JPS60142788A - Feature amount evaluating method in graphic recognition - Google Patents

Abstract

PURPOSE:To attain identification by applying the best evaluating method (discrete distribution) in the split performance to a category even if an object category is in any distribution state so as to evaluate and select the proper feature amount of plural categories. CONSTITUTION:The evaluation is conducted along the internal operation of a discrete distribution calculating section 9 of a feature amount F2. A category managing section 15 generates two category combinations from Ca-Cf so as not to be duplicated to a feature amount distribution data 14 of each category stored in a form compressed into a mean value and a standard deviation. The distribution data (mean value and standard deviation) of the two categories are transferred to a inter-distribution distance calculating section 16. A category combining and management section 15 repeats the operation until all the combinations of categories are outputted at each start command from a discrete propriety discriminating section. Thus, a proper feature amount is evaluated and selected and the identification is attained to plural given categories.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to figure recognition, and particularly to a method for evaluating feature amounts of a figure to be recognized.

[Background of the invention]

To identify and recognize objects, a recognition dictionary is created and a decision tree structure is used.
0 and the recognition target is identified according to a recognition dictionary created in advance.

The recognition dictionary is created by extracting trace amounts at various times for each recognition target and then analyzing the distribution data of each category. In this case, it is desirable to automatically create the recognition dictionary using a standardized method in order to save the trouble of analysis and prevent the analyst's subjectivity from being mixed in. The structure of the recognition dictionary is a structure (
decision tree structure) is the most promising. However, when creating a dictionary with a decision tree structure, a problem that always arises is which feature t should be used at each node of the tree. In other words, in order to divide the category candidates of each node of the tree into several small groups of category candidates, it is necessary to evaluate and rank various features (features are similarly used in other structural dictionaries). However, here we have a decision tree structure dictionary in mind).

Conventional feature ranking methods include the separability method (for example, "Applied Image Analysis" by Junpei Tsujiuchi, Kyoritsu Shuppan) °, the stability coefficient method (
For example, Japanese Patent Application Laid-Open No. 57-25078) and the variance ratio method (for example, "Mathematical Study on Feature Extraction in Pattern Recognition" by Mr. Otsu, Electric Research Institute Research Report No. 818). Each will be briefly explained below.

Assume that the category group to be attempted to be divided (identified) is Ca-Cd, as shown in FIG. 1, and the feature amounts prepared for this are Fl-Fs. And Fl, F2,
Assume that data is collected several times for each feature of Fs, and a distribution as shown in FIG. 1 is obtained. child\
For example, in the case of a circular part, the feature value is the outer circumference of the part. the amount of light falling on a circular object;
In other words, due to variations in brightness, etc., the
A distribution curve as shown in (C) is obtained.

(1) Separation method: Separation is called for Fl-Fs.

The evaluation value is calculated, and the characteristic value of the degree of separation is used for the category group in Figure 1. Here, the degree of separation is expressed by equation (1).

+1 to σ - σ where μk : Average value of category Ck in feature FI μ1c41: Average value of categories Ck and 1 (category with the next largest average value) adjacent to Ck σk : Category in feature F1 ++ to standard deviation σ of Ch: standard deviation of categories Ck and I adjacent to Ck (2) Stability coefficient method: Calculate the stability coefficient eFt-Fs.
Priority is given to feature quantities with larger values. Furthermore, the stability coefficient is (
2) It is expressed by the formula.

(“+1 for k・μ, +1 for lσ, and +1 for σ are the same as the degree of separation)”.

(3) Variance ratio method: Variance ratio f:F+ -Fs'! , and give priority to features with larger values. In addition, the variance ratio is 1 μ here; the average value for the category σ2 ; the square of the standard deviation of When evaluated, it turns out as shown in the figure. According to this, F3 can be divided into four categories (Fs is the most effective), whereas
All of the methods (1), (2), and (3) prioritize (select) Fl and F2. (1).

(2) becomes large when one category is extremely far away from the distribution group of other categories, such as CD in FIK, and (3) becomes large when all categories are largely divided into two, such as F2. This is because it becomes larger.

Here, if you create a decision tree dictionary using each method, the second
The results are as shown in Figures (-i), Φ), and (C). First, when the feature values are ranked for Ca, Cb, C6, and Cd based on the degree of separation and stability coefficient, Fl becomes the most dominant. Therefore, on the axis of Fl, the parts that can be safely divided are between Cc and cd, so Ca and Cb.

It can be divided into CC and CD groups. Next, Ca and Cb.

When ranking the feature values for Cc, F2 becomes the most likely. If the process is repeated in the same way, the result will be as shown in the figure (the same applies to the dispersion ratio). Since the recognition processing time is proportional to the number of features used when outputting one recognition result, the number of features needs to be as small as possible.

Therefore, it is better to recognize only F3 as shown in FIG. 2(C). In other words, the conventional method does not properly represent the segmentation performance of the feature values, making it impossible to rank the feature values stably.・Inefficient decision tree dictionaries are created, resulting in increased recognition processing time. ing.

[Purpose of the invention]

The purpose of the present invention is to provide a method for ranking features, which is a problem when creating a recognition dictionary with a decision tree structure.

It is an object of the present invention to provide a method for determining all evaluations of feature quantities with the best division performance for a target category, regardless of the distribution state of the target category.

[Summary of the invention]

The present invention calculates a frequency distribution for each of a plurality of feature quantities for a plurality of categories corresponding to a figure to be identified, and selects two categories that do not interfere with each other from the distribution of categories determined for each of the feature quantities. Calculate the number of discrete distributions, which is the number of combinations, and find the value of the number of distributions that is as large as possible among the calculated numbers of discrete distributions? The feature of this method is that it selects feature quantities that have the same characteristics and employs them as feature quantities for figure identification to perform figure identification.

[Embodiments of the invention]

First, I will explain the basics. The evaluation of feature amounts depends on how many categories that are the object of recognition can be divided into when using a certain feature amount. This will be explained using FIG.

Suppose (7) that the category 2 to be identified at an arbitrary node 1 of the decision tree is C-=Ct. At this time, by using a certain feature (3), C
One group of a, Cb, Cc (4), Ca, cd
.

Assume that it is divided into three groups: one group of Ce (5) and one group of CI, Cf, and Cg (6).

The details of the division are shown below.

(1) C was distinguished (divided) from Cd, C-, Cf, and Ct.

(2) Ch was distinguished (divided) from Cd, C-, Cf, and Cg.

(3) C was distinguished (divided) from Co, Cf, and Cg.

(4) Differentiated (divided) into Cd1i, Cf, and Cg (
C.

(5) C, is not distinguished (divided) from CI, Cb, and Ce, but there is overlap) 0 (6) Cf, Cg are distinct ( (split) not (ca, Cb
.

Cc and cd, but there is some overlap) In other words, what is the division of categories based on feature amounts?

It is composed of many combinations of divisions of two categories. Therefore, we decided to rank the feature amounts based on the total number of combinations D- of the two divided categories (8). For example, when calculating the total number of combinations for the example shown in Figure 1, Fl is DC-3°F2 is De=4-
Fs becomes D-=6 (e Osaka).

In other words, the result is that F3 is the most effective feature quantity. This is not a phenomenon specific to the distribution example in Figure 1; no matter how complex the target category is distributed, the 5-division performance and the total number of combinations (hereinafter referred to as the number of discrete distributions) correspond appropriately. There is. The present invention was born from a new discovery of this phenomenon.

Embodiments of the present invention will be described using FIGS. 4 to 6.

Now, consider an example in which binary patterns of categories Ca to Cf are recognized as shown in FIGS. 4(a), 4(b), and 4(c). At this time, the feature values F1 to Fh are calculated several times for C and −Cf.
It is assumed that the frequency distributions shown in FIGS. 4(b) and (C) are obtained by calculating . Here, for convenience, Fl is the perimeter of the outer shape of the binary pattern, F2 is the perimeter of the hole, and below, F3 to Fk
The explanation has been omitted. Using this example, the method for ranking feature values will be explained. In other words, in this example, Fig. 4 (a
) shows an example of the identification figure (like a circular part with a hole (9)).

Although there are various possible feature quantities, this example shows the case where the outer circumference length and the hole circumference length are respectively Fl and Fz.

FIG. 5 shows an outline of the feature ranking method. This method consists of the following parts.

A split category code set storage unit 7 that stores the code set of the category that is the target of splitting. Feature amount F
Feature value F for calculating the number of discrete distributions of the category on the l axis
1, a discrete distribution number calculation unit 8 (the same applies hereafter; 9 to 11 are discrete distribution number calculation units on each feature axis). A discrete distribution number comparison unit 12 that arranges the number of discrete distributions on each feature axis in descending order and outputs a sequence of feature names (codes) in the same order. The ranking of the feature quantities is performed by the feature quantity name string storage section 13 that stores the results.Next, the operation of each section will be explained.

Regarding the category described in the code set storage unit 7 of the category to be attempted to be divided (initial state is C, -Cf), each of the discrete distribution number calculation units 8 to 11 of the feature quantities F1 to (1G Fk) Each discrete distribution number is calculated and the value is output to the discrete distribution number comparison unit 12.Next, the discrete distribution number comparison unit 12 calculates each discrete distribution number in descending order of the steering center, and in the same order, calculates each column of feature quantities. The number of discrete distributions is a concept newly devised by Imaji, and is a total combination of two categories whose frequency distributions (probability density distributions) do not interfere with each other. The method of calculating the discrete distribution number is explained in Fig. 6.Each discrete distribution number calculation unit 8, 9゜10.11...
・It consists of the following parts.

Categories (C, -Cf) stored in the feature distribution data storage unit 14 and code set storage unit 7 that store feature distribution data for each category (average value and standard deviation of the feature values for each category) A category combination management unit 15 that generates a combination of two categories from among them without overlapping and outputs distribution data of the two categories, and distribution data of the two categories output by the category combination management unit 15. From this, the results of the inter-distribution distance calculation unit 16, which calculates the distance between the distributions of the two categories, and the appearance range parameter storage unit 171, which stores the parameters for calculating the distance between the distributions, are determined. Discrete pass/fail judgment unit 18/Discrete pass/fail judgment unit 18 for determining whether two categories are discrete.
A discrete determination parameter storage unit 19 stores a threshold value DW used for discrete determination, and a discrete distribution number counter 20 sequentially counts the number of times the discrete pass/fail determination unit determines that two categories are discrete. be.

Next, the operation of each part will be explained (the explanation will be given along with the internal operation of the discrete distribution number calculation unit 9 of the feature t F 2).

For the feature distribution data 14 of each category that is stored in a format compressed into mean values and standard deviations, the category combination management unit 15 creates combinations of two categories (for example, the Figure 4 Cf and Cc)
is generated and transferred to the distribution distance calculation unit 16 as distribution data (average value and standard deviation) 1 of the two categories. Thereafter, the combination management section 15 repeats the above operation for each activation command from the discrete pass/fail determining section until all category combinations are output.

The inter-distribution distance calculation section 16 calculates the inter-distribution distance expressed by the following formula using the average value and standard deviation of both categories transferred from the combination management section.

θ + l'r2 The lower and upper limit of the appearance range of the category with the larger average value among the two categories (θ in Figure 4 (C)
c1. θex) θ1.1 or 2 ~ Smaller average value h of the two categories
The lower upper limit value of the appearance range of the category on the side (θf1.θfz in Figure 4 (C)) The lower upper limit value θIf +2 of the appearance range of each category is ・θII+2=mean value ±α×standard deviation ・・・・・・・・
. . 0 expressed as (5) where α is supplied from the appearance range parameter storage unit 17. Normally, when the feature amount is normally distributed, if α-3 is used, it is possible to express the appearance range of 99.7 yen of a certain category.

Next, in the discrete pass/fail judgment section 18, the inter-distribution distance outputted from the distance calculation section 16 is determined by the discrete judgment parameter storage section 19.
If the threshold value Dw supplied from
1 is added to , and the category combination management section 15 is activated again. Similar processing is performed for the next combination of categories. In addition, if the distance between the distributions is smaller than the threshold DW, the two categories are considered not to be discrete and cannot be divided, and the category combination management unit 15 is activated without performing any other processing. Similar processing is performed for combinations of categories.

In the example of FIG. 4, α=3. When processing is performed with D-=0,
In the case of Fl, the combination of discrete categories is (Ca, C
d), (Ca, C.).

(Ca, Cf), (Cb, Cd), (Cb, Ca).

(Cb, Cf)l (Ca, Cd), (C6, C・),
(C.C.

Cf) and the number of discrete distributions is 9. In the case of F2, it is 12. Therefore, F2 is prioritized over Fl. In other words, it is better to use F2 than to use feature amount F1.

In this example, the judgment of whether or not the categories are discrete, which is necessary when calculating the number of discrete distributions, is performed using the above-mentioned inter-distribution distance. This has the following effects.

■The distribution distance can be determined by an extremely simple calculation from the mean value and standard deviation of both categories, and the appearance range parameter α.

■ By appropriately changing the appearance range parameter α, which determines the width of the appearance range of a category, and the discrete judgment parameter Dw', which is used as a threshold for determining whether or not to use discrete Because the degree can be adjusted,
A flexible system can be realized. For example, if there are only identical parts in one category, such as industrial parts,
Variations in the calculated values of feature quantities are related only to random errors due to unevenness in illumination and lenses, so data collected several times will almost always have a normal distribution. Therefore, in such a case,
It is also possible to set α-3, Dw=O [however, only if a risk rate of 0.3% is allowed]. However, in the case of cases with large deformations, such as the sorting of fruits and fish species, the collected data does not necessarily have a normal distribution, and it may be desirable to include enough deformed examples, so α and l) w should be increased. It is better to do so. (For example, α=4, DW=0.3, etc.).

In the above-mentioned embodiment, there are several recognition targets for each class as shown in Figure 4, and ranking is performed to obtain effective feature quantities during recognition to output all of the recognition targets to which a certain input pattern belongs. explained the method, but by just slightly changing the operation of the "Category combination management section" 15 in Figure 6, when distinguishing between one type of non-defective product category and several types of defective product categories, as in defective product inspection. A method can be used to rank features to obtain effective features.

As shown in FIG. 1, it is now assumed that category Cb is a non-defective item and the other items are defective. For example, as an example of tablet inspection, Ca
is [kake J, Cc is ``bocchiko'' is a tablet with a small lump attached to it, ca is ``stick''(``<peck'' is a tablet with two sticks attached to each other), F! is the area, F2 is the perimeter, and F3 is the shape factor (perimeter 2/area). In this case, there are two types of recognition: a good product or a defective product, so the feature value is C, which is a good product.
It is good to be able to distinguish between B and other defective products C-, Cc, and ca. That is, the greater the number of categories that are discrete from non-defective products (Cb), the better. Therefore, the category combination management department operates by selecting good products (Cb) and other categories (Cb).
By sequentially generating combinations of (m, Cc, Cd) and outputting the distribution data of both, the method for ranking the feature values can be easily realized. For example, in Figure 1, the number of discrete distributions in summer F is 1°F2 is 2. F3 is 3 - Fs can be judged to be the most effective method According to this method - In defective product inspection where there are many defective product categories for one non-defective product category - Feature value ranking for creating a recognition dictionary can easily provide a method, and the output collection order can appropriately determine the quality of the division performance.

According to the present invention, feature values used in a recognition dictionary with a decision tree structure are used for ranking features using the conventional separability method. Compared to the case of selection, it is possible to reduce the number of steps required to obtain one recognition result, and as a result, the recognition time can be significantly shortened. For example, if we consider the case of creating a judgment book structure dictionary in the example shown in Figure 1, in the conventional method, the degree of separation etc. is assigned to the maximum feature amount and each node. Two or more features will be used for this purpose. On the other hand, according to the discrete distribution number method of the present invention, the collection amount used is 1, as shown in the figure.
Become an individual. Normally, the calculation time for feature quantities is very long, about 20 milliseconds/item, so it is reasonable to assume that the recognition processing time is almost determined by the number of features.Therefore, with this method, it is 1/1/2 of the conventional method. It becomes possible to recognize and output two or more processing times.

〔Effect of the invention〕

According to the present invention, it is possible to evaluate and select the optimum feature amount for a plurality of given categories and perform identification.

[Brief explanation of the drawing]

Figure 1 is an explanatory diagram of the results of the feature quantity distribution example and conventional feature ranking. Figure 2 is an example of a decision tree recognition dictionary created using the conventional method and the method of the present invention. Figure 3 is the result of the present invention. An explanatory diagram of the discovery of the causative event, Figure 4 is a feature value distribution diagram for detailed explanation of the present invention, Figure 5 is a conceptual diagram of an embodiment of the present invention, and Figure 6 is an embodiment of the present invention. Conceptual diagram of the center. 7...Divided category set, 8-11...Discrete distribution number calculation unit for feature quantity・12...Discrete distribution number comparison unit・13
. . . Feature ranking sequence, 14. Each category distribution data, 15. Category combination management unit, 16. Inter-distribution distance calculation unit, 17. Appearance range parameter storage unit,
18...Discrete judgment part ・19...Discrete judgment Parano 11 figure μaw+ μg-'f Me' 7%Wazawa Fs ■ ■ 'V13 figure Chi S Otomi Y4 figure ¥5 ward

Claims (1)

[Claims] 1. In a method for evaluating feature quantities as identification indicators in pattern recognition, the frequency distribution is determined for each of a plurality of feature quantities (n) for a plurality of categories (m) corresponding to a figure to be identified. The number of discrete distributions, which is the number of combinations of two categories without mutual interference, is calculated from the distribution of categories (m) considered for each of the features (n).
A method for evaluating feature quantities in figure recognition, characterized in that a feature quantity having the largest possible value of the calculated discrete distribution number is selected from among the feature quantities (n).