JPH033089A - Method for recognizing pattern - Google Patents

Abstract

PURPOSE:To automatically form a dictionary by forming the appearance frequency distribution of pattern feature points by preparing many samples, normalizing respective frequency values of the distribution and registering the normalized values as membership values. CONSTITUTION:Many samples for a certain pattern inputted from a pattern input part 2 are prepared and the pattern feature points of respective samples are extracted. A mesh area is allocated to a plane area on which the pattern is to be plotted and a mesh area is set up also in a dictionary storage part 1 corresponding to the allocated mesh area. Then, the appearance frequency which is the number of generations of each feature point is found out to form the appearance frequency distribution corresponding to the mesh area, the distribution is normalized, the normalized value is used as a membership value, a distance corresponding to the degree of separation between the pattern to be recognized and the pattern to be collated is calculated by a distance computing part 5, and the proper number of patterns is selected in the ascending order to obtain a recognized result. Consequently, a recognition method with high recognition accuracy, a dictionary constituting method for reducing storage capacity and an automatic dictionary forming method can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention relates to a method for recognizing patterns such as characters and figures.

B6 Summary of the Invention The present invention is a method for recognizing patterns such as characters and figures using a dictionary to be checked, which uses a dictionary in which positional features of feature points of a pattern are expressed as fuzzy sets (Fanoy sets). , which performs calculations on the membership values corresponding to the positions of the feature points of the pattern to be recognized, and obtains recognition results based on the calculation results, making it possible to recognize patterns with high accuracy and at high speed. be.

C9 Conventional technology When recognizing patterns such as characters and figures, for example, based on black and white binary image data (input pattern) obtained by operating documents, drawings, etc. with an input device such as an image scanner, Standard patterns of characters and figures stored in advance (
Dictionary patterns) and positions are compared, and the one that most overlaps among the dictionary patterns is finally determined as the recognition result.

As a specific example of the recognition method, for example, the human cover pattern and the dictionary pattern are each divided into small square grids (mesh), and if the grid points are letters or figures, they are marked black (1), otherwise they are marked (0). In the case of expression, the degree of overlap between the two patterns of human input and dictionary can be determined using the Hamming distance. This is done by setting 0 if the human power and dictionary are black and black or white and white, and setting it as l if they are black and white or white and black, and finding the sum. In other words, the more they overlap, such as black on black or white on white, the smaller their sum (Hamming distance) becomes. Due to this property, from among a large number of dictionary patterns, the one that most overlaps with the input pattern, that is, the one with the smallest Hamming distance, is selected as the recognition result.

D1 Problem to be solved by the invention With this method, if the input pattern is misaligned with the dictionary pattern, has a different size, or has a different inclination, the degree of overlap will change, making recognition difficult. It has the disadvantage of being prone to errors.

There is also a problem in terms of storage capacity. The space required for one character or one graphic pattern is not limited to 8 x 8 pixels (alphanumeric characters) or 24 x 24 pixels (kanji) like character fonts for display, but at least that space is required to avoid line breaks and collapse as much as possible. Particularly in the case of Chinese characters or complex figures, which are characterized by a space 3 to 5 times larger, stable processing cannot be expected unless at least 128 x 128 pixels are secured.

In the case of kanji, there are more than 6,000 types of dictionaries for characters and graphic patterns, including the second level, and in the case of figures, there are often more than 1,000 types.

For such a huge number of dictionary types, 128
If x12g pixels are secured, 2 bytes are required as memory. In other words, the above-mentioned kanji require a storage area of 12M bytes or more, and graphics require a storage area of 2M bytes or more.

This capacity is too large to be ignored due to the hardware configuration, and it also affects dictionary search and matching processing in terms of processing time.

Furthermore, in terms of dictionary types, it is necessary to manually register each dictionary for the huge number of types mentioned above, which poses the problem of requiring an enormous amount of time.

In summary, the following three major issues remain in the prior art.

■ Establishment of a recognition method with high recognition accuracy ■ Establishment of dictionary construction method for reducing storage capacity ■ Establishment of automatic dictionary generation method The present invention aims to solve these problems.

E9 Means to solve the problem Letters and graphic patterns have positional relationships that are described according to human historical or empirical rules, but the points that characterize the patterns (feature points) themselves have ambiguity.

For example, in the case of a typical feature point of the character "Ao" indicated by a circle in λ in Figure 2 (this will be referred to as a vertex), the position of the vertex itself may be shifted to the left (same as Diagram b
) or swing to the right (C in the same figure). It may also break at the apex (d in the same figure).

However, it is impossible for only the vertex portion to be located below the other feature points, as shown in e of the figure.

In the end, the apex of the letter "A" has regularity on an extremely ambiguous scale, in that it is located above other feature points in terms of positional relationship and near the middle.

The present invention takes into consideration that there is ambiguity in the positional relationships of feature points in character and graphic patterns, and expresses the positional relationships by incorporating the concept of fuzzy sets to express the positional relationships for each pattern. A two-dimensional fuzzy dictionary is registered, a pattern to be recognized is compared with the two-dimensional fuzzy dictionary of each pattern, and a recognition result is obtained based on the comparison result.

Specifically, the present invention creates a large number of samples for a certain pattern, extracts feature points of the pattern in each sample, and places a×b (a.

b is an integer), a mesh area is set in the dictionary storage unit correspondingly, and the appearance frequency, which is the number of times the feature point appears, is determined for each mesh in this mesh area, and the mesh area is Create an appearance frequency distribution corresponding to the area, normalize each frequency on the appearance frequency distribution using a reference frequency, use this normalized value as a membership value, and store the dictionary based on this membership value. Create a two-dimensional fuzzy dictionary by assigning membership values to each mesh in the mesh area within the department,
The two-dimensional fuzzy dictionary obtained in this way is registered in advance in the dictionary storage unit for each pattern along with the reference number j of feature points, the feature points are extracted for the pattern to be recognized, and the total number m of feature points and each feature point are extracted. Find the mesh position (xl, y,) (1≦i≦m) in the mesh area of
The membership value r k(x r
, y 1), and calculate each membership value f k (x r, y 1) using the certainty factor rkcx,, yt) ) and the standard number of feature points.
and the total number m of feature points, calculate a distance Dk corresponding to the degree of separation between the pattern to be recognized and the pattern to be matched PTk, and calculate this distance Dk for all patterns in the dictionary storage unit. It is characterized by calculating and selecting an appropriate number of patterns in order from the smallest value to the recognition result.

Here, the meaning of the distance Dk will be explained.

The total number of feature points of the pattern (human cover pattern) currently being recognized is m, and the mesh position of each feature point is (xl, yI).
, Cxt, yt)... (xa, ya). First, with a certain pattern PTk registered in the dictionary storage unit as a matching target, the membership value fh(xt,
yI). This fh(xl, yi) is (xt
, yI) is the pattern P
This indicates the degree of certainty that it is one feature point of Tk. Therefore, in the present invention, for example, the certainty factor! , 0 to fh
(xi.

yl) divided by the total number of feature points m of the human cover turn or the reference number j of feature points related to the pattern PT, the value obtained by subtracting the value is the difference between the feature points located at (x l, y t) and the pattern PT. The concept of distance is used, and the total value of the distances for all feature points of a human cover pattern is taken as the distance Dk between the pattern and the pattern PT to be matched. An example of the calculation formula for Dk is as follows: Dh=Σ(1, Ofk(xt, y□))/m1=1 when j and m are equal. Also, when " and [n are 5'4, j>m, 1) h = (Σ(1, Ofk(xl, yI)) + (
j m))/jrn>j, 1)k-Σ(1,0-fk(xl,yI))/m. These equations can be summarized as follows.

Dh-(wax(a,j)-Σ fh(X+,ya)
/max (a+, j) - (1) i - where I lax (m, D are the larger values of m, j).

F. Example The specific procedure of the method of the present invention will be explained below.

(1) Dictionary creation unit: To create a dictionary, prepare multiple samples for the same character or map.

These become dictionary creation patterns when creating a certain dictionary.

(2) Creation of feature point appearance frequency distribution (two-dimensional histogram) Samples are repeatedly manually extracted to extract feature points from each sample. On the other hand, a×b pieces (a
, b is an integer), and corresponding mesh areas are set in the dictionary storage unit. Then, the appearance frequency, which is the number of times the feature point appears, is determined for each mesh in this mesh area, and an appearance frequency distribution corresponding to the mesh area is created as shown in FIG.

For example, when a mesh area of 10x10 is allocated to the plane area where the pattern is drawn, the mesh area of 10x10 is also allocated in the dictionary storage unit.
Set the mesh area of lO.

At this time, the position of a certain feature point is z (x, y) - (5°3
), the position Z(5,
It is the frequency of 3) plus 1.

If the feature point position is z (x, y), the frequency P(x, y) at the corresponding position Z (x, y) on the two-dimensional histogram is
, y) are expressed by the following equation.

P (x, y) −P (x, y) + 1(P(x, y
) initial value 0) (3) The feature point appearance frequency distribution obtained by defining the membership value in a two-dimensional fuzzy set and creating a dictionary shows the tendency of where feature points tend to appear. There is. In other words,
When the frequency distribution of multiple samples is taken for the same character or figure, feature points tend to appear in similar positions, indicating that a peak often exists near these points.

From this, conversely, the feature point position can be estimated by extracting the peak position.

The present invention applies this idea to setting membership values in fuzzy sets.

Now, in the frequency distribution, the larger the number of samples for one pattern, the higher the overall frequency becomes, making it difficult to make objective judgments, so it is necessary to normalize the frequency distribution.

On the other hand, the membership value in a fuzzy set is an ambiguous scale expressed as a subjective quantity in the real number domain of the [0, I] interval.

The definition of the membership value is based on the real number area in the [0, I] interval and the fact that feature points often correspond to peaks of the frequency distribution.

In other words, the peak value of the frequency distribution can be an indicator of an ambiguous measure of feature point position, so if it is normalized in the real number interval from 0 to 1, it can be used as it is as a membership value for a two-dimensional fuzzy set. .

Therefore, the frequency of appearance of feature points is normalized and registered in the dictionary storage unit as follows.

When the number of feature points of all samples is K, the number of peaks up to the reference point is obtained by multiplying by a certain - constant ratio Ck.

k = Ch.K Next, in the feature point appearance frequency distribution, search is performed in descending order of frequency, and the feature point having the k-th frequency value is determined, and this is set as the frequency P of the reference point.

Based on Pk, the feature point frequency distribution is normalized using the following equation.

If the frequency on the distribution is P1 and the frequency after normalization is M, then M=1.0...When P≧P, P/P
When P<Pk, normalized results can be obtained by performing this on all frequencies on the distribution. This result is in the range 0 to 1 and is registered in the dictionary storage unit as a two-dimensional fuzzy dictionary.

This completes the dictionary creation procedure for one pattern.

To illustrate an example of the two-dimensional fuzzy dictionary obtained in this way, Fig. 3 shows the feature point frequency distribution of the character rAJ, and Fig. 4 shows a normalized version of the distribution shown in Fig. 3. In this example, a 9×9 mesh area is set in a certain area. Here, the feature point corresponds to a contour pixel obtained by extracting, for example, a white pixel adjacent to a black pixel as a contour pixel, and compressing information by performing linear approximation processing or the like on these pixels.

Next, an example of a method of inputting a pattern to be recognized and comparing it with a group of patterns in the dictionary storage section will be described. In the dictionary storage unit l in FIG. 5, a fuzzy dictionary and a reference number j of feature points are registered for each pattern. Base Q of feature points
, gI number J is, for example, the average number of feature points of the pattern samples used when creating the fuzzy dictionary.

First, a pattern is created manually in the pattern manual section 2, and the total number of feature points (total number of feature points) m of the pattern is determined by the feature point number adding section 3. Next, the membership value adding unit 4 calculates the main positions (x,, yI) (1516
membership value f − (x i, y
1) Find 'ft, add all of these fm(x+, Yi), and do that. Then, the maximum value detection unit 6 determines jlm
The larger one of these 5ax(j, s) is output, and the distance #1 calculation unit 5 executes the calculation of equation (1) described above to obtain the distance #D,
seek.

This trial is performed for all the registered patterns, and the short distance determining section 7 picks out two patterns in descending order of Dk, which are output as recognition results from the recognition result output section 8.

In the above, as a two-dimensional fuzzy dictionary, the array of membership values shown in Figure 4 obtained by normalizing the appearance frequency distribution shown in Figure 3 may be used as is, but the following processing may be applied. You can use the one obtained by

First, the mesh that has the maximum value among the appearance frequencies of each mesh is used as the search start point for the first area, and meshes included in the first area are searched according to predetermined rules. Using the mesh with the maximum appearance frequency as the starting point for searching the second region, the meshes included in the second region are similarly searched, and by repeating this process, the mesh region can be divided into one or more regions. Separate into On the other hand, in order to normalize the frequency of appearance, a reference frequency is determined based on, for example, the total number of frequencies of appearance.

Next, for each separated area, membership values are determined by normalizing each appearance frequency for the mesh area that includes only that separated area using the reference frequency, and then for each row of each mesh in the mesh area, the membership value is determined. At the same time, the maximum value is determined from among the membership values included in each column, and these maximum values are grouped for each separation area as shown in Figure 6. What! Create a dimensional fuzzy dictionary.

Next, set the X coordinate xl to the i-th mesh column from the end of the mesh area, and set the Y coordinate y to the j-th row from the end.

are assigned respectively, and the membership values corresponding to The smaller value of X(i) and Y(j) is assigned to the two-dimensional membership value r(i, j) at the mesh position of x i, y J), and the two-dimensional membership value for the separation region is Obtain the set of membership values for .

Then, for the membership value F(i, j) at the mesh position (x i, y , 1), f of each separation region
(i, j), and assign this F(i, j
), a two-dimensional fuzzy dictionary for all regions of each separated region is obtained as shown in FIG.

G. Effects of the Invention According to the present invention, the appearance frequency distribution of feature points of patterns such as characters and figures is created by preparing a large number of samples, each frequency of this appearance frequency distribution is normalized, and its value is calculated. Since the information is registered together with the membership value, it is possible to automatically create a dictionary, and the manual work required for dictionary registration is greatly reduced. Errors in recognition are also less likely to occur.

Furthermore, as a dictionary space, one of the character fonts for display
．． A mesh space of about 2 to 1.5 times is sufficient for recognition. For example, a mesh space of about 9 x 9 is sufficient for alphanumeric characters, and about 32 x 32 for complex kanji. For example, set the membership value from 0 to l to 0. ! When expressed as a value carved in , the data can be expressed as an integer from θ to IO, so it can be composed of 4 bits.

Therefore, the amount of memory is 32×32× per i kanji character.
It is 4 bits - 512 bytes, which requires about I/4 of the conventionally required 2 bytes, and it is possible to reduce the storage capacity.

In addition, when recognizing an incoming Kabataan, the confidence can be obtained by simply finding the membership value corresponding to the mesh position of the feature point of the Kabataan from a dictionary, and the value obtained by subtracting that confidence from the confidence of 1.0 is the feature This distance is taken as the distance between each point to the pattern to be matched, and this distance is used to calculate the distance Dk between the two patterns using equation (1), and the smaller the value, the higher the degree of pattern matching is recognized. Therefore, calculations are simple and high-speed pattern recognition can be realized. Since the dictionary itself has ambiguity, it is extremely resistant to misalignment.

[Brief explanation of the drawing]

Figure 1 is a conceptual diagram showing the frequency of feature point appearance, Figure 2 is an explanatory diagram showing the ambiguity of feature points, Figure 3 is a data diagram showing the frequency distribution of feature points, and Figure 4 is the features after normalization. FIG. 5 is a block diagram of a circuit for implementing an embodiment of the present invention; FIG. 6 is a data diagram of a one-dimensional fuzzy dictionary; and FIG. 7 is a data diagram of a two-dimensional fuzzy dictionary. be. 1... Dictionary storage unit, 3... Feature point number addition unit, 4.
...Membership value addition section, 5...Distance calculation section. Fig. 2: Explanation of ambiguity of feature points Fig. 1: ef scattered point appearance robustness Fig. 3: minutiae appearance frequency distribution diagram Fig. 4: data of two-dimensional fuzzy dictionary ×

Claims (1)

[Claims] (1) In a method of recognizing patterns such as characters and figures using a dictionary to be matched, a large number of samples are created for a certain pattern, feature points of the pattern in each sample are extracted, and the pattern is drawn. ax in the plane area
At the same time as assigning b mesh regions (a and b are integers), a corresponding mesh region is set in the dictionary storage unit, and the appearance frequency, which is the number of times the feature point appears, is determined for each mesh in this mesh region. , create an appearance frequency distribution corresponding to the mesh area, normalize each frequency on the appearance frequency distribution using a reference frequency, set this normalized value as a membership value, and use this membership value as Based on this, a two-dimensional fuzzy dictionary is created by assigning a membership value to each mesh in the mesh area in the dictionary storage unit, and the two-dimensional fuzzy dictionary thus obtained is stored in advance in the dictionary storage unit for each pattern along with the reference number j of feature points. After registering, extract the feature points for the pattern to be recognized, find the total number of feature points m and the mesh position (x_1, y_1) (1≦i≦m) of each feature point in the mesh area, and create a dictionary. With a certain pattern PT_k in the storage unit as a matching target, the mesh position (x_
1, y_1), the membership value f_k(x_
1, y_1), and each membership value f_k(x_1, y_1) is subtracted from the confidence level 1.0 to obtain the total value Σ(1.0-i=1 f_k(x_1, y_1) and the feature point Based on the reference number j and the total number m of feature points, a distance D_k corresponding to the degree of separation between the pattern to be recognized and the pattern PT_k to be matched is calculated, and this distance D_k is calculated from all the feature points in the dictionary storage unit. A pattern recognition method characterized by calculating the patterns, and selecting an appropriate number of patterns from among them in order of the smallest value as a recognition result.