JPH04590A - Method for character recognition - Google Patents

Abstract

PURPOSE:To stabilize extracted features, to make it unnecessary to use a dictionary corresponding to the change of character inclination and to improve a processing speed, compact size and recognition accuracy by extracting an average inclination from the horizontal and vertical stroke components of a character pattern and determin ing a division area. CONSTITUTION:A character inclination extracting part 8 extracts respective strokes of horizontal and vertical subpatterns from a subpattern extracting part 7. The inclinations of respective trokes are calculated from the coordinates of both the ends of these strokes, average the inclinations and extracts respective average inclination values of both the horizontal and vertical subpatterns. A marginal distribution part 9 scans the character pattern stored in a pattern register 4 based upon these average inclination values and extracts the marginal distribution in both the horizontal and vertical directions. A division point detecting part 10 finds out the centroid coordinates of the peripheral distribution and determines the division point coordinates in the (x) and (y) axis direction. A feature matrix extracting part 11 divides a circumscribed frame and extracts a feature value expressing the character line variables of the subpatterns in respective areas. Thereby, the extracted features can be stabilized even in an inclined character.

Description

Detailed Description of the Invention (Field of Industrial Application) The present invention relates to a character recognition method that recognizes input characters from a character pattern obtained by photoelectrically converting characters on a medium, which is particularly fast and has good recognition accuracy. It concerns a character recognition method.

(Prior art) Conventionally, this type of character recognition method was used, and
There was one described in Publication No. 123171.

Conventionally, when recognizing characters on a medium, features are extracted by the movement of character line elements due to differences in the writer in handwritten characters, and the movement of character line elements due to differences in fonts (fonts) in printed characters. Since the characteristics change, it was necessary to prepare a dictionary that can accommodate the changes in the characteristics. As a result, the dictionary capacity increases, and the time required to check the dictionary also increases, resulting in a decrease in processing speed and an increase in the size of the device.

Therefore, in the technique of the above-mentioned document, a character pattern obtained by photoelectrically converting characters on a medium such as a form is scanned in the horizontal and vertical directions, the number of black bits is counted for each scanning line, and the character pattern is The distribution (marginal distribution) of the number of black bits on the X-axis and Y-axis set for is calculated. Furthermore, the barycenter coordinates of the peripheral distribution are determined, and based on the barycenter coordinates, division coordinates and division boundaries are set to determine division areas, and the inside of the circumscribed rectangular frame of the character pattern is divided. Therefore, it is possible to make the division boundary, and therefore the division area, follow the movement of character line elements in the horizontal and vertical directions. As a result, sub-patterns representing line elements in each of the horizontal, vertical, right diagonal, and left diagonal directions are extracted from the character pattern, and a feature matrix is created in which the amount of character line elements of each sub-pattern for each divided area is used as a feature element. By extracting the extracted feature matrix and comparing it with a dictionary prepared in advance, fluctuations in characteristics such as stroke position fluctuations based on individual differences can be absorbed, and processing speed can be increased by reducing dictionary capacity. I am trying to achieve this. Hence the characters! ! There was a large effect on raw horizontal and vertical movement.

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

Figures 2 (a), (b), (c), and (d) are diagrams for explaining the conventional character pattern division method;
) is an example of a character pattern with no slope [○J, the same figure (b) is an example of the italic character rQJ, the same figure (c) is the peripheral distribution diagram of the same figure (a), the same figure (d) is the same figure ( It is a peripheral distribution map of b).

In the conventional character recognition method, the peripheral distribution SXY is obtained by scanning the input character pattern in the horizontal and vertical directions. is divided into multiple areas. However, as shown in Figure 2(b), if a vertical division boundary S is determined for an italic character whose vertical line elements have a slope, the characters without slope as shown in Figure 2(a) When compared with the case, each division point coordinate D(1
)~D(5) and D(11)~D(15) are different, and for characters with no slope, the division boundary S is set parallel to the vertical line element, but the vertical line element has a slope. For italic characters, the dividing boundary S is set so as to diagonally divide the character line elements. Therefore, although it is the same character, the marginal distribution 5
XY(1) and SXY(2> are significantly different, and therefore the extracted features are significantly different, reducing character recognition accuracy.In this way, in conventional character recognition methods, vertical lines Since there is no way to accurately divide italic characters that naturally have a slope, we had no choice but to prepare a dictionary that accommodates various transformations in order to absorb variations in such characteristics. , the dictionary capacity has increased and it takes time to collate, so the hardware scale (equipment scale) has increased.
However, there are problems such as an increase in processing time and an increase in processing time, and it has not yet been possible to obtain a product that is technically satisfactory.

Such a problem similarly occurs in the case of a character in which horizontal line elements are inclined, such as a handwritten character with an upward slope to the right.

The present invention provides a character recognition method that solves problems E of the prior art, such as an increase in processing time and an increase in the size of the device.

(Means for Solving the Problems) In order to solve the above-mentioned Mu, the present invention divides a character pattern obtained by photoelectrically converting characters on a g typeface into a plurality of regions, and for each divided region. In this character recognition method, the character is recognized by extracting the features, fS, and then comparing the extracted features with a dictionary prepared in advance.

That is, both or either of the horizontal stroke component and the vertical stroke component are extracted from the character pattern, and for each stroke component, the stroke is calculated by weighting the slope of each stroke using the length of the stroke as a weight. The character pattern is extracted as the average slope of the component, the character pattern is scanned according to the average slope to extract the marginal distribution, the dividing point (dividing point coordinates) is detected using the marginal distribution, and the dividing point and the average slope are The character pattern is divided into a plurality of areas based on the above.

(Operation) According to the present invention, since the character recognition method is configured as described above, when a character pattern is input, both or one of the horizontal stroke component and the vertical stroke component of the character pattern is extracted, The average slope of each stroke is extracted from the horizontal stroke component, and the average slope of each stroke is extracted from the vertical stroke component.

Here, the average slope of each stroke component is determined by weighted averaging of the slopes of each stroke included in the stroke component using the length of the stroke as a weight. Normally, long strokes reflect the slant of the character well, and short strokes have large variations in the wX slant of the character.
By calculating the average inclination using a weighted average, the inclination of long strokes is well reflected in the calculation of the average inclination, and the influence of the inclination of short strokes can be suppressed. The character pattern is scanned according to the average slope obtained in this way to extract the marginal distribution, the division point coordinates are determined using the peripheral distribution, and the character pattern is created based on the division point coordinates and the previously determined average slope. is divided into multiple areas. Thereafter, features are extracted for each of the divided regions, and the input characters are recognized by comparing the features with a dictionary.

In this way, since the divided regions are determined by following the slope of the character pattern, the extracted features are stable even for character patterns that have slopes in vertical and horizontal line elements, so the deformation of the character slope There is no need to prepare a compatible dictionary. Thereby, it is possible to increase the processing speed, reduce the size of the device, and improve recognition accuracy. Therefore, the above problem can be solved.

(Embodiment) FIG. 1 shows an embodiment of the present invention, and is a block diagram of a character recognition device for explaining a character recognition method.

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 electric signal (digital signal). is connected. Row buffer 2 is, for example, @2
It has a size of 048 x height 128 bits, and is configured to store the digital signal of one line of character image.
A pattern register 4 is connected to its output side via a character cutting section 3.

The character cutting section 3 has a function of storing a digital signal for one character (hereinafter referred to as "character pattern J") from the output of the line buffer 2 into the pattern register 4.

The pattern register 4 has a storage capacity of, for example, 64 x 64 bits, and a character frame detection section 5, a line width measurement section 6, and a sub-pattern extraction section 7 are connected to its output side. The extraction unit 8 is connected, and the marginal distribution extraction unit 9 is also connected.

The character frame detection unit 5 has a function of scanning the character pattern in the pattern register 4 to detect a circumscribed frame, that is, a character frame, and providing the detection result to the peripheral distribution extraction unit 9 and the like. The line width measurement unit 6 has a function of measuring the line width of the output of the pattern register 4 and providing the measurement result to the sub-pattern extraction unit 7.

The sub-pattern extractor 7 scans the pattern register 4 in a plurality of directions to extract vertical, horizontal, right diagonal, and left diagonal sub-patterns, and includes a vertical sub-pattern extractor 7a, a horizontal sub-pattern extractor 7b, and a right-diagonal sub-pattern extractor. 7c, and a left diagonal sub-pattern extraction section 7d. Each of the extraction units 7a to 7d has a memory for storing patterns.

The character slope extraction unit 8 extracts slopes of the vertical subpatterns and horizontal subpatterns extracted by the vertical subpattern extraction unit 7a and the horizontal subpattern extraction unit 7b of the subpattern extraction unit 7, and provides the extraction results to the marginal distribution extraction unit 9. It has a function.

The peripheral distribution extraction section 9 has a function of scanning the character pattern stored in the pattern register 4 using the slope obtained by the character slope extraction section 8 and extracting its peripheral distribution.

A division point detection section 10 and a feature matrix extraction section 11 are connected to the output side of the marginal distribution extraction section 9, and an identification section 12 is further connected to the output side thereof.

The division point detection unit 0 has a function of detecting division point coordinates for dividing the inside of the circumscribed frame into a plurality of regions. The feature matrix extraction section 11 uses the vertical 1. Feature quantities are extracted from each of the horizontal, right diagonal, and left diagonal sub-patterns, and a feature matrix F (k
) and provides it to the identification unit 12. The identification unit 12 has a dictionary memory that stores a standard character feature matrix (standard character mask) G(k) and character names of standard characters having this feature matrix G(k>). By comparing the feature matrix F(k> extracted by the extraction unit 11 with the feature matrix G(k) in the dictionary memory, the character/figure in the area within the circumscribed frame from which the feature matrix F(k) was obtained is recognized. It has the function of executing the command and outputting the character name 0IJT.

Next I: Regarding the character recognition method using the character recognition device configured as above, the processing of each refined block (I>~
(IX) will be explained.

(I) When the optical signal IN of the character image written on the character pattern generation processing form is input to the photoelectric conversion unit 1, the photoelectric conversion unit 1 converts the optical signal IN into 2
The digital signal of the value, that is, the character line portion is converted to °1゛ (this is called "focus bit J"), the background part is converted to "0" (this is converted to "white bit J". Photoelectric conversion section) −
The digital signal of the character image for hf, -1 line converted by is stored in the line buffer 2.

The character cutting section 3 cuts out one character's worth of digital signals (c character pattern) from the digital signals of the hf and H character images stored in the line buffer 2, and stores them in the pattern register 4. In this embodiment, the form format is specified in advance, and the character cutting section 3 has an IJ memory in which addresses indicating character positions within the line buffer 2 are stored. Therefore, the character cutting operation is executed by reading the contents of the line buffer 2 specified by the address.

(II) Character frame detection processing The character frame detection unit 5 scans the pattern in the pattern register 4 and detects the left end coordinate X', right end coordinate Xr, upper end coordinate yt, and lower end coordinate yb of the pattern. The circumscribing frame, that is, the character frame is (X, I), Yt), (XCYb>, (Xr, Y
It becomes a rectangular frame connecting the four points t), (Xr, Yb).

After detecting the character frame, in order to normalize the feature amount,
Calculate the size of the required character frame. That is, assuming that the size of the character frame in the direction (horizontal direction) parallel to the X axis of Pattern I/Jister 4 is WPh, WPh=Xr-X, l! +l and the size of the character frame in the vertical direction (vertical direction) is WPv, and WPv=Yt-Yb+1 is outputted by hx. Furthermore, the sizes of the character frames in the right diagonal and left diagonal 45'' directions are set as WPr and WPl to calculate WPh+WPv WPr=WPr=2 These calculation results are given to the marginal distribution extraction section 9 and the feature matrix extraction section 11. Ra nine.

(III) Line Width Measurement Process The line width measurement unit 6 inputs the digital signal from the pattern register 4 and calculates, for example, the number Q of states in which all points in a 2×2 window are black bits, and the number of all black bits. Count the number A,
The line width WL is calculated according to the following well-known formula (1>).

WL=A/(A-Qn...
(1) (TV) Sub-pattern extraction processing In the sub-pattern extraction section 7, the vertical sub-pattern extraction section 7a,
Horizontal sub-pattern extraction section 7b, right diagonal sub-pattern extraction section 7
cF1. and left diagonal sub-pattern extraction section 7d,
Set it on pattern register 4.

The direction perpendicular to the X-axis direction (vertical direction) and parallel direction (horizontal direction), the direction 45° counterclockwise from the The pattern register 4 is scanned with the 45° diagonal direction) as the main scanning direction, and vertical, horizontal, right diagonal, and left diagonal sub-patterns corresponding to each main scanning direction are extracted.

That is, the vertical sub-pattern extraction unit 7a scans the entire original pattern with the vertical direction as the main scanning direction, and detects continuous black bits (black runs) on the vertical scanning line. and,
A black run of length j that satisfies the following equation (2) is extracted from the detected black runs.

j≧N・WL...(2) However, Il
l: Length N of black runs in the main scanning direction; arbitrary constant for each sub-pattern (for example, 2) Next, the vertical sub-pattern extraction unit 7a extracts the black runs that satisfy the equation (2) and the black runs that refine the sub-patterns. It is regarded as a run and stored in a vertical sub-pattern memory (not shown). Black runs that do not satisfy equation (2) are regarded as white bits.

Similarly, the horizontal, right diagonal and left diagonal sub-pattern extraction section 7b
, 7c, 7d/1i, scan the original pattern horizontally, diagonally to the right, and diagonally to the left in the main scanning direction, and select black runs that satisfy equation (2) from among the black runs on each scanning line in the main scanning direction. Runs are extracted, and unextracted black runs are stored in horizontal, right diagonal, and left diagonal subpattern memories (not shown) as black runs for refining subpatterns.

General examples of sub-patterns extracted in this way are shown in FIGS. 3(a) to 3(e). Figure 3(a)-(
e) is a diagram showing an example of a character pattern and its subpatterns, and FIG. 3(a) is an original pattern consisting of a binary pattern within a circumscribing character frame, and vertical subpatterns obtained from such an original pattern (vsp) is shown in Fig. 3(b), the horizontal subpattern (H8P) is shown in Fig. 3(c), the right diagonal subpattern (R8P> is shown in Fig. 3(d), and the left diagonal subpattern (LS
P) are shown in FIG. 3(e), respectively. Character line element (this is [stroke j
Extract 〉. Next, calculate the slope of each stroke from the coordinate values of both ends of each extracted stroke, and average them. From the vertical sub-pattern, the average slope of the vertical stroke component θV, and from the horizontal sub-pattern, the average slope of the horizontal stroke component. θh are each extracted.

Next, a specific method for extracting the average slopes θ■ and θh will be explained.

First, horizontal scanning is performed over the entire vertical sub-pattern to detect points of change from white bits to black bits and from black bits to white bits. Then, the coordinates of both ends of the vertical stroke (vertical stroke) are extracted from the relationship between the change point coordinates of the scanning line one line before and the current scanning line.

Both ends pi of the extracted vertical stroke: mark (VXS,
1' with VMS, ) and (VXE, VYE,)
1 1'
], the average slope θV of the vertical stroke component is calculated using the following equation (3). However, j-1,...
...Pv, Pv is the number of strokes extracted from the vertical sub-pattern, and VYS, <VYE.

−]1 θV= Here, VLG・ is obtained from the following equation (4), VLG
.

HAX < l VXEi-VXSi l,
l VYEi-VYS+ l )+ ...(4) This formula (4) is the distance between two points, that is, the coordinates of both ends (VXS
i , VYS i ) and (VXEi, VYEi)
This is an approximation formula in which the distance between the two end coordinates representing the length of the vertical stroke is the sum of 1/2 of the smaller of the horizontal and vertical coordinate differences between the two points and the other one.

Further, the average slope θh of the horizontal stroke component is extracted from the horizontal sub-pattern as follows.

Vertical scanning is performed on the horizontal sub-pattern, and the coordinates of both ends of the horizontal stroke are extracted. The coordinates of both ends are (HXSj, HYS, > and (HXE,
, HYE , ), water J
J Average slope θ of 1,000 stroke components
Calculate h using the following equation (5). However, j-1,...
, ph, ph is the number of strokes extracted from the horizontal sub-pattern, HXS, <HXE.

J θh= ・・・・・・(5) Here, HLG is obtained from the following equation (6).

H.L.G.

HAX < l HXEj-t(χSj l,
l HYEj-HYSj l )+ (6) Note that when the number of strokes is O, the average slope is set to zero. That is, when Pv=O, θV=○, and when ph=o, θh=o.

(VI) Marginal Distribution Extraction Process The marginal distribution extraction section 9 scans the character pattern stored in the pattern register 4 and extracts its marginal distribution. below,
Figures 4(a) and (b) show how to extract the marginal distribution.
).

FIGS. 4(a) and 4(b) are diagrams showing a method of determining dividing areas for a character pattern with an inclination, and FIG. 4(a) is an example of a character pattern in the case of the number The same figure (
b) is a diagram showing the peripheral distribution extracted from the character pattern in Fig. 4(a).

In FIG. 4(a), arrow P indicates the start address (xa).

The scanning path when starting vertical scanning from YT) is shown.

In addition, in Figure 4 (a), the X axis and yI are placed on the character pattern.
! I! In addition, in FIG. 4(b), the X-axis is taken in the horizontal direction, and the peripheral distribution SX corresponding to the X-axis direction is taken in the vertical direction.
(x) is taken.

The peripheral distribution extraction unit 9 determines the scanning path of the character pattern stored in the pattern register 4 based on the average slope θV and average slope θh obtained by the character slope extraction unit 8, and uses the scanning path to perform vertical scanning and horizontal scanning. Scanning is performed, and by each scan, a peripheral distribution 5X (x>) in the X-axis direction and a peripheral distribution sy (y) in the y-axis direction are respectively extracted.

Next, a specific method for extracting the marginal distribution sx (x>, SY (y>) will be described.

First, when vertical scanning is performed, a start address (xa), for example, is set as a scanning start point on the upper side of a character pattern.

YT), and scanning is performed in the positive direction of the y-axis from the start address (xa, YT>), the scanning path PS coordinate (J, 3'j,) is calculated by the following equation (7). It is determined.

X 1 ” X a , y1 = YT x , = xa + θv
1+1 (7) Here, θV is a real number, and in the case of FIG. 4(a), it is set to θv=5/20, for example. Also, θvX (y
, -YT) are rounded down to the decimal point, and all coordinate values are integers.

Next, vertical scanning is performed along the scanning path P determined by equation (7), the number of black bits in the scanning path P is counted, and the value of the marginal distribution SX (
xa> is obtained. xa is changed within the range of XL≦xa≦XR. That is, the scanning start point is (XL, YT), (XL+
1. YT), ... (XR, YT) respectively,
At each scanning start point, the number of black bits on the scanning path according to the above equation (7) is counted, and the marginal distribution is extracted. As a result, the marginal distribution SX (x>, x=XL, . . . , XR) is extracted.

Also, using the average slope θh, the peripheral distribution 5Y(
y) is extracted as follows.

First, when a start address (XL, ya>, for example) is set as a scanning start point on the y-axis, the horizontal scanning path is determined by the following equation (8).

x1=XL yl:ya xj=xj-1+1 y,=ya+θllx (x, -XL)J
J...(8) Here, θh is a real number, the decimal part of the calculation result of θhX(x−XL) is rounded down, and all coordinate values are integers.

In equation (8), the value of ya is changed within the range of YT≦ya≦YB. allchi (XL, YT), ・ , (XL.

ya), . . . , (XL, YB) are set as scanning start points. Then, by counting the number of black bits on the scanning path according to equation (8) above for each scanning start point, the marginal distribution S
Y (y), y=YT,..., YB are obtained.

(VII) Division point determination processing The division point detection unit 10 determines the barycenter coordinates of the marginal distribution SX (x>, SY (y)) extracted by the marginal distribution extraction unit 9,
Next, the division in the X-axis direction is performed by dividing the sum of the first moments in the entire range of the character frame in the X-axis direction or y-axis direction or in the range divided by the barycentric coordinates already determined by the sum of black bits in that range. This is to determine the point coordinate DX (m>) and the division point coordinate DY (n) in the y-axis direction. However, mn is the barycenter coordinate number assigned in order of the size of the coordinate value, and m = 1 to (Nχ -1), n=1 to (NYl).In this example, NX=NY-4.

Next, a specific method for determining the dividing point coordinates DX (m) and DY (n) will be explained with reference to FIGS. 4(a) and 4(b).

In order to obtain the division point coordinates DX (m), first, the range XL to XR of the character frame in the X-axis direction is targeted.

By dividing the sum of the linear moments of the peripheral distribution SX (x) of the character pattern by the sum of black bits in that range, the barycenter coordinate DX(2) of the center barycenter coordinate number at the center is determined by the following equation (9).

...(9) However, since the value on the right side of equation (8) is a real number, round up the decimal point to make DX(2>) an integer.

Next, each of the M boundaries, XL to DX (2
>, DX (2) Two centroid coordinates D for ~XR
X (1) and DX (3) are expressed by the following equations (10) and (
11).

X2X[χ=XL...(10) As above, the center of gravity coordinates DX(1), DX(2), D
X (3) is determined, and the coordinates are determined as the target division point coordinates on the X axis.

Similarly, regarding the dividing point coordinates DY(n), first, from the peripheral distribution SY(, y) on the y-axis, the range YT of the character frame
〜YB as a target, detect the center of gravity coordinate DY(2), and then divide it by the center of gravity coordinate DY(2> to create each range YT〜DY(2>, DY<2)〜YB part 1 The center of gravity coordinates DY (1), DY (3) of the surrounding area 1sY(, y>) are detected for zh.

As described above, the center of gravity coordinates DY(1), DY(2), D
Find Y (3) and use the coordinates as the target dividing point on the y-axis! Decided as a target.

(VIIL) Feature matrix extraction processing Character frame detection unit 5 defines the circumscribed frame of the character pattern at coordinates XN, Xr,
When Yt and Yb are detected, and the dividing point detection unit 10 further detects the target dividing point coordinates for the character pattern, the feature matrix extraction unit 11 extracts feature amounts from each of the vertical, horizontal, right diagonal, and left diagonal sub-patterns. Extract and create a feature matrix.

That is, the feature matrix extraction unit 11 extracts one circumscribed frame area from the target dividing point coordinates and the coordinates Xf1. It is divided into NxXNY partial regions by XrYt and Yb, and a feature amount representing the character dose of the subpattern in each partial region is extracted. Then, a feature matrix consisting of NχXNYX4 feature quantities extracted from each subpattern within one circumscribed frame area is created as follows:
It is extracted as a feature matrix of the region within the circumscribed frame.

First, feature extraction from the horizontal sub-pattern (H8P) will be explained.

The feature matrix extraction unit 11 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. Horizontal sub-pattern H8 within the part @ area for each partial area
P's black and soto number BH (i.

j).

The division area is determined from the division point coordinates as follows. First, the dividing point coordinates DX(n) on the X axis are set at the upper side of the circumscribing frame of the character pattern. As the starting point of this coordinate, the following formula (1
Let the left side of the coordinate series obtained in 2) be a division boundary S as shown in FIG. 4 (a>).

x (+, = D X (n) 3'0 = y'r x , = DX (n) + θv
i-1+1 (12) The horizontal division boundary S is similarly set using DY(n) and θh.

For each divided area divided as described above, count the number of black bits BH (i, j>) of the horizontal sub-pattern H3P.
H(i, j> is the number of black bits of the partial area in the i-th row and j-th column regarding one circumscribed frame area.Furthermore, the following formula (1
According to 3), the special @1F regarding the partial area of the jth row and the jth column
Calculate H(i,j).

Count the black bit numbers BV (i, j), BR (i, j), and BL (i, j) of the VSP, H3P, and LSP in the partial area of column j, and calculate the number according to the following equations (14) to (16). vsp regarding the partial area in the i-th row and the j-th column.

H3P, LSP feature quantity FV(i, j>, FR(j, j
>, FL(i, j>).

......(13) However, i=1.2. ......, NXj=1.2.・
..., NY WL, line width wph, character width (=Xr-XU +1), and H8
Similarly to the case of P, the i-th row WL-WPU-=・-
(16') However, WPv; character height (=Yb-Yt+1)WPr=WPJ
2 = (WPv+WPh)72 As above, VSP, H3P, H8P,
Feature 1 of LSP is extracted and a feature matrix F (k>(k=1.2.
...NXXNYX4) is obtained. The feature matrix extraction unit 11 extracts the feature matrix F (k) for each circumscribed frame area, and sends the extraction result to the identification unit 12 .

(IX) The identification processing identification unit 12 extracts the extracted feature matrix F(k> and the feature matrix G(k
), the feature matrix F(k)
Recognize the characters and figures in the area within the circumscribed frame obtained. In this recognition, the feature matrix F (k) is calculated according to the following equation (17)
Find the distance between
The character code defined in the JIS standard) OUT is output as the recognition result.

As described above, this embodiment has the following advantages.

In this example, the vertical sub-pattern and horizontal sub-pattern extracted from the character pattern are used to extract the average slope θ■ of the vertical stroke component from the vertical sub-pattern and the average slope θh of the horizontal stroke component from the horizontal sub-pattern, respectively. The character pattern is scanned according to the average slopes θV and θh to extract the peripheral distribution of the character pattern. Furthermore, the division point coordinates are determined using the peripheral distribution, and the division areas within the circumscribed frame of the character pattern are determined based on the division point coordinates and the average slopes θV and θh, and the feature matrix for each division area is determined. is extracted. Therefore, the extracted features are stable for character patterns that have an inclination, such as handwritten characters in which the horizontal line elements are inclined upward to the right, or printed characters in italic characters in which the vertical line elements are inclined to the right. Therefore, the identification unit 12 selects a dictionary corresponding to the deformation of the character slope.
There is no need to prepare one internally, and by reducing the dictionary capacity, it is possible to shorten the collation time and thereby increase the processing speed, and at the same time, the hardware size is reduced and character recognition with high recognition accuracy is possible.

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) In this embodiment, a case has been described in which the slope of a stroke is extracted in both the horizontal and vertical directions. However, since handwritten characters generally have many characters that slope upward to the right, it is sufficient to extract only the slope of horizontal strokes. Moreover, when the target is an italic character of a printed character, it is sufficient to extract the inclination of a vertical stroke. In this way, the tilt extraction direction can be selected as appropriate depending on the object to be read, and the configuration can be simplified.

(ii) The configuration block diagram in FIG. 1 may be configured not only by individual circuits but also by program control using a computer.

(Effects of the Invention) As described above in detail, according to the present invention, horizontal stroke components and/or vertical stroke components are extracted from a character pattern obtained by photoelectrically converting characters on a medium. For each stroke component, the weighted average of the slope of each stroke using the length of the stroke as a weight is extracted as the average slope of the stroke component,
scanning the character pattern according to the average slope to extract a peripheral distribution, detecting a dividing point using the peripheral distribution, dividing the character pattern into a plurality of regions based on the dividing point and the average slope, Features are extracted for each divided region.

Therefore, it is not possible to extract stable features even when dealing with characters such as handwritten characters in which horizontal line elements are tilted upward to the right, or italic characters in which vertical line elements are tilted to the right such as printed characters. can. Therefore, there is no need to prepare a dictionary that accommodates the deformation of the character slope, and character recognition with high processing speed and small hardware scale is possible with high recognition accuracy.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a character recognition device for explaining a character recognition method according to an embodiment of the present invention, and FIG. 2(a) to
(d) is a diagram for explaining the conventional character pattern division method, Figures 3 (a) to (e) are diagrams showing examples of character patterns and their sub-patterns, and Figures 4 (a) and (b) are diagrams for explaining the conventional character pattern division method. , is a diagram illustrating a method for determining divided regions of a character pattern having an inclination. DESCRIPTION OF SYMBOLS 1... Photoelectric conversion part, 2... Line buffer, 3... Character cutting part, 4... Pattern register, 5... Character frame detection part, 6... Line width measurement part, 7 . . . sub-pattern extraction unit, 8 . . . character slope extraction unit, 9 . . . marginal distribution extraction unit,
10... Division point detection unit, 11... Feature matrix extraction unit, 12... Identification unit.

Claims (1)

[Claims] A character pattern obtained by photoelectrically converting characters on a medium is divided into a plurality of regions, and features are extracted for each divided region, and then the extracted features and a previously prepared In a character recognition method that recognizes the character by checking with a dictionary, extracting both or either one of a horizontal stroke component and a vertical stroke component from the character pattern, and determining the slope of each stroke for each stroke component. A weighted average using the stroke length as a weight is extracted as the average slope of the stroke component, the character pattern is scanned according to the average slope to extract a peripheral distribution, and a dividing point is detected using the peripheral distribution. . A character recognition method, characterized in that the character pattern is divided into a plurality of regions based on the dividing point and the average slope.