JPH0346080A - Feature extraction device - Google Patents

Abstract

PURPOSE:To extract a feature with few fluctuation by converting a distance between the remarking point of a subpattern and a character line to a scale following the local position fluctuation of the character line by using the projection distribution (marginal distribution) of the character line of a character/graphic pattern, and extracting the sum of K-the power of a converted distance as the feature of a character. CONSTITUTION:A feature extraction device 10 is comprised of a subpattern extraction part 12, a character projection part 14, a centroid coordinate system extraction part 16, and a feature extraction part 18. And the distance between the remarking point of the subpattern and the position of the character line is expressed in a conversion value setting a centroid coordinate number attached on a centroid coordinate system in sequence of size of a coordi nate value as the scale by using the projection distribution (marginal distribution) of the character line of the character/graphic pattern, and first and second features can be extracted from the sum of K-th power of the distance expressed in the conversion value. The scale of the conversion value follows the local position conversion of the character line, therefore, the fluctuation of the character recognized as the same character can be suppressed even when the fluctuation of the position of the character line occurs, which stabilizes the feature. In such a way, the feature extraction device with few quantity of fluctuation can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a feature extraction device, and particularly to a device used for feature extraction for character recognition.

(Prior art) Conventional character recognition devices often employ a method of extracting strokes from character patterns and performing recognition processing using the positions and lengths of these extracted strokes, mutual positional relationships between strokes, etc. ing. In this type of device,
(1) By calculating the curvature of a contour point series detected by tracing the contour of a character figure, dividing the contour series using points with large values of curvature as dividing points, and combining the divided series. Extract strokes or (
2) Extract strokes by performing thinning processing on character and figure patterns to create a skeleton, detecting the connectivity of the skeleton pattern and points of sudden angle changes obtained by tracing the skeleton pattern, and extracting strokes. Characters and shapes were identified by extracting geometric features. However, method (1) has the disadvantage that when the character/graphic pattern becomes large or complex, the amount of processing increases, resulting in a decrease in processing speed.
The method requires complicated processing of thinning the character/figure pattern, and this thinning causes distortions, whiskers, etc. in the pattern, and the process to remove these distortions, whiskers, etc. complicates the entire process. There was a drawback.

Therefore, in order to eliminate these defects, scanning lines are drawn in each scanning direction from each point in the character figure pattern (called a point of interest), and the number of intersections between the scanning line and the character line is calculated as the point of interest for the scanning line. A method of extracting points as features has been proposed.

However, the number of intersections extracted as a feature changes and varies as the slope of the character line changes, and therefore the feature becomes unstable. Furthermore, the mere number of intersections alone is insufficient to reflect the structure of characters.

As a way to avoid these problems, JP-A-60-57
There is a feature extraction method disclosed in Japanese Patent No. 474. In this method, a subpattern representing a stroke component in a predetermined direction of a character figure pattern is extracted, and each point or arbitrary point of the subpattern is set as a point of interest, and the character line distribution when viewed from the point of interest in another predetermined direction. Extract features that represent the state. Therefore, the distance M% between the point of interest and all the character lines located on one side of the point of interest on the scanning line in the other direction passing through the point of interest is calculated, and the distance from the character line on one side is calculated. Let the sum of the products be the first feature regarding the point of interest, and roughly the point of interest,
Find the distance to the character line located on the other side of the point of interest on the scanning line in the other direction passing through the point of interest, and calculate the sum of the products of the distances to the character line on the other side to the point of interest related to the point of interest. It is extracted as the second feature. The character/graphic pattern can be recognized based on the characteristics of the character/graphic pattern extracted using these first and second features.

(Problem to be Solved by the Invention) However, in the prior art disclosed in Japanese Patent Application Laid-Open No. 60-57474 mentioned above, when the first and second features are extracted from printed characters or handwritten characters, There was a problem in that the characteristics of the characters changed with changes in the size of the characters and the position of the character lines.

In the case of printed characters, the size of the type and the inclination of the printed characters change, and as a result, the position between the point of interest and the character line changes, and the first and second features change. Furthermore, in the case of handwritten characters, the size of the characters and the position of the slanted character line of the characters change depending on the writer of the characters, and as a result, the distance between the point of interest and the character line changes, and the first and second characteristics change.

In character recognition, the word I! corresponds to such variation in characteristics. If the dictionary is not prepared, the recognition rate will decrease, and if the dictionary N corresponding to the variation of the feature is prepared, the amount of the dictionary will increase, and as a result, the time required for dictionary matching will become longer and the recognition speed will decrease.

An object of the present invention is to provide a feature extraction unit W that can extract features with little variation in order to solve the above-mentioned conventional problems.

(Means for Solving the Problem) In order to achieve this object, the feature extraction device of the present invention scans a character figure pattern in a plurality of predetermined directions and detects a cross section of a character line in each scanning direction. , a sub-pattern extraction unit that extracts sub-patterns representing stroke components in each scanning direction by extracting a character line whose cross-sectional length is sufficiently longer than the character line width of the character-figure pattern; a character projection unit that projects each character figure pattern onto a projection axis to obtain first and second projection distributions of character lines; Find the coordinates of the center of gravity, set the coordinates of the center of gravity of the first projection distribution within the character circumscribing frame as the coordinates of the first division point, and divide the area of the character circumscription frame in the direction of the first projection axis using the coordinates of the division point to perform the first division. Set the area, find the barycentric coordinates on the projection axis of the projected distribution in the first divided area, and use the barycentric coordinates of the first projected distribution as new point coordinates in the first divided area. Repeat the process of determining the barycenter coordinates of the first projection distribution a predetermined number of times to obtain the barycenter coordinate series on the first projection axis, and set the barycenter coordinates of the second projection distribution within the character circumscribing frame as the coordinates of the first other division point, Divide the area of the character circumscribing frame at other dividing point coordinates in the direction of the second projection axis to set a second divided area,
Find the barycenter coordinates on the projection axis of the second projection distribution in the second divided area, and use the barycenter coordinates of the second projection distribution as the coordinates of another new division point for the second projection distribution in the second divided area. a barycenter coordinate series extraction unit that repeats the process of obtaining the barycenter coordinates a predetermined number of times to obtain a barycenter coordinate series on the second projection axis; Then, the subpattern is scanned in another direction that is perpendicular to the scanning direction when extracting the subpattern and passes through the point of interest, and the position of the character line on the scanning line is detected. 11
Based on the relationship, these points of interest and character line positions @ are converted into conversion values using barycenter coordinate numbers added to the barycenter coordinate series in order of the size of the coordinate values, and the points of interest and one of the points of interest are converted. The conversion word m for each of all character line positions on the scanning line on one side is calculated based on the conversion value, and the sum of the powers of each conversion distance on one side (K is a constant) is calculated as the first The conversion distances between the point of interest and all character line positions on the scanning line on the other side of the point of interest are determined based on the conversion values, and the conversion distances of each of these conversion distances on the other side are calculated. The feature extraction unit extracts the sum of powers (K is a constant) as the first feature regarding the point of interest.

(Operation) According to the feature extraction device having such a configuration, the projection distribution (marginal distribution) of character lines of a character figure pattern is determined. and,
The coordinates of the center of gravity of the peripheral distribution within the character circumscribing frame are set as the coordinates of the first dividing point, the area of the character circumscribing frame is divided at the coordinates of the dividing point to set divided regions, and the coordinates of the center of gravity of the peripheral distribution within the divided region are determined. Then, the process of determining the barycenter coordinates by setting a new divided region with the obtained barycenter coordinates as new division point coordinates is repeated a predetermined number of times, thereby obtaining a barycenter coordinate series consisting of a predetermined number of barycenter coordinates.

Then, the distance between the point of interest and the character line position is expressed as a converted value using the barycentric coordinate number added to the barycentric coordinate series in order of the size of the coordinate value, and the sum of the distances expressed by this converted value is Extract the first and second features from. Since the scale of the conversion value follows the local positional variation of the character line, even if there is a variation in the position of the character line, the variation in the characteristics of characters to be recognized as the same character can be reduced, and the characteristics are stabilized.

As the number of barycenter coordinates in the barycenter coordinate series increases, variations in the characteristics of characters to be recognized as the same character can be reduced and the characteristics can be made more stable.

(Embodiments) Hereinafter, embodiments of the present invention will be described with reference to the drawings. It should be noted that the drawings are only schematic illustrations to the extent that the present invention can be understood, and therefore, the connections, input/output signal flows, arrangement positions, shapes, and dimensions of each component are limited to the illustrated examples. It's not a thing.

FIG. 1 is a functional diagram 7099 for explaining the present invention in detail, and shows an example of a character recognition device configured using the feature extraction section N of the present invention.

In the figure, reference numeral 10 indicates a feature extraction device according to an embodiment of the present invention, and this feature extraction section 10 includes a sub-pattern extraction section 12,
It consists of a character projection section 14, a barycentric coordinate sequence extraction section 16, and a feature extraction section 18.

The sub-pattern extraction unit 12 scans the character/figure pattern in a plurality of predetermined directions, detects the cross-section of the character line in each scanning direction, and extracts a character line whose cross-sectional length is sufficiently longer than the character line width related to the character/figure pattern. By extracting, sub-patterns representing stroke components in each scanning direction are extracted.

The character projection unit 14 projects character graphic patterns onto first and second projection axes that intersect, for example, are perpendicular to each other, to obtain first and second projection distributions of character lines.

The barycenter coordinate series extraction unit 16 obtains the barycenter coordinates on the projection axis of the first and second projection distributions within the character circumscribing frame of the character graphic pattern. Then, the barycenter coordinates of the first projection distribution in the character circumscribing frame are set as the first dividing point coordinates, and the region Kt of the character circumscribing frame is divided in the direction of the first projection axis by the dividing point coordinates to set the first divided area. Then, find the barycenter coordinates on the projection axis of the projection distribution in the first divided area, and use the barycenter coordinates of the first projection distribution as the new division point coordinates to calculate the barycenter coordinates of the first projection distribution in the first divided area. The barycentric coordinate series on the first projection axis is obtained by repeating the process of obtaining , a predetermined number of times. In addition, the centroid coordinates of the second projection distribution within the character circumscribing frame are set as the first other dividing point coordinates, and the area of the character circumscribing frame is divided in the direction of the second projection axis using the other dividing point coordinates, and the second division is performed. Set the class ti, find the coordinates of the center of gravity on the projection axis of the second projection distribution in the second divided area,
The process of determining the barycenter coordinates of the second projection distribution in the second divided area is repeated a predetermined number of times by using the barycenter coordinates of the second projection distribution as new other division point coordinates to obtain the barycenter coordinate series on the second projection axis. demand.

The feature extraction unit 18 takes each point or arbitrary point within the character circumscribing frame as a point of interest for each of the extracted sub-patterns, and scans the sub-pattern in another direction that is orthogonal to the scanning direction at the time of extracting the sub-pattern and passes through the point of interest. Then, the character line position M% on the scanning line is detected. Then, based on the order 1!1111 of these points of interest and character line positions and the barycenter coordinate series, the barycenter coordinate numbers are obtained by adding these points of interest and character line positions "M" to the barycenter coordinate series in order of the magnitude of the coordinate values. Convert to a converted value to be used as a scale.

Then, the conversion distance between the point of interest and each of all character line positions on the scanning line on one side of the point of interest is determined based on the conversion value, and the sum of the powers of each conversion distance on one side (
K is a constant) is extracted as the first feature regarding the point of interest. In addition, the conversion path MIFr conversion value between the point of interest and each of all character line positions on the scanning line on the other side of the point of interest is calculated, and the sum of the powers of each conversion distance on the other side (K is a constant) is extracted as the second feature regarding the point of interest.

Further, in FIG. 1, numeral 20 indicates an example of a character recognition device configured using the feature extraction device 10, and this character recognition device N2
0 includes a preprocessing unit 22 for obtaining a character/figure pattern to be recognized, a character frame detection unit 24 for detecting a character circumscribing frame of the character/figure pattern, and a character frame detection unit 24 for scanning the character/figure pattern to detect line widths related to the character/figure pattern. Line width calculation section 26 and feature extraction section! 1
0, a feature vector extraction unit 28 that extracts feature vectors of recognition targets, an identification unit 30 that identifies recognition targets based on the feature vectors, and a word 'N32 that stores feature vectors of standard character patterns.

Hereinafter, the feature extraction device 10 of this embodiment and the character recognition device 12 of the example shown in FIG. 1 will be explained in more detail.

(Preprocessing unit) In this example, the preprocessing unit 22 is a photoelectric conversion unit 3 that outputs a character/figure pattern of a digital signal quantized into black and white binary.
4, and a pattern register 36 for storing character/graphic patterns.

The photoelectric conversion unit 34 decomposes a predetermined reading area of an information medium such as a form into, for example, 128 x 128 pixels, converts the reflected light from the information medium into a black and white 2Ja digital signal (character/figure pattern) for each pixel, A signal is output to the pattern register 36 and line width calculation section 26. The white bits and black bits of the character figure pattern represent the character line and character background, respectively. The photoelectric conversion section 34 in this example optically reads each negative character and outputs a character/figure pattern for one character.

The pattern register 36 stores the character/figure pattern from the photoelectric conversion section 34. An X-Y coordinate system is virtually set on the pattern register 36, and pixel data at pixel positions expressed in this X-Y coordinate system can be written to or read from the pattern register 36. It gives them the freedom to do what they want. The pattern register 36 has a memory capacity of, for example, 128×128 bits, corresponding to the number of pixels in the reading area. Note that the average character-to-character size is 60x60 bits.

FIG. 2 is a diagram showing an example of a character/figure pattern. In the figure, 08G is a text/figure pattern stored in the pattern register 36, and in this example, the text/figure pattern is a pattern containing only one unit of text/figure. Further, the X and Y axes in the figure are coordinate axes set on the pattern register 36, and the origin is at the lower left of the two-dimensional plane in the pattern register 36.

(Character frame detection unit) The character frame detection unit 24 scans the character figure pattern in the pattern register 36 and determines the left end coordinate XL and right end coordinate X of a rectangular frame (character circumscribing frame) circumscribing one unit of character figure of the character figure pattern.
The R1 upper end coordinate Y□ and lower end coordinate Y are detected. In FIG. 2, the character circumscribing frame is shown by a dashed line with the symbol G. Character circumscribing frames are (XL, YT), (XL, Yl
It becomes a rectangular frame connecting four points: l ), (×6, YT), and (Xs , Ye ).

The character frame detection unit 24 detects these coordinates XL, Xs, Yt
and Y8 are output to the barycenter coordinate sequence extraction section 16 and the feature vector extraction section 28.

(Line Width Calculation Unit) The line width calculation unit 26 inputs the character/figure pattern from the photoelectric conversion unit 34 and calculates, for example, the number Q of states in which all points in a 2×2 window are black bits, and the character/figure pattern in the character/figure pattern. The number of all black bits is counted, and the line width W is calculated according to the conventionally well-known formula (1).

w=A/(A-Q) -.-(1) Note that if the line width is known in advance, the line width calculation section may be omitted.

(Sub-pattern extraction unit) Sub-pattern extraction unit 12 Scans the character/figure pattern in multiple directions, detects the number of consecutive black bits in each scanning line, and extracts the number of consecutive black bits in each scanning direction based on the number of consecutive black bits and the line width. The sub-pattern extraction unit 30 extracts a plurality of sub-patterns corresponding to the 45° counterclockwise from the axis (45° diagonal to the right) and 45° clockwise (45° diagonal to the left)
5° direction) is set as the main scanning direction in the pattern register 36.
The character/figure pattern is scanned and vertical, horizontal, diagonal, and left diagonal sub-patterns corresponding to each main scanning direction are extracted.

Therefore, the sub-pattern extraction section 12 includes a vertical sub-pattern extraction section, a horizontal sub-pattern extraction section, a right diagonal sub-pattern extraction section, and a left diagonal sub-pattern extraction section, although not shown.

The vertical sub-pattern extraction section scans the entire character/figure pattern of the pattern register 36 with the vertical direction as the main scanning direction, and detects continuous black bits (black runs) on the vertical scanning line. This black run represents the cross section of the character line, and a black run with a length l that satisfies the following equation (2) is extracted from the detected black rough.

l≧N-w (2) However, l is the length of the black run in the main scanning direction (cross-sectional layer of the character line), N
is a constant set arbitrarily and suitably for each sub-pattern. In this embodiment, for example, N=2 for all subpatterns.

The vertical sub-pattern extraction section regards the black runs that satisfy equation (2) 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 (2) are considered white hits.

Similarly, the horizontal, right diagonal, and left diagonal sub-battery extraction sections scan the character/figure patterns in the pattern register 36 using the horizontal, right diagonal, and left diagonal directions as the main scanning directions, and detect black runs on the scanning lines in the respective main scanning directions. A black run that satisfies equation (2) is extracted from among them, and the extracted black run is regarded as a black run constituting a subpattern and is stored in horizontal, right diagonal, and left diagonal subpattern memories (not shown).

An X-Y coordinate system is set on each sub-pattern memory in the same way as the pattern register 36, and black runs forming each sub-pattern are stored at pixel positions corresponding to pixel positions on the pattern register 36. Note that each sub-pattern memory has a memory capacity of, for example, 128×128 bits, similar to the pattern register 36.

Fig. 3 is a diagram showing sub-patterns extracted from the character/figure pattern shown in Fig. 2, and Fig. 3 (A), (B), (C)
and (0) respectively indicate vertical sub-pattern vSP, horizontal sub-pattern SP, right diagonal sub-pattern R8P and left diagonal sub-pattern LSPt. This is the coordinate axis set to .

(Pattern Projection Unit) In this embodiment, the first and second projection axes are the X-axis and Y-axis set on the pattern register 36.

The character projection unit 14 of this embodiment projects the character figure pattern within the character circumscribing frame onto the X-axis and Y-axis that are orthogonal to each other, and obtains a black bit distribution 5X(x) as a first projection distribution and a black bit distribution as a second projection distribution. Find bit distribution 5Y(y) lj&.

The black bit distributions 5X(x) and SY(y) are obtained according to the following equation (3).

y=Yt However, x and y represent the coordinates in the X-Y coordinate system on the pattern register 36, and take integer values from 0 to 127, for example, and P(
x,y) represents the black bit (significant color) when P(x,y) = 1 and the white bit (background color) when P(x,y) = O, and XL, XR, YT and Y8 are the same. Represents the coordinates of the left end, right end, top end, and bottom end of the circumscribed frame of the character figure of the unit.

(Barycenter coordinate series extraction unit) The barycenter coordinate series extraction unit 16 of this embodiment extracts the barycenter coordinate series X (M,) of the black bit distribution 5X(x) as the barycenter coordinate series on the first projection axis, and The barycenter coordinate series Y of the black bit distribution SY(y) is the barycenter coordinate series on the second projection axis.
(M,)1Fr 0 barycentric coordinate series X(M,), Y
(M,) is the sum of the first moments in the area within the character circumscribing frame or the divided area, and the number of black bits in that area (
It is obtained by dividing by the black bit distribution 5X(x), 5Y(y). However, M and M are barycenter coordinate numbers that are attached to the barycenter coordinates in order of the size of the coordinate value, and in this example, 0M, which is attached in order of the coordinate value from small to large, is 1≦M.
, ≦MXt, MX is an odd number and represents the total number of barycenter coordinates detected as the barycenter coordinate series X(M,), and M is a natural number that satisfies 1≦M, ≦MY. is an odd number and represents the total number of barycenter coordinates detected as the barycenter coordinate series Y (M,). MX and MY are set to any suitable value, but preferably a value close to the size of the character figure pattern. For example, if the size of the character/graphic pattern is 60 x 60 bits, then approximately 15 or 31 pieces may be used.

Below, we will explain how to detect the barycenter coordinates on the projection axes of the black bit distributions 5X(x) and 5Y(y).In the following explanation, we will assume that MX=MY=7 and use the seven barycenter coordinates to simplify the explanation. It is assumed that X(M,) and seven barycenter coordinates Y(M,) are detected.

In order to obtain the barycenter coordinate series
-X, by dividing the sum of the -th moments of the black bit distribution 5X(x) by the total number of black bits within the range, the barycenter coordinates x (4) of the center barycenter coordinate number 4 are obtained.

Next, the center of gravity coordinates X (4) 1Fr are used as the coordinates of the first division point, and the area of the character circumscribing frame x, ~ x 1. Divide l by the coordinates ×(4), and divide the region XL-X(4) and ×(4) as the first divided region.
) to X+t.

Then, the barycentric coordinates of the black pit distribution in the division class @x, ~X(4) x (2), and the barycenter coordinates of the divided area x (4) to X5 x
(6) Find.

Next, add the barycenter coordinates ×(2) and ×(6) as new division point coordinates, and divide the division point coordinates X(2), X(4), and ×(
6), divide the area XL~x of the character circumscribing frame, and divide the area XL-X(2) as the first divided area and the area X(2)~X(
4), X (4) ~X (6), and ×(6) ~x,
Set I.

Then, the coordinates of the center of gravity in the divided area XL~X(2) x (1)
, the barycentric coordinates in X (2) to X (4) × (3), the barycenter coordinates in × (4) to X (6) × (5), and × (6
) to x, barycentric coordinates in l×(7).

Similarly, the barycenter coordinate series extraction unit 16 extracts the barycenter coordinate series Y
To find (M,), first, divide the sum of the -th moments of the black bit distribution 5Y(y) in the entire range Y@~YT of the circumscribed character frame in the X-axis direction by the total number of black bits within the range. By doing this, the center of gravity coordinate Y(
Find 4).

Next, the area Y8 to YT of the character circumscribing frame is divided by the coordinate Y(4) using the centroid coordinate Y(4) as the first dividing point coordinate, and the areas Y, ~Y(4) and Y(4) are divided as the second divided area. )~YT
and set the centroid coordinates Y(2) of the divided areas Y, ~Y(4)
) and the center of gravity coordinates Y(6) of Y(4) to Y.

Next, the barycentric coordinates Y(2) and Y(6)! Add new dividing point coordinates, divide area Y11 = Yyt of character circumscribing frame,
Region Y, ~Y(2) and Y(2) as the second divided region
~Y(4), Y(4)~Y(6), and Y(6)~YT
and set the centroid coordinates Y(
1), the center of gravity coordinates Y(3) within Y(2) to Y(4),
Centroid coordinates Y(5) and Y(6) within Y(4) to Y(6)
- Find the center of gravity coordinates Y(7) in YT.

4 and 5 are diagrams showing the relationship between the character/figure pattern 0th day G, the black bit distribution 5Y(y), and the barycenter coordinate series Y(M,). (A), (8), and (C) of these figures are the character figure pattern, the black bit distribution 5Y (y) for the character figure pattern of figure (A), and the center of gravity coordinates between the character figure pattern of figure (A). The series Y(M,) is shown.

Character figure pattern 08G in Figure 4 (A) and Figure 5 (A)
shows a pattern of characters that should be recognized as the same character "upper" but whose character line positions are partially different, and as is clear from these figures, the centroid coordinate Y ( It can be understood that M,) moves.

(Feature Extraction Unit) The feature extraction unit 18 scans the sub-pattern in another direction orthogonal to the scanning direction at the time of extracting the sub-pattern, and extracts a number representing the distribution state of character lines when viewed from the point of interest in the other direction. Extract the first and second features. These first and second features are extracted for each subpattern.

Therefore, in this embodiment, the other directions for extracting the first and second features are the horizontal direction (X-axis direction) with respect to the vertical sub-pattern vSP, the vertical direction (Y-axis direction) with respect to the horizontal sub-pattern H8P, and the right direction with respect to the horizontal sub-pattern H8P. Regarding the diagonal sub-pattern R3P, there is a left diagonal direction (a direction shifted by 45 degrees clockwise from the X axis) and a right diagonal direction (a direction shifted by 45 degrees from the X axis in the direction of the half hour hand) between the left diagonal sub patterns LSP, and this implementation. Although not shown, the feature extraction unit 18 in the example includes a VSP horizontal feature extraction circuit, an H3P vertical feature extraction circuit, a SP left diagonal feature extraction circuit, and an LSP right diagonal feature extraction circuit. The vsp horizontal feature extraction circuit extracts the first and second horizontal features (horizontal features) for vSP, and the mouth SP vertical feature extraction circuit extracts the first and second vertical features (vertical features) for H4F. R
The 3P left diagonal feature extraction circuit extracts the first and second features in the left diagonal direction per SP (left diagonal feature r!1)%, and the LSP right diagonal feature extraction circuit extracts the first and second features in the right diagonal direction per LSP. Extract the features (right diagonal features).

FIG. 6 is a functional block diagram for explaining the vSP horizontal feature extraction circuit, and includes the VSP horizontal feature extraction circuit, the vertical sub-pattern memory of the sub-pattern extraction section that is mistaken for this circuit, and the barycenter coordinate series memory of the barycenter series extraction section. shows.

In the same figure, 38 is a vSP horizontal feature extraction circuit, 40 is a vertical sub-pattern memory of a sub-pattern extraction unit that stores vertical sub-patterns, and 42 is a barycenter coordinate series X (M, ) and Y
The barycenter coordinate sequence memory of the barycenter coordinate sequence extraction unit that stores (M,) is shown.

The VSP horizontal feature extraction circuit 3 includes a control circuit 44 that controls the entire feature extraction section 18, and a control circuit 44 that controls the entire feature extraction section 18, and while scanning the subpattern memory, refers to the contents of the subpattern memory and detects a change point where a white bit changes to a black bit. A change point detection circuit 46 for detecting a change point, a change point memory 48 for storing the coordinates of a black bit when changing from a white bit to a black bit as change point coordinates, and the number of change points stored in the change point memory. and a change point counter 50 that specifies the address of the minute change point memory.

Roughly speaking, the vSP horizontal feature extraction circuit 38 uses a pattern memory address counter 52 that generates X and Y coordinates, and coordinates of a changing point to determine whether the changing point is located on one side or the other side of the point of interest. A comparator 54 that compares the coordinates of the point of interest with the coordinates of the point of interest, an arithmetic circuit 56 that calculates the power of the transformed distance between the point of interest and the point of change, and a feature (
A register 458 for extracting the first feature), a register B60 for extracting the feature on the other side of the point of interest (M2 feature), and a register B60 for extracting the first and second features for each point of interest. and a vSP horizontal feature pattern memory 62 that stores the vSP horizontal feature pattern memory 62. For example, the capacity of register A58 and register B60 is 1
I have a 6-bit house.

FIG. 7 is a diagram for explaining the VSP horizontal feature pattern memory, and FIG. 8 is a diagram showing features regarding one point of interest.

As shown in FIG. 7, the vSP horizontal feature pattern memory 62
is, for example, 12 in the
For example, it has 32 memories of 8 bits and 128 bits in the Y-axis direction. In this case, as shown in Fig. 8, the feature of one point of interest can be represented by 32 bits. For example, the upper 16 bits represent the first feature of the point of interest, and the lower 16 bits represent the second feature of the point of interest. can express the characteristics of vSP horizontal feature pattern memory 62! The X and Y axes corresponding to the X and Y axes on the vertical sub-pattern 40 are set on each of the 32 memories making up the memory. The first and second characteristics can be stored.

Next, the overall operation of the vSP horizontal feature extraction circuit 38 will be schematically explained.

The VSP horizontal feature extraction circuit 38 first scans the vertical subpattern horizontally from a point (XL, Y) on the left side of the 0 character circumscribing frame (where Y, ≦y≦YT) to scan the vertical subpattern in order to detect a change point. While registering the detected change point coordinates in the change point memory 48, select the point (
x*, Y). The coordinates of a black bit when it changes from a white bit to a black bit on a scanning line are detected as change point coordinates, and the change point coordinates are detected for each scanning line.

Next, the vSP horizontal feature extraction circuit 38 scans the point of interest to extract the first and second features.

In horizontal feature extraction, in order to determine the positional relationship between the point of interest and the point of change, the X coordinate TX of the point of interest and the X coordinate Cx of the point of change are
Compare the size and size of the booth. If the coordinate TX is larger than the coordinate Cx, it means that the changing point is located on one side of the point of interest, and if the coordinate TX is smaller than the coordinate Cx, it means that the changing point is located on the other side of the point of interest. .

Also, compare the X coordinate TX of the point of interest with the barycenter coordinate series .

・X(M,)≦TX<X(M,-1)(7) To!
・When TX=XL, TX,=0 ・When rx=x*, TX,=MX+1 However, here, M is an integer in the range of 0≦M2≦MX, and X(0)=XL and X (Mx ◆I)=X*.

In the same way as for the point of interest, compare the X coordinate Cx of the changing point with the barycenter coordinate series, and convert the coordinate Cx to @(conversion value CXH) using the barycenter number M as a scale according to the comparison result. .

In this embodiment, to calculate the transformed distance of the point of interest and the point of change to the power of 2, 2 is used. For horizontal feature extraction, D=
There are l CXH-TX and l Te.

Then, based on the determination result of the positional relationship between the point of interest and the point of change, the first feature is determined by sequentially adding the transformed distances of the points of interest and points of interest located on one side of the point of interest. , the change point located on the other side of the point of interest, and the power of the conversion path HD of the point of interest are sequentially added to obtain the second feature.

Next, a more specific operation of the vSP horizontal feature extraction circuit 38 will be explained using an example.

When scanning of the vertical sub-pattern begins, pattern memory address counter 52 generates X and Y coordinates for horizontal scanning of the vertical sub-pattern, and outputs these coordinates to vertical sub-pattern memory 40 and change point memory 48. The vertical sub-pattern memory 40 outputs pixel data of the scanning point specified by the input X and Y coordinates to the change point detection circuit 46.

The change point detection circuit 46 determines whether the input pixel data is a white hit or a black bit, and when a black bit is detected next to the white bit, the black bit is detected as a change point.

When the change point detection circuit 46 detects a change point, the change point counter 50 increments the count value i by 1 to count the number of change points detected on one scanning line. The initial value is zero at the start of line scanning)
Then, the count [li is outputted to the change point memory 48 as an address specifying the storage location of the change point coordinates. When the change memory 48 receives the specified address i, it stores the coordinates input from the pattern memory address counter 52 in the specified address i as the coordinates of the first detected change point on one scanning line.

When the detection of the change point within the character circumscribing frame on one scanning line is completed, the point of interest on the scanning line is scanned.

In this example, all points within the character circumscribing frame are set as points of interest, and therefore, each point on the scanning line is set as a point of interest, and the points of interest are sequentially scanned horizontally.

FIG. 9 is an explanatory diagram of the horizontal feature extraction process for one point of interest after detecting a change point.

Hereinafter, with reference to FIG. 9, an example of the horizontal feature extraction process for one point of interest will be described.

When the detection of the change point on one scanning line is completed, the process of extracting the horizontal feature regarding one point of interest on the scanning line starts (5TART).The count ii of the change point counter 50 changes at the time of 5TART. This value represents the total number of changing points on the scanning line for which point detection has been completed.

When the process starts, the control circuit 44 initializes the register A 58 and the register B 61 (Sl, S2), and at the same time, the comparator 54, the arithmetic circuit 56, and the vSP horizontal feature pattern memory 62 select the target point from the pattern memory address cursor 52. Enter the coordinates of

Next, the control circuit 44 determines whether extraction 1Fr of horizontal features between one point of interest has been completed (S3).

If the count value i of the change point counter 50 is i=O in the determination in S3, the horizontal feature extraction for one point of interest is completed, and the values stored in the register A58 and register 860 at this time are respectively related to the point of interest. Since these represent the first and second features, the control circuit 44 stores the values stored in the register A58 and register B60 in the storage locations designated by the coordinates of the point of interest in the vSP horizontal feature pattern memory 62, respectively. Thereafter, the horizontal feature extraction process for one point of interest is ended (END).

If the count value i of the change point counter 50 is 1≠O in the determination in S3, this indicates that horizontal feature extraction between one point of interest has not been completed, so the change point memory 48 stores the change at this time. The changing point coordinates of the address specified by the count value l of the point counter 50 are input to the comparator 54 and the arithmetic circuit 56.

Next, the arithmetic circuit 56 converts the coordinates of the point of interest and the point of change based on the input coordinates of the point of interest and the point of change and the barycenter coordinate series from the barycenter coordinate series memory 42 (S4
), the conversion word MD of the point of interest and the point of change is determined based on the roughly determined conversion value.

Further, the comparator 54 compares the input coordinates of the point of interest and the point of change, and inputs the comparison result to the control circuit 44 .

Next, the control circuit 44 determines whether the change point is on one side or the other side of the point of interest based on the comparison result from the comparator 54 (S5).

If CX<TX in the determination in S5, the change point is considered to be located on one side of the point of interest, and the control circuit 44 uses the calculation circuit 56
The converted distance calculated by DK is stored in the register 458.
86), and then performs the process of S8.

If Cx≧TX in the determination in S5, it is assumed that the change point is located on the other side of the point of interest, and the control circuit 44 uses the arithmetic circuit 56
The converted distance R to the power (D) calculated by is stored in register B60.
(S7), and then the process of S8 is performed.

In S8, the control circuit 44 calculates the conversion distance between the next other change point and the point of interest and uses the change point counter 50 to extract features.
The count [1i is decreased by 1, and then the process returns to S3.

The count value 1 of the change point counter 50 is sequentially incremented from the total number I of change points detected on the scanning line until i=0, and the conversion word jlfffD of each change point and the point of interest is raised to the power of the conversion word jlfffD in order. Alternatively, the first or second characteristic m can be extracted by adding it to the value stored in the register B60.

The above process is performed for each point of interest on the scanning line to extract the first and second features, and once the first and second features have been extracted for all points of interest on the scanning line, the next A point of change is detected on the scanning line, and first and second features of the point of interest are extracted.

The H3P vertical, R8P left diagonal, and SP right diagonal feature extraction circuits have the same configuration as the vSP horizontal feature extraction circuit shown in FIG. First and second features between the points of interest of the subpattern and the left diagonal subpattern are extracted.

Hereinafter, the first and first feature extractions that close to the points of interest of the horizontal sub-pattern, right diagonal sub-pattern, and left diagonal sub-pattern will be schematically explained.

The H9P vertical feature extraction circuit first scans the horizontal sub-pattern to detect changing points. A point on the bottom edge of the character circumscribing frame (
Vertical scanning of the horizontal sub-pattern is started from X, VB) (where XL≦X≦Xl+), and while registering the coordinates of the detected change point, it scans to the point (X, YA) on the upper side of the character circumscribing frame.

Next, the point of interest of the mouth SP vertical feature oil extraction circuit is scanned.

In vertical feature extraction, the Y coordinate TY of the point of interest and the Y coordinate of the point of change are
The magnitude relationship with the coordinate CY is compared to determine whether the changing point is located on one side or the other side of the point of interest.

Also, the Y coordinate TY of the point of interest is compared with the barycenter coordinate series Y (M,), and the coordinate TV is converted into [(conversion!TV, ) with barycenter coordinate number M and 1 scale according to the comparison result.

・When X (M,)≦TY<X (M,◆1)・T
When V=Y8, TYH”0 ・When TV=YT, TV, = MY+ 1 However, here, M is an integer in the range of 05M, ≦MY, and Y(0)”Ya and X(MY◆I)= Let it be YT.

Similarly, the Y coordinate CY of the change point and the barycenter coordinate series Y(M
, ), and the coordinate CY is converted into a value (converted value CV+) using the barycenter number M as a scale according to the comparison result.

Then, the transformed distance of the point of interest and the point of change is calculated.

For vertical feature extraction, D = l CY, -Tv.

It is.

Then, based on the determination result of the relationship between the positions of the point of interest and the point of change, the powers of the conversion word MD are sequentially added by f to obtain the first and second features.

Further, the R3P left diagonal feature extraction circuit first scans the right diagonal sub-pattern to detect a change point. The left diagonal scan of the right diagonal sub-pattern is started from the points on the left side and the bottom side of the character circumscribing frame, and is scanned to the points on the top and right side of the character circumscribing frame while registering the detected change point coordinates.

Next, the SP left diagonal feature extraction circuit scans the point of interest.

In extracting the left diagonal feature, the magnitude relationship between the coordinates TX or TV of the point of interest and the coordinates Cx or CY of the point of change is compared, and it is determined whether the point of change is located on one side or the other side of the point of interest. do.

In addition, the coordinates TX and Y of the point of interest and the barycentric coordinate series X (M,
) and Y (M,), coordinates TX and TY'8
According to the comparison results of 1 and 2, the barycenter coordinate numbers M and M9 are converted into converted values TX and TY using M9 as a scale.

Similarly, the coordinates Cx and CY of the change point are compared with the barycenter coordinate series X (M,) and Y (M,), and the coordinates Cx and CY are converted into converted values Cx8 and CYH according to the comparison result. do.

Then, the power of the converted word HD at the point of interest and the point of change is calculated.

In extracting the left diagonal feature, the conversion word MD is expressed as in the following equation.

Then, based on the determination result of the magnitude M relationship between the point of interest and the point of change, the powers of the conversion word MD are sequentially added to obtain the first and second characteristics.

In addition, the LSP right diagonal feature extraction circuit first scans the left diagonal subpattern in order to detect a change point.The left diagonal subpattern is first scanned from points on the lower side and on the right side of the 0 character circumscribing frame, and the detected change point is detected. While recording the coordinates, scan to the points on the left and top sides of the character circumscribing frame.

Next, the LSP right diagonal feature extraction circuit scans the point of interest.

To extract the right diagonal feature, in the same way as the left diagonal feature extraction, it is determined which side of the point of interest the change point is located on, one side or the other, and the coordinates TX and TY of the point of interest are determined. The coordinates Cx and CY of the change point are converted into the converted value TX and the coordinate Y
s and converted values CXH and CY.

Then, in the same manner as in the case of left diagonal feature extraction, the power of the conversion word MD of the point of interest and the point of change is determined, and the first and second features are determined.

Note that the conversion distance can generally be expressed as shown in the following equation.

However, in horizontal feature extraction, the Y coordinates of the point of interest and the point of change are equal, and in vertical feature extraction, the X coordinates of the point of interest and the point of change are the same, so the equation representing the transformation distance can be simplified as described above.

(Feature matrix extraction unit) The feature matrix 2 extraction unit 28 divides the inside of the character circumscribing frame into MXN partial regions based on the coordinates of division points for creating a feature matrix. Then, for the vertical, horizontal, right diagonal, and left diagonal sub-patterns, MxNx two dimensions whose elements are the arithmetic mean of the first feature and the arithmetic mean of the second feature of the points of interest existing in each partial region. The horizontal, vertical, left diagonal, and right diagonal feature vectors, which are feature vectors of

Although the dividing point coordinates for creating the feature map may be obtained by a conventionally known method, in this embodiment, the barycenter coordinate series X(MP) and '/
CM, )lFr The inside of the character circumscribing frame is divided using the dividing point coordinates for creating the feature matrix, and M=MX+1 and N=MY+1.

Hereinafter, a method for obtaining a feature vector will be explained by taking as an example the extraction of a horizontal feature vector obtained from horizontal features.

The inside of the character circumscribing frame is divided into MXXMY partial regions using the barycenter coordinate series. Then, the arithmetic mean of the first feature and the arithmetic mean of the second feature are determined for each partial region of the vertical sub-pattern. ILl,1. , and HR-, (m=1.2,...MX and n=1.2,
..., MY), position (x
, y), the first feature and the second feature of the point of interest in (l
If X*V and rX+V, and the number of points of interest in the partial region of the m-th row and n-th column are T Ill + n, then )IL□ and HR, , .

By definition like Ichiji expression Wear.

HL, , , and HR, found in each partial region,
,. is (MX+ 1) X (MY+ 1) X 2
Represents a dimensional horizontal feature vector.

The feature vector extraction unit 28 refers to the mouth SP vertical feature pattern memory, the 8SP left diagonal feature pattern memory, and the LSP right diagonal feature pattern memory in the same way as the horizontal feature vector, and extracts (MX+1) x (MY+1) x 2 dimensions. Find vertical, left diagonal, and right diagonal feature vectors.

These (MX+1)x (MY+1)x2-dimensional horizontal,
Vertical, left diagonal, and right diagonal feature vectors are recognized (7)
(MX+1)x (MY+1)x2x4-dimensional feature vector.

(Identification unit) The identification unit 30 recognizes the recognition target (7) (MX+1) x (Y+1)X2X4 extracted by the feature bellator extraction unit 28.
The dimensional feature vector is compared with the standard character feature vector stored in the dictionary 32 to identify the recognition target.

The dictionary 32 is extracted by the feature vector extraction unit 28 (MX
+1)x(MY+1)x2x4-dimensional feature vectors, feature vectors of a plurality of standard characters written in the same format, and character names of the standard characters, such as character codes defined in the JIS standard, are stored.

The identification unit 32 calculates the Euclidean distance M between the feature vector to be recognized and the feature vector of the standard character, and outputs the name of the standard character with the minimum distance as the recognition result C.

The present invention is not limited only to the embodiments described above, and therefore, the configuration, operation, flow of operation, input/output signals, flow of input/output signals, numerical conditions, shape, dimensions, and arrangement position of each component. Il can be changed as desired.

(Effects of the Invention) As is clear from the above description, the feature extraction device of the present invention can detect the projection distribution (
(marginal distribution), convert the distance M between the subpattern's point of interest and the character line into a measure that follows local positional fluctuations of the character line, and extract the sum of the powers of this converted distance as a character feature. Therefore, even if there is a change in the character line position, the change in the characteristics of characters that should be recognized as the same character can be reduced, and the characteristics are stabilized.

Furthermore, the feature extraction process can be performed easily and the processing time can be shortened.

Therefore, by configuring the character recognition bag M using the feature extraction bag H of the present invention, the recognition rate does not deteriorate without having to prepare a large amount of dictionary data corresponding to variations in the character line position, and therefore the character recognition Therefore, it is possible to provide a character recognition device that can perform dictionary matching with high accuracy and shorten dictionary matching time compared to conventional ones.

[Brief explanation of drawings]

Fig. 1 is a functional block diagram providing a detailed explanation of the present invention, Fig. 2 is a diagram showing an example of character/figure Bakun to be recognized, and Figs. 3 (A) to (D) are characters shown in Fig. 2. Diagrams showing sub-patterns extracted from figure patterns, Figures 4 (A) to (C)
) and FIGS. 5(A) to 5(G) are diagrams showing the relationship between the character figure pattern, the projection distribution 5Y(V), and the barycentric coordinate series Y(M, ), and FIG. 6 is the diagram of the vSP horizontal feature extraction circuit. 7 is a diagram for explaining the feature pattern memory, FIG. 8 is a diagram for explaining the feature regarding one point of interest, and FIG. 9 is a diagram for explaining the feature of one point of interest. FIG. 4 is a diagram illustrating a process of extracting first and second features related to FIG. IO...Feature extraction device, 12...Sub pattern extraction unit 14...Character projection unit, 16.-Barycenter coordinate series extraction unit+8-...Feature extraction unit. Figure 9: Extraction process of first and second features between one point of interest

Claims (1)

[Claims] (1) Scan the character/figure pattern in a plurality of predetermined directions, detect the cross-section of the character line in each scanning direction, and extract the character line of the cross-section whose cross-sectional length is sufficiently longer than the character line width related to the character/figure pattern. a sub-pattern extraction unit that extracts sub-patterns representing stroke components in each scanning direction; a character projection section for determining the projection distribution; determining the barycentric coordinates on the projection axis of the first and second projection distributions within the character circumscribing frame of the character figure pattern; and determining the barycentric coordinates of the first projection distribution within the character circumscribing frame of the character figure pattern set the coordinates of the first dividing point, divide the area of the character circumscribing frame in the direction of the first projection axis by the dividing point coordinates, and set a first divided area;
Find the barycentric coordinates on the projection axis of the projection distribution in the first divided area, and set the barycentric coordinates of the first projected distribution as new dividing point coordinates to determine the barycentric coordinates of the first projected distribution in the first divided area. The barycentric coordinate series on the first projection axis is obtained by repeating the obtaining process a predetermined number of times, and the barycentric coordinates of the second projection distribution within the character circumscribing frame are set as the coordinates of the first other division point, and the area of the character circumscribing frame is A second divided area is set by dividing it in the direction of the second projection axis using other dividing point coordinates, and the center of gravity coordinates on the projection axis of the second projection distribution in the second divided area are determined. The process of calculating the barycenter coordinates of the second projection distribution in the second divided area is repeated a predetermined number of times by using the barycenter coordinates of the two projection distributions as new other division point coordinates, and the barycenter coordinate series on the second projection axis is obtained. A coordinate series extraction unit, for each of the extracted sub-patterns, takes each point or arbitrary point within the character circumscribing frame as a point of interest, and converts the sub-pattern into another direction that is perpendicular to the scanning direction at the time of extracting the sub-pattern and passes through the point of interest. to detect the character line position on the scanning line, and based on the positional relationship between these points of interest and character line positions and the barycentric coordinate series, coordinate values of these points of interest and character line positions are added to the barycentric coordinate series. Convert the barycenter coordinate numbers assigned in order of size to a conversion value as a scale, and calculate the conversion distance between the point of interest and each of all character line positions on the scanning line on one side of the point of interest as the conversion value. The sum of the K powers of each transformation distance on one side (K
is a constant) as the first feature regarding the point of interest, and calculate the converted distance between the point of interest and each of all character line positions on the scanning line on the other side of the point of interest based on the converted value. , the sum of the K powers of each transformation distance on the other side (K
is a constant) as a second feature regarding the point of interest.