JP2643293B2 - Dictionary creation method - Google Patents

Description

The present invention relates to a method for creating a dictionary used for recognizing patterns such as characters and figures.

B. Summary of the Invention The present invention is used for recognizing patterns such as characters and figures, and in a method of creating a dictionary to be compared with the pattern, the method for expressing a positional relationship between feature points of the pattern. The concept of fuzzy set (fuzzy set) is adopted, mesh areas are allocated, the appearance frequency of feature points is obtained for each mesh, and the mesh areas are separated to normalize the entire mesh area for each separated area. A dictionary can be automatically generated by creating a one-dimensional fuzzy dictionary in addition to a two-dimensional fuzzy dictionary for all areas of each separation area, and shortening the manual work time for dictionary registration. The storage capacity can be further reduced.

C. Prior Art When recognizing patterns such as characters and figures, for example, based on black and white binary image data (input pattern) obtained by operating a document or drawing with an input device such as an image scanner. The position and the like are compared with a standard pattern (dictionary pattern) of characters or figures stored in advance, and the most overlapped dictionary pattern is finally obtained as a recognition result.

As a specific example of the recognition method, for example, the input pattern and the dictionary pattern are each divided into small square lattices (mesh), and black (1) when the lattice points are characters or figures,
Otherwise, in the case of expressing white (0), the degree of overlap between the two patterns of the input and the dictionary can be obtained using the Hamming distance. In this case, 0 is set when the input and the dictionary are black and black or white and white, and 1 is set when the input and dictionary are black and white or white and black. That is, the more they overlap, such as black and black, and white and white, the smaller the sum (Humming distance). Due to this property, a pattern that most overlaps with the input pattern among a large number of dictionary patterns, that is, a pattern with the smallest Hamming distance is used as a recognition result.

D. Problems to be solved by the invention With this method, if the position of the input pattern is shifted from the position of the dictionary pattern, the size is different, or the inclination is different, the degree of overlap will change,
There is a disadvantage that recognition errors easily occur.

There is also a problem in terms of storage capacity. The space required for one character or one graphic pattern is not limited to the space of 8 x 8 pixels (alphanumeric characters) or 24 x 24 pixels (kanji) like a character font for display. At least 3 to 5 times the space is required. In particular, in the case of kanji or complicated figures, stable processing cannot be expected unless about 128 × 128 pixels or more are secured.

The dictionary type of characters and graphic patterns is more than 6000 characters including the second level in the case of Chinese characters, and more than 1000 characters when there are many figures.

For such a huge dictionary type, 128
Assuming that × 128 pixels are secured, 2K bytes are required as memory. In other words, the previous kanji is over 12MB,
A figure requires a storage area of 2 Mbytes or more.

Due to the hardware configuration, this capacity is so large that it cannot be ignored, and also affects the search and matching processing of the dictionary due to processing time and the like.

Further, in terms of dictionary types, it is necessary to manually register the dictionaries one by one for each of the enormous types described above, and there is a problem that the time required for the work is enormous.

In summary, the prior art has the following three major problems.

Establishment of a recognition method with high recognition accuracy Establishment of a dictionary construction method for reducing storage capacity Establishment of an automatic dictionary generation method The present invention focuses on solving particular problems.

E. Means to solve the problem Characters and graphic patterns have a positional relationship described according to human historical or empirical rules, but the points that characterize the pattern (feature points) themselves have ambiguity. .

For example, in the case of a representative feature point of the character “A” indicated by a circle in FIG. 3A (this is referred to as a vertex), the position of the vertex itself is shifted left as viewed from other feature points ( FIG. B), and swings to the right (FIG. C). In addition, it is cut off at the vertex (d in the figure).

However, it is unlikely that only the vertex is located below other feature points as shown in FIG.

As a result, the vertex of the character "A" has regularity on a very ambiguous scale that it is closer to the center and closer to the center in terms of positional relationship than other feature points.

The present invention takes into account the fact that the positional relationship between the characteristic points of a character or a graphic pattern has ambiguity, expresses the positional relationship by introducing the concept of an ambiguous set (fuzzy set), and expresses this in a dictionary. You are trying to register.

FIG. 1 shows the flow of the method of the present invention. First, a large number of samples are prepared for a certain pattern, preprocessing such as noise processing is performed, and then feature points of the pattern in each sample are extracted. And mx in the plane area where the pattern is drawn
In addition to allocating n (m and n are integers) mesh regions, an appearance frequency that is the number of appearances of the feature point is obtained for each mesh in the mesh region, and an appearance frequency distribution corresponding to the mesh region is created. .

Thereafter, a mesh having the maximum value among the appearance frequencies of the respective meshes is used as a search start point of the first area, and a mesh included in the first area is searched according to a predetermined rule. Similarly, a mesh having the maximum value of the appearance frequency is set as a search start point of the second region, and a mesh included in the second region is similarly searched. By repeating such processing, the mesh region is converted into one or more regions. To separate. On the other hand, in order to normalize the appearance frequency, a reference frequency is obtained based on, for example, the total number of appearance frequencies.

Subsequently, for each separation region, the appearance frequency of the mesh region including only the separation region is normalized using a reference frequency to obtain a membership value, and thereafter, each line of each mesh of the mesh region is included in the line. The maximum value is determined from the membership values that are included in each column, and the maximum value is determined from the membership values included in that column for each column, and these maximum values are grouped for each separation area to create one-dimensional fuzzy data. Then, this is registered as a one-dimensional fuzzy dictionary.

Furthermore, a one-dimensional fuzzy dictionary is created based on the one-dimensional fuzzy data as follows. That is, the X coordinate xi is assigned to the i-th mesh column from the end of the mesh area, and the Y coordinate yj is assigned to the j-th row from the end, and xi, y is calculated for the one-dimensional fuzzy data created for each separation area.
X (i) and Y (j) are the membership values corresponding to j, respectively.
Is displayed, the smaller one of X (i) and Y (j) with respect to the two-dimensional membership value f (i, j) at the mesh position of (xi, yj) for each separation region. Assign to get a two-dimensional set of membership values for the isolation region.

Next, the maximum value of f (i, j) of each separation region is assigned to the membership value F (i, j) at the mesh position (xi, yj). A two-dimensional fuzzy dictionary is created for all areas of each separation area.

F. Examples Specific procedures of the method of the present invention will be described below.

(1) Preparation for dictionary creation In order to create a dictionary, a plurality of samples are prepared for the same character or graphic.

These become patterns for creating a dictionary when a certain dictionary is created.

(2) Creation of feature point appearance frequency distribution (two-dimensional histogram) Samples are repeatedly input and feature points are extracted.

The same m × n (m, n is an integer) mesh space as the dictionary space is taken, and the frequency of occurrence of these feature points is distributed as shown in FIG.

For example, when the dictionary space is a 10 × 10 mesh space, the histogram also takes a 10 × 10 mesh space. At this time, if the position of a certain feature point is obtained as z (x, y) = (5,3), one is added to the frequency of the position Z (5,3) on the histogram.

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 represented by the following equation.

P (x, y) = P (x, y) +1 (Initial value of P (x, y) 0) (1) (3) Area separation in feature point appearance frequency distribution For the same character or figure If a frequency distribution of a plurality of samples is taken, a feature point is likely to appear at a similar position, and a frequency peak is often present in the vicinity thereof.

Generally, the peak value is sharp and high in the case of a fixed pattern such as a typeface, but tends to be gentle and low due to variations in the case of an irregular pattern such as a handwritten character.

Therefore, it is necessary to reflect in the dictionary in consideration of an area where the feature points are likely to appear.

The method of the present invention separates regions using a feature point appearance frequency distribution. The region is separated by repeating the following procedure.

Detection of Search Start Point The point Z (x, y) having the maximum peak value (maximum appearance frequency) in the feature point appearance frequency distribution is set as the search start point in the first area A (the point indicated by a circle in FIG. 4). Equivalent).

Detection of Region Near Start Point Assuming that the frequency at the search start point Z is P and the frequency at any of the eight neighboring points Zi about Z is Pi, the region A is expanded under the following conditions. In addition, as shown in FIG. Adjacent horizontally and diagonally.

If Pi ≠ 0 P ≧ Pi, Zi∈A (2) where i = 0, 1,..., 7 Actually, the search start point Z is the maximum peak value, and P <Pi does not lie in the vicinity of it. Therefore, in the normal case, the neighborhood point Zi satisfying Pi を 満 た す 0 is included in the area A.

 This process is performed for all eight neighborhoods.

Movement of search point and detection of area If Zi is a point in area A, the search point is moved to Zi.

This is set to Z again, and its frequency is set to P. In addition, an arbitrary neighboring point in the vicinity of the new search point Z is set to Zi again, and its frequency is set to Pi.

At this time, whether or not the neighboring point Zi is included in the area A with respect to the search point Z is determined by the following conditional expression.

Pi ≠ 0, P> Pi If P ≧ Pi ₋₁ , Zi∈A (3) P ≧ P _{i + 1} where i = 0,1,..., 7 When i = 0, Pi ₋₁ = P performed for 8 near all against _{7 i = 7 P i + 1} = P 0 the new search points Z when.

Repetition of Search Point Movement and Area Detection The area search can be advanced by repeatedly performing the processing of the item.

As a result, if the processing is performed until the neighboring point is no longer included in the area A, the area A is finally determined.

 This is shown in FIG.

In the figure, a feature point appearance frequency distribution having two peaks is shown, in which the maximum peak value is the frequency surrounded by a circle, and that position is the search start point. 8 of this starting point
The neighboring points are included in the area A by all the procedures.
Since all of them become new search points, the procedure for each search point can be repeated. As a result, a new search point is found, and the processing continues in the order of...

In this way all search points are found,
The arrows in FIG. 4 show how the search points move in the following procedure.

With the processing up to this point, it is determined that the area following the maximum peak point is surrounded by a rectangular solid line.

Post-Processing for Regions Obtained from Feature Point Frequency Distribution As post-processing for the regions obtained, clear processing is performed as follows in consideration of boundary conditions. This area is copied to the temporary storage memory.

When a point at the boundary of the region is in contact with another region, it may not be surprising that it is included on that side, and it is an ambiguous part.

Therefore, considering this point, by examining whether or not the vicinity point indicated by 4 in FIG. 6 is in contact with another area,
The following processing is performed.

Let Z be the region boundary point and Zi be any of the four neighboring points.
And The frequencies at those positions are P,
Pi.

 At this time, it is defined as follows.

Definition: If any one of the four neighboring points is not in the self-area and the point is not 0, the point is included in other areas.

When the area boundary point comes back into contact with this definition, a process of reducing the frequency of the area boundary point is performed by the following equation.

P = P-1 (4) The frequency is cleared to 0 for all points other than the area boundary and all points that are not in contact with other areas among the boundary points.

This is shown in FIG. The processing of the rectangular area surrounded by the dotted line corresponds to this processing.

By performing the clearing process in consideration of the boundary in the detection of all regions (repeating process of ~),
This means that the already detected area has been removed from the feature point appearance frequency distribution.

Therefore, when the maximum peak point is obtained based on the frequency distribution of the remaining feature points, this becomes the area start point for the next area.

In other words, the process is repeated to return to the above process and search for the next area.

The area detection processing ends when the area is no longer found after the repetition of the above processing. In the example of FIG. 7, since the other area disappears when the second area is detected, the frequency becomes all 0 as shown in FIG. 8, and the detection processing ends.

(4) Definition of Membership Value in Fuzzy Set and Creation of Dictionary The feature point appearance frequency distribution obtained in the above shows the tendency at which positions feature points are likely to appear.

It was also shown that a region including a feature point can be detected and separated based on the frequency distribution.

If a frequency distribution of a plurality of samples is obtained for the same character or graphic, characteristic points are likely to appear at similar positions, and a frequency peak is often present near the characteristic points.

Conversely, by extracting the peak position, the feature point position can be estimated.

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

Here, a method of defining the membership values of the normalized one-dimensional and two-dimensional fuzzy sets will be described below step by step.

Normalization of feature point frequency distribution The frequency distribution needs to be normalized because the larger the number of samples for one pattern, the higher the overall frequency and the more difficult it is to make objective judgments. is there.

With the hint that the feature points often correspond to the peaks of the frequency distribution, the normalization processing is performed as follows.

First, the concept of normalization will be described. The above hint is based on a one-dimensional histogram in which the frequency of appearance of feature points is plotted on the horizontal axis, and the frequency (frequency) of the frequency is plotted on the vertical axis.
The higher the feature point appearance frequency is on the right side of the horizontal axis, so that the feature point appearance frequency occupying the right side of a certain reference point on the horizontal axis indicates that only those feature points with high expectations are gathered.

However, such a histogram is not actually taken, and the reference point is easily obtained as follows.

Assuming that the number of feature points (total value of appearance frequency) of all samples is K, a value obtained by multiplying by a certain ratio Ck is obtained as a peak extraction number k up to a reference point.

k = Ck · K (5) Next, in the feature point appearance frequency distribution, search is performed in descending order of frequency, and the one that takes the value of the k-th frequency is obtained, and this is set as a reference frequency Pk. This is the same idea as the one obtained in the one-dimensional histogram in order from the rightmost side of the horizontal axis.

Based on Pk, normalization is performed for each of the separation areas separated in the above item (3) as follows.

If a certain frequency in the separation region is Pr (i, j) and the normalized frequency is Mr (i, j), then Pr (i, j) = 1.0 ... Pr (i, j) ≧ Pk (I, j) / Pk ... Pr (i, j) <Pk (6) Here, Pr (i, j) and Mr (i, j) represent the frequency and the frequency at the position (xi, yj) of the mesh region, respectively. The frequency after normalization,
1 ≦ i ≦ m and 1 ≦ j ≦ n. Also, 1 ≦ r ≦ R, where R is the number of regions.

In this way, for each separation region, the frequency of all meshes in the mesh region including only the separation region (mesh region other than the separation region, the frequency of which is cleared to 0) is normalized.

The normalized frequency Mr (i, j) obtained by the definition of the membership value in the one-dimensional fuzzy set is in a real number section of 0 to 1, and the appearance frequency of the feature point becomes higher as it approaches 1.0,
It is highly likely that the included positions are points that express the features well. In other words, the closer the value is to 1.0, the higher the degree of certainty is. Therefore, it can be used as an index indicating an ambiguous measure of the feature point position, and can be used as it is as a membership value.

Here, this idea is further developed, and based on Mr (i, j) obtained for each separation area, the projection is performed on the x-axis and the y-axis.
A method for obtaining a membership value in a one-dimensional fuzzy set will be described.

Assuming that the membership value in the x direction, that is, the membership value corresponding to the coordinate xi is Xr (i), the membership value is obtained by the following calculation.

Xr (i) = max (Mr (i, j), Mr (i, 2), ..., Mr (i,
n)) (7) Assuming that the membership value corresponding to the y-direction membership value, that is, the coordinate xj is Yr (j), the value is obtained by the following calculation.

Yr (j) = max (Mr (i, j), Mr (2, j), ..., Mr (m,
j)) (8) where max (a, b, c) represents the maximum value among a, b, c. Also, 1 ≦ i ≦ m and 1 ≦ j ≦ n.

As a result, m and n sets of membership values for each one-dimensional direction of x and y are obtained in one certain separation region r, and one-dimensional fuzzy data is obtained.

Creation of Two-Dimensional Fuzzy Dictionary Using the above membership values, a two-dimensional membership value in one region r is determined as follows.

fr (i, j) = min (Xr (i), Yr (j)) (9) where min (a, b, c) represents the minimum value among a, b, c.

For example, when fr (4,0), Xr (4) is compared with Yr (0) and the smaller one is selected. When Xr (4) is smaller, fr is smaller.
(4,0) takes the value of Xr (4). Similarly, when fr (6,3), Xr (6) is compared with Yr (3), and if Yr (3) is smaller, fr (6,3) takes the value of Yr (3). In this way, by obtaining m × n, a set of two-dimensional membership values for the region r is obtained. This is obtained for all regions.

According to the concept of the fuzzy set, if a union for all areas is obtained as follows, one dictionary can be obtained for one pattern.

F (i, j) = max (f ₁ (i, j), f ₂ (i, j),..., F _R (i, j)) (10) For example, for F (2,5) When f ₁ (2,5) to f _R (2,5)
, And the one that takes the maximum value is F
(2,5). That is, if the maximum value from f ₁ (2,5) to f _R (2,5) is f ₄ (2,5), F (2,5) becomes f (2,5).
_{4 You} just need to take the value of (2,5).

As described above, a dictionary can be created from a plurality of samples of one pattern.

(4) Creation of a plurality of pattern dictionaries When handling a plurality of patterns, (1) to (3) may be repeated.

(5) Example of dictionary creation FIG. 9 is a diagram showing the characteristic point frequency distribution of the character “A”. In this example, a 9 × 9 mesh area is set in a certain area, and the uppermost position of the character “A” is determined. These are feature points. FIG. 10 shows that the mesh area is divided into a plurality of areas by performing the processing described in the above item (3) on the feature point frequency distribution shown in FIG. 9, and further described in the above item (4). FIG. 9 is a diagram showing one-dimensional fuzzy data created by executing the above-described processing. FIG. 11 is a two-dimensional fuzzy dictionary created by executing the processing described in the above item (4) on the one-dimensional fuzzy data shown in FIG.

Next, a circuit for executing the method of the present invention will be described. As shown in FIG. 12, a scanning circuit 2 scans a sample pattern 1 such as a character or a graphic, and then, applies a noise to the scanning result. Preprocessing such as processing and size normalization is performed by the preprocessing circuit 3. The scanning circuit 2 scans a large number of samples under the control of the controller 6. The data output from the preprocessing circuit 3 is input to a feature point extraction circuit 4, which extracts feature points, normalizes their positions, and outputs them to a feature point appearance frequency distribution creating unit 5, The number of feature points is passed to the feature point counter circuit 9. The feature point appearance frequency distribution creator 5 counts up the frequencies stored in the temporary storage memories 50 and 51 based on the feature point appearance positions obtained by the feature point extraction circuit 4 to create a feature point appearance frequency distribution. The controller 6 repeats the input to the scanning circuit 2 according to the number of samples in the dictionary, controls the generation processing of the feature point appearance frequency distribution, and advances the processing to the reference frequency detection circuit 10 after the completion of the generation processing. The reference frequency detection circuit 10 determines the number of peaks for obtaining the reference frequency based on the total number of feature points (total value of appearance frequency) obtained from the feature point number counter circuit 9 and the set value given from the controller 6. From the characteristic point appearance frequency distribution stored in the temporary storage memory 51, the peak frequencies are sequentially read in descending order, and the frequency at which the given number of peaks is reached is detected as the reference frequency.

On the other hand, the area detection circuit 7 finds a mesh position serving as a separation area search start point from the feature point appearance frequency distribution stored in the temporary storage memory 50, detects a separation area satisfying a predetermined condition, and further increases the separation area. When no more data can be found, the process proceeds to the area normalization circuit 11. The separation areas detected by the area detection circuit 7 are sequentially stored in the temporary storage memory 52 by the area separation circuit 8. Then, the area separating circuit 8 performs post-processing of the area in consideration of the condition of the area boundary, that is, the area of the above item (3) with respect to the appearance frequency distribution in the temporary storage memory 50.

After the processing in the area detection circuit 7 and the reference frequency detection circuit 10 is completed in this way, the area distribution stored in the temporary storage memory 52 by the area normalization circuit 8 is used to calculate the reference frequency detected by the reference frequency detection circuit 10. In addition to the normalization processing, the normalized frequency distribution in each area is stored in the temporary storage memory 52 again.

Further, the x-direction membership value calculator 12 obtains the x-direction membership value for each region by the equation (7) using the normalized frequency distribution of the region unit stored in the temporary storage memory 52, and temporarily stores the value. Stored in the memory 53, the y-direction membership value calculator 13 uses the normalized frequency distribution of the area unit stored in the temporary storage memory 52,
The y-direction membership value is obtained for each area by the equation (8) and stored in the temporary storage memory 53.

Based on the x-direction membership value and the y-direction membership value stored in the temporary storage memory 53 in this way, the combining circuit 14 combines the two-dimensional membership values in area units according to equation (9) to obtain two-dimensional membership values. Fuzzy dictionary memory 54
To memorize.

G. Effects of the Invention According to the present invention, an appearance frequency distribution of feature points of a pattern such as a character or a graphic is created by preparing a large number of samples, and each frequency of the appearance frequency distribution is normalized. Because values are registered as membership values, dictionaries can be created automatically, manual work for dictionary registration is greatly reduced, and there are variations in the position and inclination of input patterns. In addition, recognition errors are less likely to occur.

Furthermore, as a dictionary space, character fonts for display
If there is a mesh space of about 1.2 times to 1.5 times, it can be sufficiently recognized. For example, about 9x9 for alphanumeric characters, complicated kanji
Only 32 × 32 mesh space is required. For example, when the membership value is expressed by a value obtained by dividing 0 to 1 by 0.1, the data expression can be made up of 4 bits because it is sufficient to have an integer of 0 to 10. Therefore, the amount of memory is 32 x 32 x 4 bits per kanji character = 512 bytes, which is 1/4 that of the conventionally required 2K bytes.
And the storage capacity can be reduced.

[Brief description of the drawings]

FIG. 1 is a flowchart showing the flow of the method of the present invention, FIG. 2 is a distribution diagram showing a feature point appearance frequency distribution, FIG. 3 is an explanatory diagram showing the ambiguity of feature points, and FIG. FIGS. 5 and 6 are explanatory diagrams showing points near the mesh, FIG. 7 is an explanatory diagram showing the state of detection and separation of the second region, and FIG. FIG. 9 is a data diagram showing a feature point appearance frequency distribution after the first and second region separation, FIG. 9 is a data diagram showing a feature point appearance frequency distribution, FIG. 10 is a data diagram showing one-dimensional fuzzy data, and FIG. Is a data diagram showing two-dimensional fuzzy data, and FIG. 12 is a block diagram showing a dictionary creation circuit. 1 ... input pattern, 2 ... scanning circuit, 3 ... preprocessing circuit, 4 ... feature point extraction circuit, 5 ... feature point appearance frequency distribution creator, 6 ... controller, 7 ... area detection circuit, 8
……………………………………………………………………………………………………………………………… ………………………………………………………….
... Reference frequency detection circuit, 11 ... Regularization circuit, 12 ... x
Directional membership value calculator, 13... Y-direction membership value calculator, 14... Two-dimensional membership value synthesis circuit, 50 to 54.

Claims (2)

(57) [Claims] 1. A method for creating a dictionary used for recognizing a pattern such as a character or a figure, which is to be compared with the pattern, comprising: creating a large number of samples for a certain pattern; Mxn (m, n are integers) in the plane area where the pattern is drawn
And assigning an appearance frequency that is the number of appearances of the feature points to each mesh in the mesh area, and creating an appearance frequency distribution corresponding to the mesh area. The mesh having the maximum value is used as a search start point of the first region to search for a mesh included in the first region according to a predetermined rule, and then the maximum value of the appearance frequency is determined from the meshes other than the first region. The mesh to be taken is used as a search start point of the second region, a mesh included in the second region is similarly searched, and the above-described processing is repeated to separate the mesh region into one or more regions. In the mesh area including only the separation area, a membership value is obtained by normalizing each appearance frequency using a reference frequency, and then the mesh area is determined. For each row of each mesh, the maximum value is obtained from the membership values contained in that row, and for each column, the maximum value is obtained from the membership values contained in that column, and these maximum values are divided into the separation areas. A dictionary creation method characterized in that one-dimensional fuzzy data is created for each group. 2. An x coordinate xi is assigned to an i-th mesh column from an end of a mesh area, and a Y coordinate yj is assigned to a j-th row from an end. For one-dimensional fuzzy data created for each separation area, xi, When the membership values respectively corresponding to yj are indicated as X (i) and Y (j), (x
The smaller of X (i) and Y (j) is assigned to the two-dimensional membership value f (i, j) at the mesh position of (i, yj), and the two-dimensional member for the separation region is assigned. A set of ship values is obtained, and the maximum value of f (i, j) of each separation region is assigned to the membership value F (i, j) at the mesh position (xi, yj). 2. The dictionary creation method according to claim 1, wherein a two-dimensional fuzzy dictionary is created for all areas of each separation area by a set of (i, j).