JPH0540852A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To improve the production efficiency of a pattern recognizing dictionary and also to shorten the pattern recognizing time. CONSTITUTION:In a dictionary production state, the essential components of a 64-dimensional learning sample inputted through an input part 3 are analyzed. Then the essential component vectors covering up to the k-th (k<64) axis and an average vector are registered in a essential component dictionary 12. At the same time, the membership functions of each axis are registered into a membership function dictionary 13. Meanwhile the membership functions of the essential component axes following the (k+1)-th one are supported as a normal distribution of dispersion sigma, and only this sigma is registered into the dictionary 12. If an unknown pattern is inputted, a feature extracting part 5 extracts the features of the pattern and a essential component evolving part 6 evolves the essential components in each category. A membership value calculation part 7 calculates the membership value and then calculates the resemblance of the categories from the product of the membership value. Under such conditions, the membership functions of the (k+1)-th and its subsequent ones are shown in the normal distribution of the dispersion sigma. Therefore the resemblance can be easily known among categories.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character pattern recognizing device, and more particularly to a pattern recognizing device for recognizing a pattern by utilizing a feature space in multidimensional data analysis.

[0002]

2. Description of the Related Art Conventionally, among the recognition algorithms for character patterns and the like, the identification function is as important as the feature extraction.

The most basic discriminant function is the Euclidean distance, which is expressed by the following equation.

[0004]

[Equation 1]

Here, x is an input pattern vector, x ₁ is each component thereof, and m ₁ is each component of the standard pattern vector of the category to be recognized. The number of dimensions of the feature amount is n.

Since this discriminant function is simple and has a high processing speed, it is often used in the world of pattern recognition. However, it is unreasonable to represent the category to be identified by one standard pattern vector (usually the average vector is used), and the data distribution becomes complicated, and in the case of multiple categories such as character recognition. Had a performance problem.

Therefore, the Mahalanobis distance is used in consideration of the distribution of data.

Now, this will be described by taking the case of recognizing the character pattern 20 shown in FIG. 7 as an example, and the learning character pattern 20 is, for example, as shown in FIG.
It is divided into 64 pixels of D ₁ , D ₂ , D ₃ , ..., D ₆₄ and obtained as a vector expression in the 64-dimensional feature space 30 as shown in FIG.

Then, a large number of sample data are obtained for each character pattern. FIG. 9 shows the character pattern 2 shown in FIG.
It is an example of distribution of 0 sample data.

Then, principal component analysis is performed on this learning sample to establish the appearance of samples on each principal component axis.
2 (C) is assumed to have a normal distribution 25.

Here, if an unknown pattern is input,
For example, in the case of multidimensional data diffraction in 64 dimensions, the appearance probability is obtained on each of the 64 axes, and the unknown pattern is identified as the character pattern with the highest probability of occurrence.

That is, FIG. 10 shows the principal component expansion in the 64-dimensional feature amount space, φ1 is the eigenvector of the first principal component, and φ2 is the eigenvector of the second principal component.
In this case, the coordinates are exchanged as shown in FIG.

In this case, the following equation will be calculated.

[0014]

[Equation 2]

Here, φ 1 and λ 1 are an eigenvector and an eigenvalue obtained from the covariance matrix of the learning data of the identification category, and they respectively match the principal component vector of the data distribution and the variance value on the principal component axis.

On the other hand, there is also a method in which the probability density function is approximated from the distribution of the learning data and the Bayes decision mechanism is used to identify the probability density function. For example, assuming that the data is normally distributed on each principal component axis after the data distribution is principal component expanded, the probability density function at a certain principal component axis i is

[0017]

[Equation 3]

If the principal component axes are statistically independent, then the probability density function of x in the entire pattern space is

[0019]

[Equation 4]

And taking the logarithm of both sides

[0021]

[Equation 5]

It becomes Here, considering the following discriminant function, the probability density decreases monotonically with respect to the probability density function, and the smaller this value, the higher the probability density value.

[0023]

[Equation 6]

This equation is called the Bayesian discriminant function,
Since the formula reflects the probability density function of the data,
If a good sample can be collected, it is theoretically the optimal discriminant function.

[0025]

However, in the conventional method as described above, the sample appearance probability on each principal component axis is assumed to be a positive distribution centered on the average value. Often does not have a positive distribution like the function 26 shown in FIG. 12B. In this case, the assumed probability density function and the appearance probability of the actual sample are different, and the accuracy of pattern recognition is reduced. There was a problem.

Further, in the recognition of characters and the like, the appearance probabilities due to the distribution of learning samples and the distribution of the degree of similarity felt by humans are different. There was a problem of doing.

Therefore, in recent years, the distribution function of similarity is shown in FIG.
As shown in (C), a fuzzy pattern recognition method represented by a membership function 27 has been proposed.

This is because when learning sample data is obtained, a membership function is created for each pixel based on this sample data, and when an unknown pattern is input, this membership function is applied to identify the pattern. Is to do.

That is,

[0030]

[Equation 7]

As a pattern recognition based on the following equation 8, the pattern is recognized in the category having the maximum value of equation 8.

[0032]

[Equation 8]

However, in the pattern recognition method using such a fuzzy pattern recognition method, the above-mentioned inconvenience can be avoided, but when the dimension number of the feature amount becomes large as described above, for example, 64, the identification data Since it is necessary to create a membership function for all principal component axes for all categories (hereinafter referred to as a dictionary), the dictionary capacity becomes large, and principal components must be expanded for all axes during pattern recognition. Therefore, there is a problem that the processing time takes a lot of time.

The present invention has been made in view of the above-mentioned conventional problems, and an object thereof is to provide a pattern recognition apparatus capable of reducing the dictionary capacity and significantly reducing the identification time. To do.

[0035]

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention provides a pattern recognition apparatus that identifies patterns based on similarity on a plurality of principal component axes by multidimensional data analysis. The similarity on the axis is characterized by being calculated by applying a membership function relating to a preset similarity.

[0036]

In the present invention, since the membership function is created only for the similarity on the main principal component axis to identify the pattern, the labor for creating the membership function is reduced and the pattern recognition is performed. Processing time is also reduced.

[0037]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the electrical construction of an embodiment to which the present invention is applied.

First, the configuration will be described. 1 is a central control unit including a CPU and the like, and the central control unit 1 has a bus line 2
The input unit 3, the preprocessing unit 4, the feature extraction unit 5, the principal component expansion unit 6, the membership value calculation unit 7, the similarity calculation unit 8, the determination unit 9, and the output unit 10 are connected via the Line 2 has a work area 11 composed of RAM, RAM and ROM
Principal component dictionary 12 and membership function dictionary 13
Are connected. The input unit 3 is composed of an image sensor and the like, and the output unit 10 is composed of a display unit and the like. Further, the preprocessing unit 4, the feature extraction unit 5, the principal component expansion unit 6, the membership value calculation unit 7, the similarity calculation unit 8, and the determination unit 9 are included.
Is composed of a CPU and the like.

The above is the configuration of the present embodiment, and its operation will be described below.

In the following description, the case of recognizing a character pattern will be described.

First, also in this embodiment, the dictionary creating process is performed as shown in FIG.

This will now be described with reference to FIGS. 4 to 6. First, the character pattern 20 to be registered in the dictionary.
Is read from the input unit 3 such as a scanner as shown in FIG. 4 (a) and input (step 210).

When the character pattern 20 is input in this way, the preprocessing of step 220 is performed to remove noise components, and as shown in FIG. 4B, enlargement or reduction processing to a standardized constant size is performed. To obtain the character pattern 21.

Next, the feature amount extraction processing of step 230 is performed. As shown in FIG. 4C, the feature amount extraction processing is performed as shown in FIG.
A character pattern 22 obtained by applying a Gaussian filter to the character pattern 21 is used.

That is, also in this embodiment, as shown in FIG. 7, the character pattern 20 is divided into, for example, 64 pixels,
The feature quantity on the principal component axis for each pixel is extracted in the feature quantity space in the multidimensional data analysis. Such a feature quantity extraction process uses the Gaussian filter 40 as shown in FIG. Use the sprinkled one.

The Gaussian filter 40 is composed of a plurality of square filter regions each having the same size as each pixel of D ₁ , D ₂ , D ₃ etc. shown in FIG. Has a center filter 41
Has a plurality of filter regions extending radially. By the way, in this case, the pixel pattern on which the center filter 41 is overlapped is weighted with, for example, 100, and the peripheral pixel portions are weighted with the respective weightings shown in FIG. To process.

As a result, as shown in FIG. 4C, filter regions 42, 4 which are bolder than the original character pattern 21 and have a small weighting value around the character contour.
A character pattern 22 having a thin character pattern due to 2, 42, 42 will be obtained.

Therefore, in the processing of step 230, the learning character pattern 20 input in the form of FIG. 4A is first exchanged for the character pattern 22 as shown in FIG. become.

Then, in the subsequent processing of step 240, the principal component analysis is performed based on the sample data, and the average vector (center value), the principal component vector up to a certain value k axis, (k
Approximate values σ of variances on the principal component axes up to the +1) axis and up to 64 axes are obtained. In this case, the pixels for which the principal component vector is obtained are selected in order from the axis having the longest principal component axis. This is because a longer principal component axis means a higher degree of data dispersion, and there is a high possibility that the pixel is a characteristic portion of a character pattern. Further, as described in detail later, in the case of a character pattern, for example, as shown in FIG. 7, most of the 64 pixels have the character pattern 20 as the main component vector not calculated for the (k + 1) axis and thereafter. This is because it is a blank part that does not include any feature quantity.

Next, in step 250, for the principal components up to the k axis, a membership function corresponding to the character transformation on the principal component axes is created. FIG. 6 is an explanatory diagram showing a method of creating a membership function in this case.

The figure shows learning data when the center vector position 0 of the principal component vector for which the membership function is to be obtained is shifted for each unit distance in the plus direction and the minus direction. As shown in the figure, regarding the main components, even if the center vector position is shifted by one unit distance, it looks greatly different. Therefore, a membership function is created based on this data. Then, in step 260, the data obtained in step 240 is registered in the principal component dictionary 12, and the data obtained in step 250 is registered in the membership function dictionary 13. The above processing is performed and registered for each category (character).

The above is the details of the dictionary creation processing.

Next, when the dictionary is created in this way,
Based on this created dictionary, pattern recognition processing as shown in FIG. 3 is performed.

That is, first, when an unknown pattern is input by the input unit 3 such as an image sensor (step 3
10), preprocessing is performed as in the case of the dictionary creation processing (step 320), and the feature amount extraction processing of the input data is performed (step 330).

Next, the extracted feature quantity is subjected to principal component expansion (step 340), the membership function of each principal component axis is up to the k axis, and the normal distribution function of the variance σ is used after the (k + 1) axis. Find the membership value.

Then, the product of the whole is taken as the similarity.

That is,

[0059]

[Equation 9]

[Mathematical formula-see original document] Equation 10 is calculated as

[0061]

[Equation 10]

By the way,

[0063]

[Equation 11]

Is

[0065]

[Equation 12]

Since it can be expanded to, equation 10 can be expressed by equation 13.

[0067]

[Equation 13]

In this way, the unknown pattern is identified in the category having the highest similarity.

As described above, in this embodiment, 64
An unknown pattern of dimensions is identified as follows.

(1) When creating a dictionary, the learning sample is subjected to principal component analysis, and up to the kth axis, the principal component vector, the membership function on that axis, and the average vector are registered as a dictionary. Then, after the (k + 1) th axis, the membership function on the principal component axis is assumed to be a normal distribution with variance σ, and only σ is registered.

(2) When an unknown pattern is input here, after extracting the feature amount, the principal component is expanded for each category, the membership value is obtained on each principal component axis, and the product of them is calculated. Find the similarity to a category.
At this time, in the present embodiment, the membership functions after the (k + 1) th axis are all represented by the normal distribution with variance σ, and therefore the equation for calculating the similarity is as shown in Formula 10. Also, since the expression 10 can be transformed into the expression 13, the expression 13 is actually used.
By calculating, the degree of similarity to that category can be obtained. Then, the unknown pattern is identified in the category having the highest similarity among the calculated similarities.

By the way, in this case, as shown in FIG.
Most of the pixels of No. 4 do not include the character pattern 20, and in general, the higher-order principal components often have a variance close to 0 and have no meaning. Therefore, the recognition performance is hardly changed even if it is expressed by a normal distribution with a constant variance as in this embodiment.

Therefore, as compared with the conventional fuzzy pattern recognition method, the method according to the present embodiment has a dictionary capacity of about k / 64 and a processing time of about k / 64 with almost no change in recognition performance. The effect is obtained.

In this embodiment, the case where the character pattern is represented by 64 pixels has been described, but it goes without saying that the number of pixels is not limited.

[0075]

As described above, in the present invention, only the similarity on the main principal component axis is calculated by applying the membership function relating to the preset similarity. It has an effect that it can be made small and the identification time can be greatly reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing an electrical configuration of an embodiment to which the present invention is applied.

FIG. 2 is a flowchart showing a processing procedure for creating a dictionary.

FIG. 3 is a flowchart showing a processing procedure of pattern recognition processing.

FIG. 4 is an explanatory diagram of a case where a Gaussian filter is applied to an input learning character pattern by noise processing or the like.

FIG. 5 is an explanatory diagram of a Gaussian filter.

FIG. 6 is an explanatory diagram showing a method of creating a membership function.

FIG. 7 is an explanatory diagram when a feature amount of a learning character pattern is extracted with 64 pixels.

FIG. 8 is an explanatory diagram of character patterns in a feature space.

FIG. 9 is an explanatory diagram of a sample distribution in a feature space.

FIG. 10 is an explanatory diagram of the case of performing principal component expansion based on each eigenvector.

FIG. 11 is an explanatory diagram of a case where coordinate conversion is performed based on two eigenvectors.

FIG. 12 is an explanatory diagram of the similarity distribution represented by a normal distribution, a non-normal distribution, and a membership function.

[Explanation of symbols]

 1 central control unit 2 bus line 3 input unit 4 pre-processing unit 5 feature extraction unit 6 principal component expansion unit 7 membership value calculation unit 8 similarity calculation unit 9 judgment unit 10 output unit 11 work area 12 principal component dictionary 13 membership Function dictionary 20,21,22 Learning character pattern 40 Gaussian filter

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] August 6, 1991

[Procedure Amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0005

[Correction method] Change

[Correction content]

Here, x is an input pattern vector, x _i is each component thereof, and m _i is each component of the standard pattern vector of the category to be recognized. The number of dimensions of the feature amount is n.

[Procedure Amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0010

[Correction method] Change

[Correction content]

Then, principal component analysis is performed on this learning sample to establish the appearance of samples on each principal component axis.
Assuming a normal distribution 25 as shown in 2 (a).

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0015

[Correction method] Change

[Correction content]

Here, φ _i and λ _i are an eigenvector and an eigenvalue obtained from the covariance matrix of the learning data of the identification category, and they respectively match the principal component vector of the data distribution and the variance value on the principal component axis. To do.

[Procedure amendment 4]

[Document name to be corrected] Drawing

[Name of item to be corrected] Figure 7

[Correction method] Change

[Correction content]

[Figure 7]

Claims (1)

[Claims] 1. In a pattern recognition apparatus for identifying a pattern by a similarity on a plurality of principal component axes by multidimensional data analysis, the similarity on a principal component axis is a membership related to a preset similarity. A pattern recognition device characterized by being calculated by applying a function.