JPH06131496A - Pattern normalization processing method - Google Patents

Abstract

PURPOSE:To improve the recognition rate of a pattern such as a character, a symbol, and a figure. CONSTITUTION:A Japanese character shown in figures features three lateral lines and two longitudinal line and a Japanese character equivalent to an English word 'tree' features one lateral line, one longitudinal line, one left downward slanting line, and one right downward slanting line. Thus, the length and directivity of each of segments constituting a character are taken into consideration. From this point of view, when the character pattern of an originally identical character is divided, parts (segments or parts of segments of the character) of a character pattern corresponding to the same parts of the characters are put in respective divided areas.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern normalization processing method among the elemental technologies of pattern recognition technology.

In the field of pattern recognition, even if the characters, symbols, and figures are originally the same, patterns that have changed in various ways due to individual circumstances such as individual differences appear, which causes difficulty in recognition. Has become. Therefore, a process for aligning various patterns as uniformly as possible, for example, a process for aligning character sizes is performed.

In particular, in the recognition of handwritten characters and the like, a process of absorbing the fluctuation of the handwriting and bringing it closer to a standard character having no fluctuation, that is, "normalization" is performed. Character normalization is a process of converting a character having a large variation into a character having a variation eliminated, and what kind of normalization has a great influence on the recognition rate thereafter.

[0004]

2. Description of the Related Art Here, a recognition technique for handwritten characters and a normalization technique directly related to the present invention among the recognition techniques will be described. The same technology is applied to symbols and figures other than letters. In the following, it is assumed that the character pattern is represented by binary (1/0) pixel data.

FIG. 6 shows an example of a handwritten character pattern of the character "day". As shown in FIG. 6, a well-balanced character pattern (A), a character in which the central horizontal line of the "day" is shifted downward (B), a rounded character (C), a small character ( Even if they have the same letter "day" such as D), etc.,
Also, the same person sometimes forms different character patterns.

As a device for recognizing such handwritten or printed patterns of various kinds, there is a recognition device having a structure as shown in FIG. FIG. 2 is a block diagram of a recognition device. The recognition device includes an input data reading unit 1 that optically reads characters and the like written on recording paper, a normalization unit 2 that normalizes the read input data, A feature extraction unit 3 that extracts the features of the normalized data, a collation unit 4 that compares and collates the feature-extracted input data with the dictionary 6, a collation result output unit 5 that outputs the comparison and collation result, and an overall control unit. 7 and 7.

Further, the dictionary 6 collects many samples for each character and the like, and stores the average of the characteristics of the samples. Since the present invention relates to the normalization unit 2, a conventional technique of the normalization unit 2 will be described here.

FIG. 7 is a block diagram of a conventional normalization unit, FIG. 8 is a diagram in which the sizes of handwritten character patterns are made uniform, and FIG. 9 is a labeling of the character pattern "day" (A). FIG. 10 shows the result of labeling the character pattern “day” (B). Further, FIG. 11 shows movement of the center of gravity and partial enlargement / partial reduction.

The conventional normalization technique is as shown in FIG.
The character size processing unit 20a, the direction label assigning unit 21a, the center of gravity calculating unit 22a, the pattern dividing unit 23a, the enlarging / reducing unit 24a, and the control unit 25a.

The function of each processing unit will be described below. (1). The character size processing unit 20a. Align the sizes of handwritten and printed patterns.

As shown in FIG. 6, the sizes of the input characters, symbols, etc. are made uniform for each type. The method of arranging the sizes of the characters and symbols is as follows. The circumscribing rectangle 10 shown in FIG. 11 is calculated for the handwritten character, the circumscribing rectangle is converted into a rectangle 11 having a predetermined size, and the conversion coefficient at that time is calculated.

The vertical of the circumscribing rectangle 10 is u, and the horizontal is v
And -The length of the rectangle 11 of a predetermined size is U, and the width is V. The vertical conversion coefficient a is a = U / u, and the horizontal conversion coefficient b is
b = V / v.

The point (x, y) within the original circumscribing rectangle 10
If the value of M is a binary value of 0/1 or a multivalue representing grayscale, the value of the pixel at the point (X, Y) in the rectangle 11 of the determined size is M. (The original value M is mapped to the new coordinate system.) Where X = ax Y = by.

(2). Direction label providing unit 21a. Focusing on a certain black pixel, character pixels above and below and to the left and right of the black pixel are examined. If there is a certain number of black pixels in the horizontal direction of the focused black pixel, and there are black pixels in the vertical direction of the focused black pixel, the number of black pixels in the horizontal direction. Is greater than the number of black pixels in the vertical direction and the difference is a certain value or more, the directionality of the focused black pixel is the horizontal (H) direction, and a mark in the H direction is attached.

Similarly, if there is a certain number of black pixels in the vertical direction of the focused black pixel and there are also black pixels in the horizontal direction of the focused black pixel, When the number of black pixels in the direction is larger than the number of black pixels in the left-right direction and the difference is more than a certain size, the directionality of the focused black pixel is the vertical (V) direction, and the V-direction mark is wear.

If black pixels are present in either the up, down, left, or right directions and the difference between them is less than a certain size, the directionality of the focused black pixel is neutral, and nothing is marked. Not attached.

FIGS. 9 and 10 show the black pixels marked with the above rule. FIG. 10 shows the labeling result of the character pattern “day” (A), and FIG. 10 shows the labeling result of the character pattern “day” (B). In FIG. 9 and FIG. 10, the portions marked with ● are neither in the H direction nor in the V direction.

(3). Center of gravity calculator 22a. The center of gravity C10 of the handwritten character pattern in FIG. 11 is obtained as follows.・ The position of the center of gravity C10 on the vertical axis is H
= 1 to K (K = 12), the position of a black pixel having a directionality (H directionality) perpendicular to each V axis with respect to the V axis of each H value is obtained, and from that position, each V axis is obtained. The upper center of gravity is found.

For example, the character pattern "day" in FIG.
In the labeling result of (B), black pixels having H directionality exist at V = 3, 10, and 13 with respect to the six V axes of H = 4 to 9.

Since the positions of the respective centroids on the six V-axes are V = 8.7, the average value is also V = 8.7. Further, the center of gravity on the H axis is V = 4 to 12 (where V = 1
Excluding 0. ), Since there are black pixels having V directionality at H = 3 and 10 with respect to each H axis of FIG.
= 6.5, the average value of the center of gravity also becomes H = 6.5.

Therefore, the character pattern "day" in FIG.
The center of gravity C10 (H, V) of (B) is the point (6.5, 8.7).
Becomes Similarly, the center of gravity C10 (H, V) of the character pattern “day” (A) in FIG. 9 is the point (6.5, 8).

(4). The pattern division unit 23a. With the center of gravity C10 thus obtained as the center, a line parallel to the coordinate axes is divided into four as shown in FIG.
,, and.

(5). Enlargement / reduction unit 24a. The coordinates of each point are converted so that the obtained center of gravity C10 becomes the center C11 (U / 2, V / 2) of the rectangle 11 having the determined size, and each region is enlarged or reduced.

The area of () is enlarged in the horizontal direction and reduced in the vertical direction. The area of ・ is reduced both vertically and horizontally.・ The area of is expanded vertically and horizontally.

The area in the vertical direction is enlarged in the vertical direction and reduced in the horizontal direction. Becomes The conversion formula in each area is as described in (1). The same method as the method described in the character size processing unit 20a is used.

This method is a normalization method in which a dense portion of a character is enlarged and a sparse portion is reduced so that the density of black pixels of the character is made as uniform as possible.

[0027]

As described above, as shown in FIG.
In the various patterns of the character "day", the center of gravity of the pattern (A) is a point (H = 6.5, V = 8), and the center of gravity of the pattern (B) is a point (H = 6.5, V). = 8.7).

When the patterns (A) and (B) are divided into four parts, that is, the upper and lower halves and the left and right halves at the center of gravity described above, in the case of the pattern (A), the line of the upper and lower halves is just "day". It matches the horizontal line in the middle of the character and becomes vertically symmetrical.

On the other hand, in the case of the pattern (B), the position V of the horizontal line at the center of the character "day" of this character pattern is V = 10, and the position V of the upper and lower halves is V = 8. .7
Therefore, the line that divides the upper and lower halves is above the horizontal line in the center of the character "day", and is vertically asymmetric.

As described above, even if the same character "day" is used, the number of pixels included in each divided area and the position of the character or the character forming the character are different depending on the pattern in the centroid dividing method as in the prior art. The number of lines and their location differ.

For this reason, there is a problem in that when the features in each divided area are compared in the feature extraction process of recognition and the comparison and collation process with the dictionary, a large difference occurs. The present invention is
It is an object to provide a means for improving the pattern recognition rate.

[0032]

The above problem can be solved by a pattern normalization processing method configured as follows.

FIG. 1 is a diagram for explaining the principle of the present invention. 1. The pattern written on the recording paper is read, the digitized numerical value is recorded on the storage device, the read pattern is scanned, the pattern is normalized based on the scan result, and the normalized pattern is recorded. Extract features,
Regarding the normalization of the pattern recognition device that compares and collates with the dictionary in which the standard character pattern is stored, the reading pattern is scanned in the two-dimensional coordinate system, and the length in the scanning direction of the pattern is set to K and digitized. The directivity of each black pixel which is an element of the pattern is checked, and the directivity is given to each black pixel. For each value on each coordinate axis of the two-dimensional coordinate system, When the number of black pixels having the directionality of an angle is C (i) and the total sum of C (i) is N, the pattern is divided into L in the direction perpendicular to the scanning line, and N / L = W is set. , The j-th division position,
Partial sum S up to value d on the coordinate axis of C (i)

[0034]

[Equation 2]

The smallest d exceeding W × j is set as a division position, and the area is enlarged / reduced in the scanning direction so that the length in each scanning direction of each divided area becomes K / L. . 2. In the above 1, as the scanning direction, horizontal and vertical axes,
It is configured to have a right-upward and right-downward axis or two orthogonal axes that are inclined at an arbitrary angle.

3. In the above-mentioned item 1, all directions are adopted as the directionality of each black pixel. 4. In the above 1, when the number C (i) of black pixels having a certain angle of directionality is obtained, the number of black pixels having a direction perpendicular to the scanning direction is adopted.

5. In the above 1, when obtaining the number C (i) of black pixels having a certain directionality, the number of black pixels having all directions other than the same direction as the scanning direction is adopted. .

[0038]

According to the conventional technique, one characteristic face of a character pattern is condensed into one point called the center of gravity, and then enlargement / reduction is performed centering on the center of gravity. Therefore, one characteristic face is calculated in the process of calculating the center of gravity of the character pattern. It disappears and recognition fails.

In the present invention, three horizontal lines and two vertical lines are used for the character "day", and one horizontal line, one vertical line and one diagonal line "lower left" for the character "tree". It focuses on the length of the line segment that makes up the character and the directionality of the line segment, such as "No" and one diagonal line that descends to the lower right.

According to the present invention, if the characters are originally the same from this point of view, when the character pattern of the character is divided, the part of the character pattern corresponding to the same part of the character ( The present invention provides a division method (normalization method) in which a line segment of a character or a part of the line segment is included.

[0041]

EXAMPLE The normalization method of the present invention will be described below by taking the character "day" as an example. In FIG. 1, both (A) and (B) are binary character patterns in which the same character "day" is read by the input data reading unit 1 and binarized.

Two orthogonal axes are a horizontal H axis and a vertical V axis. The maximum value of the character frame is K = 12 for the horizontal H axis and K = 15 for the vertical V axis. Figure 3
[Fig. 3] is a block diagram of a normalization unit, which is an enlargement of the normalization unit 2 in Fig. 2 in more detail. This normalization unit 2
A character size processing unit 20, a direction label assigning unit 21, a label totaling unit 22, a division position determining unit 23, an enlarging / reducing unit 24,
It is composed of a control unit 25.

FIG. 4 is a flowchart (1) for determining the division position, and FIG. 5 is a flowchart (2) for determining the division position. The character size processing unit 20 and the direction label assigning unit 21 in FIG.

The label totalizing unit 22, division position determining unit 23, enlarging / reducing unit 24, and control unit 25 shown in FIG. 3 will be described below. (1). Label totaling unit 22.

The character pattern is scanned in the vertical direction at each value (H = 1 to 12) on the horizontal axis, and the black pixels of the character pattern whose scanning lines have the horizontal direction are respectively scanned on the vertical axis. The values V = 1 to 15 are added to obtain an added value C (i).

In the example of FIG. 1A, V = 3, V = 8,
The added value C (i) is "6" for each of the three locations of V = 13. The total sum N of the added values C (i) is N = 18. In the example of FIG. 1B, the added value C (i) is “6” only at three locations of V = 3, V = 10, and V = 13. Addition value C
The total sum N of (i) is N = 18.

Next, the character pattern is scanned in the horizontal direction at each value (V = 1 to 15) on the vertical axis, and the black pixel of the character pattern whose scanning line has the vertical direction is displayed on the horizontal axis. Add value at each of the above values H = 1 to 12 and add value C
Find (i).

In the example of FIG. 1A, H = 3 and H = 10
The added value C (i) is "8" in each of the two positions. Therefore, the total sum N of the added values C (i) is N = 16. Figure 1
In the example (B), H = 3, H as in the example (A).
The added value C (i) is "8" in only two places of = 10. Therefore, the total sum N of the added values C (i) is N = 16.

(2). Division position determination unit 23. The following description is based on FIGS. 4 and 5. (2.1). Initial values are set in the vertical division processing. (Processing step 100) ・ The number of divisions L is set according to the knowledge (for example, L = 1
0).

N = 0, i = 1. (2.2). (1) The total sum N of all the rows of the added value C (i) of each row obtained by the label totaling unit 22 is obtained. Here, the maximum value of the row is K. That is, the maximum value K of a row is the number of scanning lines when electronically scanning a character, or the number of scanning lines when some of the scanning lines are grouped together.

N = N + addition value C (i). To add the addition value C (i) of the next row to N, i =
Let i + 1. Check if i is equal to the maximum value K of the row.

(Processing Step 101 to Processing Step 10
4) (2.3). The total sum N is divided by the number of divisions L to obtain W. W = N / L. (Processing step 105) (2.4). Initialization is performed to obtain the partial sum S (total sum is N) of the added value C (i). (Processing step 106) i = 1.

J = 1. Here, j represents the j-th divided area, that is, the j-th partial sum S. The maximum value of j is L.・ S = 0.

(2.5). Find the partial sum S. (Processing Step 107 to Processing Step 109, Processing Step 11
4 to processing step 115) S = S + addition value C (i).

In order to add the added value C (i) of the next row to S, i = i + 1. The minimum row d (the value of i at that time) for which the partial sum S satisfies S ≧ W × j is obtained, and the minimum row d is the division position.

(2.6). The last line of the character (i =
K) or the last divided area (j = L). (Processing step 110, processing step 116) (2.7). Initial values are set in the horizontal division processing. (Processing step 111) (2.8). Similar to the vertical direction, horizontal division processing is performed. (Processing Step 112 to Processing Step 113) For example, consider dividing into two in the horizontal direction (L = 2), that is, dividing into two vertically.

Addition value C (i) of black pixels having H directionality
Is added up from i = 1 to d to obtain a partial sum, and the partial sum up to d is 1/2 of the total N = 18 of C (i), which is 9 (W =
Good division can be achieved by dividing at position d beyond 9).

In the example of FIG. 1A, d = 8 is the upper / lower splitting position, and in the example of FIG. 1B, d = 10 is the upper / lower splitting position. By dividing in this manner, unlike the prior art, both of the examples of FIG. 1A and FIG. 1B are divided into upper and lower parts at the position of the horizontal line at the center of “day”.

The same can be said for values of L = 3 or more. For example, if L = 3, W = 18/3 = 6, and in the example of FIG. 1A, there are two positions of V where the partial sum exceeds 6: d (V) = 3,8. , (B), there are two locations, d (V) = 3,10.

(3). Enlargement / reduction unit 24. When the division position is obtained by the above (2), enlargement / contraction is performed using the conventional technique so that the divided areas have the same area.
Reduce and convert. The position of the horizontal line in the center of "day" of the character pattern after conversion is near the center.

(4). Control unit 25. The flow is controlled by the character size processing unit 20, the direction label assigning unit 21, the label totaling unit 22, the division position determining unit 23, and the enlarging / reducing unit 24.

After the normalization processing is performed as described above, the input data is sent to the feature extraction unit 3 by the control unit 7, feature extraction is performed, and the collation unit 4 compares and collates with the dictionary 6. The comparison result output unit 5 outputs the comparison result.

Further, as described above, since a character has diagonal character lines like "trees", the method of determining the scanning direction, the directionality of each black pixel of the character pattern, and the directionality are determined. Regarding how to obtain C (i) for each black pixel
There are various methods as follows.

(1). How to determine the scanning direction. The scanning direction is configured to have horizontal / vertical axes, upward / downward rightward axes, or two orthogonal axes inclined at an arbitrary angle. (Claim 2) (2). Direction of each black pixel in the character pattern.

As the directionality of the character pattern to be examined,
As the directionality of each black pixel, all directions are adopted. For example, at six points where three horizontal lines and two vertical lines of the character "Sun" intersect, the direction is "none" in the explanation of the prior art section (see mark ● in Fig. 9). It is assumed that the directionality of the location has both the “H direction” and the “V direction”, and the values are added when C (i) is obtained. (Claim 3) (3). How to find C (i).

There are the following two methods for obtaining C (i). (31) Adopt the number of black pixels having a direction perpendicular to the scanning direction. For example, when scanning is performed in the "V direction" with the character "tree", only the black pixels of the horizontal character line having the "H direction" are adopted, and the black line "No" in the lower left corner is not adopted. I will. Neither will it be used for the lower right and lower lines. When scanning in the "H direction", only the black pixels of the horizontal character line having the "V direction" are adopted, and the character line of "No" in the lower left direction is not adopted. Neither will it be used for the lower right and lower lines. (Claim 4) In this case, in this process, the characteristic of the character line of "NO" which descends to the lower left and the characteristic of the character line which descends to the lower right disappears in the normalization process.

To avoid this, the lower left "no"
In order to normalize the character line of (1) and the character line of lower right, the above (1). In the method of determining the scanning direction, the lower left lower direction and the lower right lower direction are adopted as the scanning direction, or the following method (32) is adopted.

(32) The number of black pixels having all directions except the same direction as the scanning direction is adopted. For example, when scanning in the "V direction" with the character "tree", the black pixels of the horizontal character line having the "H direction", the black pixels of the lower left "NO" character line, and the right The black pixel of the character line which descends and descends is adopted. When scanning in the "H direction", black pixels of horizontal character lines having the "V direction",
The black pixel of the character line of "No" which descends to the lower left and the black pixel of the character line which descends to the lower right are adopted. In this way, there is no possibility that the characteristic disappears as in the above (31) in the normalization process. (Claim 5)

[0069]

As is apparent from the above description, according to the present invention, characters, symbols, and figures are divided into a plurality of areas without losing their characteristics, and the inside of each divided area of the characters is enlarged. -There is an industrial effect that the recognition rate of handwritten or printed characters is improved by reducing the size and extracting the features.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a block diagram of a recognition device.

FIG. 3 is a block diagram of a normalization unit.

FIG. 4 is a flowchart (1) for determining a division position.

FIG. 5 is a flowchart (2) of determining a division position.

[Fig. 6] Handwritten character pattern example of the character "day"

FIG. 7 is a block diagram of a conventional normalization unit.

FIG. 8 is a diagram in which handwritten character patterns have the same size.

FIG. 9: Labeling result of character pattern “day” (A)

FIG. 10: Labeling result of character pattern “day” (B)

FIG. 11: Movement of center of gravity and partial enlargement / partial reduction

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Input data reading unit 2 Normalization unit 3 Feature extraction unit 4 Collation unit 5 Collation result output unit 6 Dictionary 7 Overall control unit 10 Handwritten character pattern circumscribing rectangle 11 Normalized character pattern character frame 20, 20a Character size processing unit 21 , 21a Direction label imparting unit 22 Label totaling unit 22a Centroid calculating unit 23 Dividing position determining unit 23a Pattern dividing unit 24, 24a Enlarging / reducing unit 25, 25a Control unit 100 to 116 Processing steps

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Haruo Akimoto 1015 Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Fujitsu Limited

Claims (5)

[Claims] 1. A pattern normalization for reading a pattern written on a recording paper, recording a digitized numerical value on a storage device, scanning the read pattern, and normalizing the pattern based on the scanning result. In the processing method, the reading pattern is scanned in a two-dimensional coordinate system, the length in the scanning direction of the pattern is set to K, and the directionality of each black pixel that is an element of the digitized pattern is checked and the direction is determined. Property is given to each black pixel, and for each value on each coordinate axis of the two-dimensional coordinate system, the number of black pixels having a directionality of a certain angle with respect to the coordinate axis is defined as C (i). When the sum of C (i) is N, the pattern is divided into L in the direction perpendicular to the scanning line, and N / L =
As W, the j-th division position is the partial sum S up to the value d on the coordinate axis of C (i). Is the minimum d exceeding W × j, and the pattern normalization is performed by enlarging / reducing in the scanning direction such that the length of each area after the division in each scanning direction is K / L. Processing method. 2. The pattern normalization processing method according to claim 1, wherein the scanning direction has a horizontal / vertical axis, a rightward / rightward downward axis, or two orthogonal axes inclined at an arbitrary angle. . 3. The pattern normalization processing method according to claim 1, wherein all directions are adopted as the directionality of each black pixel. 4. The number of black pixels having a direction perpendicular to the scanning direction is adopted when determining the number C (i) of black pixels having a certain angle directionality. Pattern normalization processing method. 5. The number of black pixels having all directions other than the same direction as the scanning direction when determining the number C (i) of black pixels having a directionality of a certain angle according to claim 1. A pattern normalization processing method characterized by being adopted.