8X8
) The area within the circumscribing frame is often divided into partial areas of approximately However, in this embodiment, regardless of the complexity of the recognition target, the area within the circumscribed frame is divided into equal parts, for example, into (4×4) parts.

The division coordinate on the X axis DX (n>, and the division coordinate D on the Y axis
Y (n) is determined by the following equations (4) and (5).

DX (n>・・・・・・(3> DY (m) However, n=1.2. ・−・−・, NX−lm=1.2
．． ------・, NY-1 In this embodiment, for example, NX=NY=4 (V) Sub-pattern extraction process FIGS. Figure (a) is an example of a character slam, and the same figure (b)
) is an example of a vertical sub-pattern. An arrow P in FIG. 7(a) indicates a scanning path when vertical sub-patterns are extracted.

In the sub-pattern extraction section 9, based on the slant degree obtained by the character slant detection section 7, a vertical sub-pattern extraction section 9a, a horizontal sub-pattern extraction section 9b, a right-diagonal sub-pattern extraction section 9C, and a left-diagonal sub-pattern extraction section 9d each extract a button register. 4 Directions almost perpendicular to the X-axis direction (vertical direction) and parallel directions (horizontal direction), 45° counterclockwise direction from the X-axis (45° diagonal right direction), and 45° clockwise direction
The button register 4 is scanned with the direction of 45° diagonally to the left (45 degrees diagonally to the left) as the main scanning direction, and vertical, horizontal, diagonal right, and diagonal left sub-patterns corresponding to each main scanning direction are extracted.

In this embodiment, for example, the italic character shown in FIG. The slope of the character is extracted from the slope of the character. Since this inclination is due to the inclination of the vertical sub-pattern, the character stamp is scanned based on the inclination to extract the vertical sub-pattern.

The operation of this vertical bump extractor 9a will be explained.

The vertical sub-pattern extraction section 9a scans the character stamp shown in FIG. 7(a) based on the degree of inclination obtained by the character inclination detection section 7, and detects continuous black bits (black runs) on the scanning line. Then, a black run with a length j that satisfies the following equation (5) is extracted from the detected black runs.

g≧N−WL (5) However, g: length N of the black run in the main scanning direction; arbitrary constant for each sub-pattern (for example, 2) The scanning path P is as follows. Vertical scanning starts from the top edge. Scan start address (Xa.

The coordinates (xH, yH) of the scanning path P from YT) can be expressed by the following equation (6).

x1=xa y1=YT x ・=Xa+THx (y 1-YT)VH=Vi-
1+1 (6) However, H is a real number, the result of HX (V·-YT) is rounded down to the decimal point, and all coordinates are integers.

The vertical sub-pattern extracting unit 9a regards the black runs that satisfy equation (5) as black runs constituting a sub-pattern and stores them in a vertical sub-pattern memory (not shown). Black runs that do not satisfy equation (5) are regarded as white bits.

Similarly, the horizontal, right diagonal and left diagonal sub-pattern extraction section 9b
, 9C, and 9d scan the original pattern with the horizontal, diagonal right, and diagonal left directions as the main scanning directions, and select black runs that satisfy equation (5) from among the black runs on the scanning lines in the respective main scanning directions. The extracted black runs are stored in horizontal, right diagonal, and left diagonal subpattern memories (not shown) as black runs constituting a subpattern.

In this embodiment, the inclination of characters is not considered when extracting horizontal, right diagonal, and left diagonal sub-patterns.

(Vl) Coordinates X where the character frame detection unit 5 defines the circumscribed frame of the character stamp in the feature matrix extraction process
, I), Xr, Yt, and Yb, and further, when the division point detection unit 8 detects the target division point coordinates for the character pattern, the feature matrix extraction unit 10 extracts vertical, horizontal, right diagonal, and left diagonal sub-patterns. Extract the special weight from each bang and create a feature matrix.

That is, the feature matrix extraction unit 10 converts one circumscribed frame area into target dividing point coordinates and coordinates XfJ, Xr.

The area is divided into NXXNY partial areas based on Yt and Yb, and feature amounts representing the character dose of the subpattern in each partial area are extracted. Then, a feature matrix consisting of NXXNYX four special weights extracted from each sub-pattern within one circumscribed frame area is extracted as an inter-feature matrix for the circumscribed frame area.

First, feature extraction from the horizontal sub-pattern (H3P) will be explained.

The feature matrix extraction unit 10 extracts target division point coordinates and coordinates
Based on JII, Xr, Yt, and Yb, the area within the circumscribing frame is divided into NXXNY partial areas and the target dividing point coordinates and coordinates X,! ! , Xr, Y↑, Yb are division point coordinates), horizontal sub-pattern H3P within the partial area for each partial area.
The number of black bits BH(i.

j).

The division area is determined from the division point coordinates as follows. First, set the division point coordinates DX (n>) on the X-axis on a straight line parallel to the X-axis passing through the midpoint of the upper and lower edges of the circumscribed frame of the character pattern. From this coordinate as a starting point, use the following equation (7 The left side of the coordinate series obtained by ) to (9) is defined as a division boundary S as shown in FIG. 7(b).

X□=DX (n> y□=(YT+YB)/2 ・・・・・・(7> From (X□, y□), the coordinate series downward is X・=D
X (n) INT (THx (y i-y□) +0.5)yi
=VH, +1 ・・・・・・(8) However, ;=1.2.3゜((YT-YB)/2-YT) From (x□, y□), the upward coordinate series is: xj=D
X (n) 10 INT (THX (V□ V j) +0.5)
yj=yj+1 1...(9) However, j=-1,-2,-3,...

(YB-(YT-YB)/2). However, INT() indicates that the calculation within the parentheses is performed using a real value, and the decimal part of the result is rounded down to an integer value.

The horizontal division boundary S is set in the horizontal direction from DY (m>).

The number of black bits BH (i, j) of the horizontal sub-pattern H3P is counted for each divided area divided as described above. This B
H(i,j) is the number of black bits in the partial area at the i-th row and j-th column regarding one circumscribed frame area. Next, according to equation (10), the inter-feature FH (
Calculate i, j>.

Roughly, in the same way as in the case of H3P, the partial area of the i-th row and j-th column (7) VSP, H3P, LSP (7) Black 1: j number of bits BV (i, j), BR (i, j ＞、BL
(i, j> is counted, and the VSP for the partial area in the i-th row and j-th column is calculated according to the following equations (11) to (13).

H3P, LSP holding weight FV(i, j>, FR(i, j
>, calculate FL(i,j).

However, i=1.2. ......, NXj=1.2. −
・・・・・・, NY WL; Line width WPh; Character width (=Xr-X, Q +1) However, WPV: Character height (=Yb-Yt+1>WPr=WPl
= (WPv+WPh>/2 or more, VSP, H3P, H3P, L for each partial area of the area within the circumscribed frame
The SP features are extracted and a feature matrix F (k) (k=1.2.
・・・・・・．

NXxNYx4) is obtained. The feature matrix extraction unit 10 extracts the feature matrix F (k) for each circumscribed frame region, and sends the extraction result to the identification unit 11 .

(VI) The identification processing identification unit 11 uses the extracted feature matrix F(k),
Feature matrix G (
k>, the feature matrix F
Recognize the characters and figures in the area within the circumscribed frame obtained in (k). In this recognition, the feature matrix F is calculated according to the following equation (14).
Find the distance between (k) and G (k), and use the character name (for example, the character code specified in the JIS standard) of the standard character in the feature matrix G (k) with the minimum distance as the recognition result. Output.

As described above, this embodiment has the following advantages.

(a> In this embodiment, in the plane expressed by the X-Y coordinate system that includes the character slam in the button register 4, the coordinates that give the maximum and minimum values of αX and βy regarding the black pixels of the character pattern are The detecting unit 7 detects the character pattern, and further detects the inclination angle of the character pattern based on the coordinates.Then, the sub-pattern extracting unit 9 extracts the sub-pattern according to the inclination angle.Therefore, even for character patterns having an inclination, The extracted features become stable.Therefore, there is no need to prepare a dictionary in the identification unit 11 that accommodates the deformation of the character slope, and by reducing the dictionary capacity, the matching time can be shortened and the processing speed can be increased. At the same time, the hardware size is small and character recognition with high recognition accuracy becomes possible.

(b) In the above embodiment, a method for extracting features and a method for recognizing italic characters in printed characters as shown in FIG. 2 has been described, but characters to be recognized are not limited thereto. for example,
In the law of handwritten characters, when a character written upward to the right is to be recognized, the horizontal side of the character is assumed to be inclined, and the recognition can be performed using substantially the same process as in the above embodiment.

Note that the present invention is not limited to the above embodiments, and various modifications are possible. Examples of such modifications include the following.

(i) The feature extraction method of the above embodiment can also be applied to a method of extracting features by scanning character stamps, such as that described in Document 2 above. For example, in the method of Document 2, the number of intersections between a scanning line and a stroke is used as a feature, so if the stroke has an inclination, the number of intersections extracted will vary greatly when the scanning direction is fixed. Therefore, by extracting the slope using the method of the above embodiment and performing feature extraction according to the slope, substantially the same effect as that of the above embodiment can be obtained.

(ii) The character inclination detection section 7 shown in FIG. 4 may be constructed of other functional blocks, or these functional blocks may be executed under program control using a computer. Furthermore, the character inclination detection process may be executed using process flowcharts other than those shown in FIGS. 6(a) and 6(b).

(Effects of the Invention) As explained in detail above, according to the first invention, in a plane expressed by an X-Y coordinate system including a character pattern,
The coordinates giving the maximum and minimum values of αX and βy are determined, and the inclination angle of the character pattern is detected based on the coordinates. next,
The character pattern is scanned according to the inclination angle, and features such as the number of intersections between scanning lines and strokes are extracted. Therefore, the extracted features are stable even for character patterns having an inclination. Therefore, there is no need to prepare a dictionary that corresponds to the deformation of the character slope, and the processing speed is fast.
Character recognition with high recognition accuracy is possible on a small hardware scale.

In the second invention, a sub-pattern is extracted according to the detected tilt angle, and then the features of the sub-pattern are extracted, so that substantially the same effect as the first invention can be obtained.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of a character recognition device using a feature extraction method showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of italic characters, and FIGS. Figure 4 is a functional block diagram of the character slant detection section in Figure 1, and Figures 5 (a) and (b) are diagrams for explaining the feature extraction method. Diagram showing a detection example, No. 6
Figure (a). (b) is a flowchart of character slant detection, Fig. 7 (a)
, (b> is a diagram for explaining the sub-pattern extraction method. 1...Photoelectric conversion unit, 2...Line buffer, 3...Character cutting unit, 4...Bang register, 5...Character frame detection unit, 6...
Line width measurement section, 7...Character slope detection section, 8...
... Division point detection section, 9 ... Sub pattern extraction section, 10 ... Feature matrix extraction section, 11 ...
. . . Identification unit, 21 . . . X coordinate generation means, 22
...Y coordinate generation means, 30, 40...
Coordinate detection means, 50... Slope calculation means.

Claims (1)

[Scope of Claims] 1. A feature extraction method for extracting features of a character pattern by scanning a character pattern obtained by photoelectrically converting characters on a medium in a plurality of directions, comprising: In the plane defined by the Y coordinate system, αx+βy (however, x,
Find the coordinates that give the maximum and minimum values (y is the coordinate value of the X-Y coordinate system), detect the inclination angle of the character pattern based on the coordinates, and scan the character pattern based on the inclination angle to determine the characteristics. A feature extraction method characterized by extracting. 2. A character pattern obtained by photoelectrically converting characters on a medium is scanned in multiple directions, and a sub-pattern representing a line element component in the scanning direction is created based on the relationship between the number of consecutive black pixels in the scanning direction and the line width. In the feature extraction method, the sub-pattern is divided into a plurality of regions and, for each divided region, a feature representing the amount of line elements of the sub-pattern is extracted.
In the plane defined by the Y coordinate system, the maximum and minimum values of αx + βy (where x, y are coordinate values in the X-Y coordinate system) regarding the black pixels of the character pattern for at least two sets of specific real numbers α and β. A feature extraction method comprising: determining coordinates that give a value, detecting an inclination angle of the character pattern based on the coordinates, and extracting the sub-pattern according to the inclination angle.