JPH01136287A - Feature extracting system - Google Patents

Abstract

PURPOSE:To stably and effectively extract a feature by extracting a character frame split point coordinate system from the centroid coordinate of projection on two axes of a character graphic pattern, and extracting the information of a pole which satisfies a prescribed condition. CONSTITUTION:Character graphic pattern information from a photoelectric transfer part 2 is inputted to a character projection generating part 5 via a pattern register 3. The character projection generating part 5 finds the distribution of the number of black bits by projecting an input character graphic pattern to desired two axes, and a centroid detecting part 6 finds the centroid coordinate of the input character graphic pattern. A split coordinate system is decided by a character frame split point decision part 7 based on the centroid coordinate. A pole counter part 10 extracts the information of the pole which satisfies the prescribed condition from an external coordinate representing the outer shape of the character graphic pattern, and the existence of a reference pattern is decided at an identification part 12 by the information of the pole.

Description

DETAILED DESCRIPTION OF THE INVENTION (Industrial Application Field) The present invention relates to a feature extraction method applied to character recognition devices, etc., and in particular to feature extraction that effectively and stably extracts differences between patterns that have only minute differences. Method In conventional character and figure recognition, strokes are extracted from character and figure patterns,
Many recognition methods are employed that use the positions, lengths, and mutual relationships between strokes of these extracted strokes. The method is (1) Calculate the curvature of the contour point series detected by tracing the contour of the character figure, divide the contour series using points with large values of curvature as dividing points, and assemble the divided series. (2) Extract the strokes by combining them, or (2) perform thinning processing on the character/figure pattern to create a skeleton, trace the connectivity and skeleton pattern of the skeleton pattern, detect sudden angle changes, etc., and extract the strokes. The strokes extracted from (1) and (2) above are identified by extracting their geometric features.

and (3) determining the distance from each side of the character rectangular frame of the character figure pattern to the character figure pattern, and using the distance, extracting the protrusion of the character figure pattern corresponding to each side of the character rectangular frame, Japanese Patent Laid-Open No. 60-230282 proposes a method of character recognition using the extracted protrusion f9.

(Problems to be Solved by the Invention) However, in method (1), as the character/graphic pattern becomes larger or more complex, the amount of processing increases and the processing speed decreases.

In the method (2), it is necessary to thin the character/figure pattern, and the thinning causes problems such as dark spots and whiskers in the pattern, which complicates subsequent processing. In method (3), the extracted features become unstable when the shape of the human characters changes, and this effect is especially noticeable between similar characters such as ``water'', ``ice'', and ``dai'' and ``thick''. .

An object of the present invention is to divide each of (MXN) regions based on the peripheral distribution of a character/figure pattern without performing unnecessarily complicated processing such as contour tracing or line thinning in conventional recognition methods. By extracting information about the extreme points that satisfy a predetermined condition from the contour coordinate series representing the contour of a character figure pattern, it is stable even when the character shape changes, especially for patterns with only minute differences. The object of the present invention is to provide a feature extraction method that effectively extracts differences.

(Means for Solving the Problems) The present invention includes a pattern register that stores a pattern obtained by reading and quantizing character figures on a medium as an original pattern, and calculates the characteristics of the character figure based on the original pattern. In order to solve the problems of the prior art with regard to the feature extraction method, (a) a means for detecting a character rectangular frame consisting of a circumscribed side of a ring pattern stored in a pattern register;
(c) means for first projecting the ring pattern onto two desired axes to obtain each black bit number distribution; Determine the centroid coordinates of each bit number distribution, and then calculate the centroid of each of the black bit number distributions for each range obtained by dividing the range limited by the character rectangular frame using each barycenter coordinate detected so far. (d) Means for obtaining people's barycenter coordinate series by repeating the process of determining coordinates multiple times; The divided coordinate numbers in the divided coordinate series on each axis are made to correspond almost equally, one barycenter coordinate in the barycenter coordinate series is associated with one divided coordinate, and the inside of the character rectangle is (M×N (c) scanning the inside of the character rectangular frame of the ring pattern in the vertical direction from each side of the character rectangular frame, and determining the character rectangular frame from the background part; (r) means for calculating the amount of change of each change point based on the change point coordinate series; (g) increase/decrease according to the amount of change; and (h) means for extracting information on the extreme points that satisfy a predetermined condition for each of the (M×N) areas.

(Operation) According to the present invention, since the feature extraction method is configured as described above, each technical means operates as follows.

The means (a) detects a character rectangular frame which is a circumscribing frame of a character graphic by scanning the ring pattern stored in the pattern register, and outputs it to each means (c) and (e). The means (b) creates a black bit number distribution in each direction of two desired axes (for example, the X-axis and Y-axis directions) from the ring pattern, and outputs it to the means (c). (c) means (a)
Receiving the character rectangular frame from the means of (b) and each black bit number distribution from the means of (b), detect the barycenter division coordinates of the ring pattern based on the marginal distribution with respect to the desired two axes, and Output to means. The means (d) uses (M
The area is divided into N) areas and the coordinates of the dividing points are output to the means (h). The means (e) detects the coordinates of the first change point when scanning in the vertical direction from each of the four sides of the character rectangular frame of the ring pattern, the top, bottom, left, and right sides, over the entire range of the above sides. Create a change point coordinate series by
Output to means of ([). The means (f) calculates the amount of change of each change point based on the change point coordinate series from the means (e), and outputs it to the first stage of (g). The means (g) detects the extreme point at which the increase/decrease changes based on the amount of change from the first stage (f), and outputs it to the means (h). The means (h) receives the dividing point coordinates from the means (d) and the pole information from the means (g), and satisfies a predetermined condition for each of the (M×N) regions. The extreme points are extracted and used as features for character/figure recognition. Therefore, it becomes possible to effectively and stably extract differences between patterns in which there are only minute differences, and the problems of the Moeki prior art are solved.

(Example) Examples of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a character recognition device to which the feature extraction method of the present invention is applied.

In FIG. 1, 1 is an optical input from a form, and 14 is a character name output of the character recognition device of this embodiment.

This character recognition device includes a photoelectric conversion section 2, a pattern register 3,
Character frame detection section 4, character projection creation section 5, center of gravity detection section 6, character frame division point determination section 7, outer coordinate series detection section 8, polar point coordinate series detection section 9, polar point counting section 10, polar point counting area table 11. It is composed of an identification section 12 and a dictionary memory 13.

Optical input 1 is input to photoelectric conversion section 2 . The photoelectric conversion unit 2 decomposes one character planned area into 128×12B pixels,
It converts each pixel into a binary digital signal (hereinafter referred to as an input character pattern), and the average size of 1
Characters are expressed by an input character pattern of about 60x60 bits. The pattern register 3 stores input character patterns in a format that can reproduce the X and Y coordinates of each pixel in the character expected area.
It has a capacity of B bits.

In this embodiment, the character frame detection section 4 detects the circumscribed frame of a character by expressing it in the left end coordinate XI, right end coordinate xr, upper end coordinate Yt, and lower end coordinate Yb in the pattern register 3.

The character projection creation unit 5 converts the human character pattern in the pattern register 3 into two desired axes, in this embodiment, the X axis and Yll'dl (
the water γ direction in the two-dimensional coordinates of the pattern register 3, respectively;
The distribution of the number of black bits is obtained by projecting it in the superimposition direction), and the distribution of the number of black bits S X (x) and S Y (y) is created. The black bit number distributions 5X(x) and 5Y(y) are expressed by the following equations.

However, x and y are respectively θ~12 in the pattern register 3.
7, where Yt and Yb are the upper and lower coordinates of the character frame in the Y-axis direction, and xl.

X is the leftmost coordinate and rightmost coordinate in the X-axis direction, and P(x,
y) means a black bit or a white bit; in the case of a black bit (significant color), P (x, y) = 1, and in the case of the own bit (background color), P (x, y) = 0.

In FIG. 2, (b) and (c) show the black bit number distributions 5X(x) and 5Y(y) for the human character patterns (FIG. 2(a)) for the kanji characters "water" and "ice", respectively.

The center of gravity detection unit 6 detects the entire range X of the character frame in the X and Y axis directions,
For each range obtained by dividing the range x1 to X,,Yt to Yゎ by ~X,,Yt-Yb and the center of gravity coordinates detected in the previous process, the black bit number distribution S of each human character pattern is calculated.
The barycentric coordinate series X (
MP), Y (M9) is calculated by dividing the sum of the first moments of each range by the sum of black bits of that range. However, Mp and Mq are the barycenter coordinate numbers assigned in order of the size of the coordinate values, and Mp=
1~MX (MX is the number of centers of gravity in the X-axis direction), Mq=1~
MY (MY is the number of centers of gravity in the Y-axis direction). It is desirable to adopt a relatively large number (compared to the number of divisions) of about 15 as the number of barycenter coordinates MX in the (M
The case of detecting P) will be described. First, the X in the text frame
Targeting the range x1 to Xr in the axial direction, the distribution of the number of black bits in the human character pattern S
By dividing the sum of the first moments of by the sum of black bits in that range, the center of gravity coordinate x (
Find 4).

Next, the two barycenter coordinates X (2) and 4) Calculate using the formula.

Next, the centroid coordinates X(2), X(4) detected so far are
), X(6) Te divided range X, -X(2)
, X(2) ~X(4) , X(4) ~X(6
), X(6) ~X, f, and the four centroid coordinates X(1), X(3), X(5), X
(7) is obtained using equation (5).

To detect the barycenter coordinate Y (M9) in the Y-axis direction, if the number of barycenter coordinates MY to be detected is 7, first, the black bit distribution S Y (y ) is detected, and then the character frame is divided into two by the barycenter coordinates Yt-Y (4),
Black bit distribution S for each of Y (4) to Y
Detect the center of gravity coordinates Y (2) and Y (4) of Y (y), and further detect the center of gravity coordinates Y (2) and Y (4) detected so far.
The respective ranges Yt~Y(2), Y(2)NY(
4), Y (4)NY (6), Y (6)N
The barycenter coordinates Y(1), Y(3), Y(5), Y(7) of the black bit distribution S Y(y) for Yb are
), a total of seven barycentric coordinates Y (1
)NY(7) is detected.

Figure 2 (b) and (c) show the centroid coordinates X (
1) ~X (7) and Y (1) NY (7) are both illustrated as black bit number distributions.

The character frame division point determining unit 7 receives the X,
Center of gravity coordinate series X (MP) in the direction outside the Y-axis.

With Y(Mq) as the divided coordinate series, barycenter coordinate number Mp9
The division coordinate series DX (k,), D is created by associating Mq almost equally with the division coordinate numbers for dividing the character frame of the human character pattern into NX, NY division unit areas, , kJ.
Y (k,) is determined.

In this embodiment, the division unit area can be divided into four formats, NX=4°5.6.8 as the number of divisions in the X-axis direction, and NY=4.5 as the number of divisions in the Y-axis direction. ,6゜8 can be taken,
The division coordinate number in the X-axis direction is (k, = 1 ~ NX-1,
NX = 4゜5.6.8), and the division coordinate number in the Y-axis direction is kJ (kJ = 1 to NY-1, NY = 4.5, 6゜8).
), the division coordinate series DX (k, ), DY (kj) for dividing the character frame into NX, NY division unit areas are determined. The center of gravity coordinate numbers MP and Mq in each direction of the X and Y axes and the divided coordinate numbers in the X and Y axes directions are, k.

The divided coordinate series D
.

In order to determine D Y (k,), a table as shown in Table 1 is prepared, and the number of divisions NN in each direction of the X and Y axes is
, NY from this table, the barycenter coordinate number MP,
Read Mq and divide the center of gravity coordinates X (M,), Y (Mq) corresponding to the center of gravity coordinate number MP9Mq into the coordinates D
Determine as X (ki) and D Y (kJ).

Table 1 This table is for the case where the number of barycenter coordinates MX and MY detected by the barycenter detection unit 6 is 7, but even in the negative case, the barycenter coordinates of the number of divisions in each direction of the X and Y axes are If extra barycenter coordinates remain at that time, it can be created by making correspondence so that one more barycenter coordinate is included starting from both end regions.

Figure 3 shows that when the number of divisions NX and NY in each direction of the X and Y axes is specified as NX=NY=5,
The following shows the correspondence between the divided coordinate series D X (k,), D Y (k,) and the 4[heart coordinate series X (Mp), Y (Mq), and also shows the correspondence between the divided coordinate series DX (ki), D
Y (kJ) TE setting area of tB area (kl.

kJ).

Note that the numbers of divisions NX and NY are determined depending on the complexity of the human characters. Also, if -1 is rejected, the number of divisions is NX.
, NY and perform character recognition again.

As described above, in the character frame division point determining section 7, the division format of the division unit area is NX=
4 formats, 4.5, 6.8, and the number of divisions in the Y-axis direction: NY=4.5, 6. '8 can take four forms. In this example, to simplify the explanation, the number of divisions is NX
The following description will be made assuming that =NY=8. In this case, in the X-axis direction, the center of gravity coordinates X(1), X(2), X(
3), X(4).

X(5), X(6), X(7)! ,: Corresponding division coordinates DX(1), DX(2), DX(3), DX
(4), DX (5).

In the DX(6), DX(7), and Y-axis directions, the center of gravity coordinates are Y(1), Y(2), Y(3), Y(
4), Y(5).

The division coordinates DY (
1), DY(2), DY(:l), DY(4), DY(
5), DY(6), and DY(7) are determined.

In this embodiment, the center of gravity coordinates and the divided coordinates are associated with each other by employing a table, but they can also be associated with each other by executing the arithmetic processing shown in the flowchart shown in FIG. Note that all division results in FIG. 4 are rounded down to the decimal point.

In FIG. 4, in step S1, the number Mct is obtained by dividing the number of centroids MX by the number of divisions NX, and in step S2. M in S3
Find the remainder R of X/N X and R2 by dividing R1 by 2. Also, in step S4, the median value of the number of divisions is determined, and in step 35. In S6, the division number and center of gravity number MP are set to 0. Also, step S7. S8. In S9, each time the division number is incremented by 1, the previously set quotient R2 is multiplied by 1, and the 1n core number MP is increased by MCt. It is checked in step SIO that the quotient R2 is not negative, and unless the quotient R2 is negative, the number of centroid numbers is set to 1 in step Sll.
Then, in step 12, the center of gravity number MP is associated with the division number, and division coordinates DX(Mp) are determined. quotient R
If 2 is negative, in step S13 it is determined whether the current division number is larger than the median value, and if it is larger, the centroid number is increased by one, and if it is smaller, the centroid number set in step S9 is determined. Determine the division coordinate DX (Mp) by
In step S14, it is detected that the division number matches (NX-1), and the process ends.

The outer coordinate series detection unit 8 targets the inside of the character frame, and scans the entire range of each of the four sides of the character frame, the top, bottom, left, and right sides, in a direction perpendicular to the sides, and detects the first black point. Detect the coordinates of the outer coordinate series GT(X), GB(X), at
, (y), and GR(y) are created. However, x and y are two-dimensional coordinates of θ to 127 in the pattern register 3, GT stores the X coordinate obtained by the above processing for the upper side, and GB stores the X coordinate obtained by the above processing for the lower side.
GL stores the X coordinate obtained by the above processing for the left side, and GR stores the X coordinate obtained by the above processing for the right side.

FIG. 5 shows the outer coordinate series from the top side for the human-powered character pattern for the kanji "water," and FIG. 8 shows the values of the outer coordinate series from the top side for the human-powered character pattern for the kanji "ice."

The polar point coordinate sequence detection section 9 detects the range X, ~X, , Yt-Y, in the outer coordinate series detected by the outer coordinate series detection section 8.
, and the increase/decrease to the outer coordinate series of each husband is (6
)2~(9), ignore the points where the value of the formula is zero, and find the series of coordinates of the poles that change in increase or decrease (TX(r++), TY(n+)), (BX(n2)
), BY(n2) ).

(LX (na), l, Y(n3)), (R
Detect X(nJjlY(n4)).

dGT(x)-GT(xtI)-GT(x) XI≦
x< x, ...(6)dGl)(x)-GII
(xt1)-GII(x) XI≦ x<xr...
(7) dcL(y)=GL(y+1)−GL(y)
yt≦V<yb...(8) dGR(y)=C
H(y+1)-GR(V) yt≦y<yb
”・(9) However, starting point (xt, GT(xt)), (x
t, GB(xt)), (GL(yt).

yt), (Gll(yt), yJ & end point (xr,
G'r (xr) ) , (Xr, GB(xr)),
(GL(yb), yb), (Gll(yb), yb)
is unconditionally the extreme point. Here n + + 2 +
n,, l n4 is the coordinate number of the pole for each outer coordinate series, 1≦n1≦Tn, 1≦02
≦Iln, 1≦n3≦Ln, l≦n4≦Rn
, Tn is the total number of poles in the outer coordinate series GT for the upper side, and Bn is the outer coordinate series G for the lower side.
Ln is the total number of poles in the contour coordinate series GL for the left side, and Rn is the total number of poles in the contour coordinate series GR for the right side. Also (TX(n+).

vv(n+)) is the pole coordinate series in the outer coordinate series GT for the upper side, (ax(n2), BY(n2
)) is the polar point coordinate series in the outer coordinate series GB for the lower side, (Lx (n3), LY (n3)) is the polar point coordinate series in the outer coordinate series GL for the left side, (RX (n4), RY (n4) )) is a pole coordinate series in the outer coordinate series GR for the right side.

FIG. 6 shows an example of the polar point coordinate series from the top side for the human-powered character pattern for the kanji character "water," and FIG. 9 shows an example of the polar point coordinate series from the top side for the human-powered character pattern for the kanji character "ice."

The polar point counting unit 10 calculates the target polar point coordinate series (PTX(pn+
), PTY(pn+)), (PBX(pn2),
PBY(pn2)), (PLX(pna)),
FLY (pn3)), (PRX (pn4)
. Referring to the area end dividing point number from the pole counting area table 11, the dividing point for the starting point dividing point number S of the reference side and the ending dividing point of the first reference side are determined for the polar point coordinate series determined by the type SE. Targeting the area surrounded by the division point for number E, the reference side, and the division point for the area end division point number, maximize the local maximum and minimum points, respectively, when the direction toward the reference side is positive. The number of points n5aax(j) and the number of minimum points n, . Count as (i).

However, in this example, the type SE is 5E=1 indicating a polar point coordinate series with the upper side as the reference edge, 5E=2 indicating a polar point coordinate series with the lower edge as the reference edge, and 5E=3 indicating the left side as the reference edge. We have decided to show the polar point coordinate series where 5E=4 and the right side is the reference side. In addition, i represents the number of each set when type SE, starting point dividing point number S, ending point dividing point number E - area end dividing point number are set as one set in the pole counting area table 11, and l≦i≦ru. ax, and i□8 is the total number of pairs in the pole counting area table 11. S, E, D are zero and indicate xl or YL,
NX+1 or NY+1 indicates xr or Yb.

Figure 7 shows the maximum and minimum points from the upper side of the kanji character "water" corresponding to Figure 6, and the maximum point counting area for S, E, and D, and the first
Figure 0 shows the maximum and minimum points S and E from the upper side of the kanji character "ice" corresponding to Figure 9.

The pole counting regions for D are shown respectively.

The identification unit 12 stores the number of local maximum points and the number of local minimum points for each target area counted by the local point counting unit 10 in a dictionary memory 13.
It is determined whether the standard pattern exists within the maximum point number range and the minimum point number range for each area of the standard pattern prepared in advance stored in , and if it exists within the range in all judgments. The character name of the determined standard pattern is output to the character name output 14.

Here, the effectiveness of the feature extraction method of the present invention regarding the number of maximum points and the number of minimum points, which are one type of feature information, will be shown below.

For example, when detecting the presence or absence of the upper left point of the kanji ``water'' and the kanji ``ice'' from the upper side, fluctuations in the strokes of the two trees on the right adversely affect the feature value. Therefore, by limiting the horizontal and directional right side range of the pole counting region by dividing points based on the peripheral distribution, the above-mentioned adverse effects can be removed. In addition, by limiting the polar point counting area in the vertical direction as well, it is possible to remove artifacts such as cuts and blurs outside the polar point counting area, and more stable feature values can be obtained.

(Effects of the Invention) As described above in detail, according to the present invention, character/figure patterns can be easily processed without performing very complicated pattern processing such as contour tracing or thinning in feature information extraction in conventional recognition methods. A character frame division point coordinate series for dividing the inside of a character rectangular frame into (MXN) areas is extracted from the center of gravity coordinates of the projection onto the two axes, and a surface By extracting the information of the extreme points that satisfy a predetermined condition for each region and using this as a feature, it is possible to provide a feature extraction method that is stable and effectively extracts differences between patterns where only minute differences exist. .

[Brief explanation of the drawing]

Fig. 1 is a functional block diagram showing an embodiment of a character recognition device to which the present invention is applied, Fig. 2 is a diagram showing the relationship between an example of a human character pattern, a barycenter coordinate series, and a divided coordinate series, Fig. 3 is a diagram showing the correspondence between the barycentric coordinate series and the divided coordinate series, Figure 4 is a flowchart of another method for determining the divided coordinate series, and Figure 5 is the outer coordinate series from the top side for the kanji character "water" in the human character pattern. An explanatory diagram of the detection method, Figure 6 is the 5th
An explanatory diagram of the maximum point and minimum point detection method for the outer coordinate series shown in the figure, FIG.
An explanatory diagram of the maximum point/minimum point counting method in FIG. 6 for 3,
Fig. 8 is an explanatory diagram of a method for detecting an outer coordinate series from the upper side for the kanji character "ice" in the human character pattern, Fig. 9 is an explanatory diagram of a method for detecting maximum and minimum points for the outer coordinate series of Fig. 8, and Fig. 10 The figure shows the pole counting area S=1. E=6. FIG. 9 is an explanatory diagram of the maximum point/minimum point counting method shown in FIG. 9 for D=3; 1... Optical input 2.-Photoelectric conversion unit 3...Pattern register 4.-Character frame detection unit 5...Character projection creation unit 6.-Gravity center detection unit 7...Character frame division point determination unit 8 . . . External coordinate series detection section 9 . . . Pole coordinate series detection section 10 . . . Pole counting section 11 - .
Character name output patent applicant Oki Electric Industry Co., Ltd. Patent application agent Patent attorney Megumi Yamamoto ×

Claims (1)

[Claims] In a feature extraction method that includes a pattern register that stores a pattern obtained by reading and quantizing a character figure on a medium as an original pattern, and extracts the feature of the character figure based on the original pattern, a) means for detecting a character rectangular frame consisting of the circumscribed sides of the original pattern stored in a pattern register; and (b) means for projecting the original pattern onto two desired axes to determine the respective black bit number distributions. (c) first on each axis;
Determine each barycenter coordinate of the black bit number distribution in the range limited by the character rectangle frame, and then divide the range limited by the character rectangle frame using each barycenter coordinate detected so far. (d) means for determining barycenter coordinates of each of the black bit number distributions as objects by repeating the process a plurality of times to obtain each barycenter coordinate series; The coordinate numbers and the divided coordinate numbers in the divided coordinate series on each axis, which are assigned in the order of the size of the coordinate values, correspond almost equally, and one barycenter coordinate in the barycenter coordinate series corresponds to one divided coordinate. (e) means for determining division coordinates for dividing the inside of the character rectangular frame into (M×N) areas by dividing the inside of the character rectangular frame into (M×N) areas; (f) means for calculating the amount of change at each change point based on the change point coordinate series; (g) means for detecting the extreme points that increase or decrease according to the amount of change; and (h) extracting information on the extreme points that satisfy a predetermined condition for each of the (M×N) regions. A feature extraction method characterized by comprising means for.