JPH0420225B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern recognition method for recognizing a brightness pattern of a detection target object from one screen of image data.

Conventionally, the object to be detected is photographed using an ITV camera.
The input screen is scanned, and the brightness and darkness of the object to be detected is determined based on a comparison between a brightness pattern consisting of a predetermined number of continuous image data in the scan direction and a standard pattern prepared in advance that indicates the brightness of the object to be detected. There are ways to recognize patterns.

However, although this pattern recognition method performs input in multiple gradations, the actual image processing is performed by binary conversion, so it is difficult to use optical Good environment and high quality 2
It is practical when it is relatively easy to obtain a valued image, but in environments where there is no special light source for detection and is affected by sunlight, the brightness pattern of the object to be detected may fluctuate. Therefore, there were problems in that proper pattern recognition was not possible and incorrect recognition was performed due to the influence of the background, noise, etc.

The present invention has been made in view of the above circumstances, and provides a pattern recognition method that can recognize the brightness and darkness pattern of a detection target object even under a bad optical environment or when the irradiation light source changes. The purpose is to

According to the first aspect of the present invention, a standard pattern B consisting of image data showing a predetermined number of intensities related to the shape of the detection target object is prepared in advance, and image data of a predetermined field of view in which the detection target object exists is set in advance. By extracting a brightness pattern A consisting of a predetermined number of image data from and determining the correlation between the standard pattern B and the brightness pattern A, it is recognized whether the brightness pattern A is a brightness pattern related to the shape of the object to be detected. When determining the correlation, patterns A' and B' are obtained by subtracting the average value of each pattern from the light/dark pattern A and the standard pattern B, respectively, and transformed around the average value, and Find the normalized patterns A″ and B″ from A′ and B′, and calculate the inner product of the patterns A″ and B″ and the pattern A′ and B″.
The inner product with B'' is obtained as an evaluation value, and based on the two evaluation values, it is recognized whether the brightness pattern is that of the object to be detected.

Further, in the second aspect of the present invention, the pattern
The above recognition is performed using the inner product of A'' and B' and the inner product of the patterns A' and B' as evaluation values.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

First, pattern recognition according to the present invention will be explained in principle. Now, consider a round bar 1 as shown in FIG. 1 as an object to be detected. A standard pattern (FIG. 2) consisting of the brightness Cb _i in the transverse direction of each round bar 1 when photographing this round bar 1 is prepared in advance. Note that the brightness Cb _i is displayed in 16 gradations.

On the other hand, when the brightness Ca _i of pixels continuous in the direction of arrow A shown in FIG. 1 is detected, the result is as shown in FIG. 3, for example. Here, since the diameter of the round bar 1 corresponds to the width of 10 pixels (see Figure 2), 10 pixels are extracted from the pixels continuous in the direction of arrow A above, and the brightness Ca _i The light and dark pattern
Round bar 1 depending on whether it approximates the standard pattern or not.
Determine whether or not there is.

Generally, a pattern expressed by n numbers is considered to correspond to one point in an n-dimensional space, so whether or not two patterns (the brightness pattern and the standard pattern described above) are similar depends on the space. It can be expressed by the distance between two points within. However, due to fluctuations in the light source that irradiates the round bar 1, the brightness shown in FIG. 3 may move vertically in parallel as a whole, or expand or contract in the vertical axis direction. Taking this into consideration, the bright and dark pattern A is designed to be insensitive to changes in illuminance.
From (Ca ₁ , Ca ₂ , ..., Ca _o ), a pattern A'(Ca' ₁ , Ca' ₂ , ..., Ca' _o ) converted around the average value is obtained. Incidentally, and Ca′ _i are expressed by equations (1) and (2), respectively.

= _o 〓 ⁱ⁼¹ Ca _i /n ……(1) Ca′ _i =Ca _i −……(2) Furthermore, a pattern obtained by normalizing pattern A′
A″ (Ca″ ₁ , Ca″ ₂ , …, Ca″ _o ) is expressed as: Find it by Similarly, standard pattern B (Cb ₁ ,
Pattern B′ (Cb′ ₁ , Cb′ ₂ , ..., Cb′ _o ) converted from Cb ₂ ,…, Cb _o ) around its average value and pattern B″ (Cb″ ₁ , Cb″ ₂ ,...
Find Cb″ _o ).

FIG. 4 illustrates the patterns A'' and B'' in an n-dimensional space. Note that the spherical surface represents a unit sphere. The similarity between these two patterns is
It is expressed as the distance l between A″ and B″. Now, as shown in Figure 4, if the length of B'' projected onto A'' is φ, then
The distance l is expressed by the following formula, l=√2(1-)...(4). On the other hand, φ is the following formula, is given by As is clear from the above equation (4), since l and φ have a certain relationship, φ is also a value related to the similarity of the two patterns. Here, if the two patterns perfectly match, φ takes a value of 1.0, if they are completely different, it takes a value of 0, and if the two patterns have the same shape but the positive and negative are reversed, it takes a value of -1.0. this is,
Cross-correlation coefficient R _fg (τ) used for time series analysis, This value has the same meaning as . Hereinafter, this φ will be referred to as pattern similarity. Based on this pattern similarity, the similarity between the standard pattern and the bright/dark pattern is determined.

The above-mentioned pattern similarity φ is an evaluation value that evaluates the similarity of the distribution shapes of bright and dark patterns. Using this evaluation value, it is possible to detect the target pattern even when the absolute brightness of the entire image increases or decreases. . The reason for this will be explained below.

Figures 5 a to e illustrate photographed brightness patterns, Figures 6 a to e illustrate patterns obtained by converting these brightness patterns around the average value, and Figures 7 a to e
is the normalized pattern of each pattern in Figure 6,
Further, FIG. 8 shows examples of standard model patterns (standard patterns).

Compared to the contrast of the bright and dark patterns shown in FIGS. 5a and 5b, the contrast of the bright and dark patterns shown in FIGS. 5c, d, and e is low. However, since the distribution shapes of the respective bright and dark patterns are similar, the normalized patterns of these patterns are almost equal, as shown in FIG.

Since the above-mentioned pattern similarity φ is the inner product of the standard pattern and the above-mentioned normalized pattern, even if the light-dark patterns have low contrast as shown in Figure 5 c, d and e, the light-dark patterns a and b It shows a high value similar to the pattern similarity φ (see the table in Figure 9). Therefore, if the evaluation value of the pattern similarity φ is used, even if the contrast of the bright and dark patterns decreases due to the disturbances described below, , the detected object can be recognized.

a Dust, etc. is present in the space surrounding the detection target.

b Strong light enters the area where the detection target is placed.

c Rust, dirt, etc. are uniformly adhered to the surface of the detection target.

d Changes in the background of the detection target.

By the way, when the contrast of the photographed light-dark pattern is lower than a certain level, the light-dark pattern is often not based on the object to be detected, but is based on, for example, a defective product, a foreign object, noise, or the like.

However, since the contrast cannot be evaluated using only the evaluation value of pattern similarity φ, even if the light and dark patterns shown in Figure 5 c to e are based on the defective products, foreign matter, noise, etc., these It is not possible to distinguish between the bright and dark pattern and the bright and dark patterns shown in a and b of the same figure.

Therefore, in this embodiment, in addition to the evaluation value of pattern similarity φ, an evaluation value of pattern strength that can evaluate the contrast is used.

Now, if the pattern strength is φ, ψ is the following formula, It is expressed as As is clear from Equations (6) and (9) above, the pattern strength ψ is the product of the pattern similarity φ by √ _o 〓 ⁱ⁼¹ (Ca′ _i ) ² , and the bright and dark patterns It represents the absolute size of the pattern, that is, the contrast of the pattern.

The magnitude of the pattern strength ψ of each light-dark pattern shown in FIG. 5 is as shown in the table of FIG. 9, so for example, the pattern strength of the light-dark pattern in FIG. If the value T ₁ between the pattern intensities is set as the threshold for the pattern intensity ψ, the light and dark patterns shown in a and b in the same figure and the light and dark patterns shown in c to e in the same figure can be distinguished based on this threshold. can do.

In this way, the pattern strength ψ can evaluate the contrast of bright and dark patterns, so the pattern similarity φ can be used to evaluate the pattern distribution shape.
By using both the pattern strength ψ and the pattern strength ψ as evaluation values, it is possible to recognize the brightness pattern of the object to be detected, for example, as shown in FIGS. 5a and 5b.

Furthermore, as described later, the pattern similarity φ
A threshold is also set for The thresholds for the pattern similarity φ and the pattern strength φ are appropriately set in consideration of the size of the bright and dark pattern, the surrounding environment, and the like.

T ₂ shown in FIG. 9 shows another example of setting the threshold value for the pattern strength ψ. This threshold T ₂ is set to a value between the pattern intensity of the light and dark pattern in Figure 5 d and the pattern intensity of the light and dark pattern in Figure 5 e, so when this threshold T ₂ is used, It is possible to distinguish between the bright and dark patterns shown in a to d and the bright and dark pattern shown in e of the figure.

In the present invention, the similarity between a standard pattern and a bright/dark pattern is determined using two evaluation values, the above-mentioned pattern similarity φ and pattern strength φ.

FIG. 10 is a block diagram showing an embodiment of an apparatus for carrying out the pattern recognition method according to the present invention. In the figure, the ITV camera 10 photographs the round bar 1,
After the brightness signal is converted into image data showing 16 levels of brightness by an A/D converter 11, the image data is sequentially added to a shift register 12. The shift register 12 has the same number of stages as the number of pixels in the transverse direction of the round bar 1, and applies the image data stored in each stage to the subtracter 13 and the average value calculation circuit 14 as a brightness pattern A. Note that the processing described below is executed until new image data is added to the shift register 12 and the brightness/darkness pattern A becomes a brightness/darkness pattern shifted by one pixel.

The average value calculation circuit 14 calculates the average value of the bright and dark pattern A using the above-mentioned equation (1), and adds this to the subtracter 13. The subtracter 13 performs the subtraction shown in equation (2) above, transforms the brightness pattern A around its average value, and outputs the transformed pattern A' to the normalization circuit 15 and the arithmetic circuit 17.

The normalization circuit 15 normalizes the pattern A′ transformed around the average value as shown in the above equation (3) by dividing it by √ _o 〓 ⁱ⁼¹ (Ca′ _i ) ² , and this normalized Pattern A'' is output to the arithmetic circuit 16.

On the other hand, the standard pattern setter 18 is set with a standard pattern B consisting of the brightness Cb _i in the transverse direction of the round bar 1 as shown in FIG. Arithmetic circuit 20
added to.

Similar to the subtracter 13, the subtracter 19 converts the standard pattern B around the average value calculated by the average value calculation circuit 20, and applies the converted pattern B' to the normalization circuit 21. Normalization circuit 2
Similarly to the normalization circuit 15, the normalization circuit 1 normalizes the pattern B' transformed around the average value, and applies this normalized pattern B'' to the calculation circuits 16 and 17.

The arithmetic circuit 16 has two normalized patterns.
The inner product of A″ and B″ is taken, and this inner product value is output to the determination circuit 22 as pattern similarity φ. The determination circuit 22 has a threshold value appropriately set in advance for determining pattern similarity, and when the input pattern similarity φ exceeds the threshold, it outputs a signal (high level signal) indicating that the patterns are similar. Note that the pattern similarity φ is 1.0 when the patterns are exactly the same as described above, so the threshold value is set to a value slightly smaller than 1.0.

Further, the arithmetic circuit 17 calculates the normalized pattern.
The inner product of B'' and the pattern A' converted around the average value is taken, and this inner product value is output as the pattern strength ψ to the judgment circuit 23.The judgment circuit 23 has an area for judging pattern similarity in advance, is set, and when the input pattern strength ψ is within that region, a signal (high level signal) indicating that the pattern strength is appropriate is output.The region is defined by the patterns of the bright and dark pattern and the standard pattern. When the intensities are exactly the same, that is, in equation (8) above, ψ=(A′・B″)=(B′・B″) =√ _o 〓 ⁱ⁼¹ (Cb′ _i ) ² ……( 10) It is set by appropriately determining the upper limit value and lower limit value based on the pattern strength ψ that becomes .Furthermore, the lower limit value may be used as the threshold value instead of the above region.This is because noise etc. Therefore, when the bright and dark pattern becomes convex and the pattern similarity φ exceeds the threshold value, the pattern strength φ at that time is almost always smaller than the lower limit value.

AND circuit 24 outputs a high level signal only when the output signals of determination circuits 22 and 23 are both high level signals. When a high level signal is output from the AND circuit 24, it is determined that the standard pattern and the bright/dark pattern are similar and that the round bar 1 exists, and the position data of the round bar is extracted from the bright/dark pattern at that time. The position data of the round bar 1 extracted in this way is added to, for example, a handling robot, and the handling robot can accurately grip the round bar 1 based on this position data.

Note that in this example, the pattern similarity φ is set to the (6th)th
It was obtained from the formula, but the value in the denominator of formula (6) √ _o 〓 ⁱ⁼¹ (Cb′ _i ) ² ...(11) is a constant value related to the standard pattern, so if this is set as 1, In other words, the value obtained by (A″・B′) also corresponds to the pattern similarity, and this may be used as the pattern similarity.In this case, 2
If the two patterns perfectly match, the pattern similarity will be the value shown in equation (11), so it is necessary to take this into consideration when setting the threshold. Similarly, the pattern strength ψ was obtained from equation (9), but since the denominator of equation (9) is a constant value related to the standard pattern, only the numerator of equation (9), that is, (A′・B ') also corresponds to the pattern strength, so this may be used as the pattern strength. In addition, the average value calculation circuit 20
An appropriate area is established based on the average value of the standard pattern B obtained in the above two evaluation values. It may be added to the judgment based on (pattern similarity φ, pattern strength ψ).

As explained above, according to the present invention, the similarity of patterns is determined based on the two evaluation values of pattern similarity and pattern intensity. Even if the detection object changes, the brightness pattern related to the shape of the object to be detected can be accurately recognized.

[Brief explanation of drawings]

Figure 1 is a plan view of the round bar, Figure 2 is a graph showing an example of a standard pattern consisting of brightness in the transverse direction of the round bar, and Figure 3 is when the round bar in Figure 1 is scanned in the direction of arrow A. Graph showing an example of the brightness of
The figure shows the relative relationship between two patterns in n-dimensional space, Figures 5 a to e are graphs illustrating light and dark patterns, respectively, and Figures 6 a to e are graphs illustrating the 5
Graphs showing the patterns obtained by converting the light and dark patterns in Figures a to e around the average value, Figures 7 a to e are graphs showing the normalized patterns of the patterns in Figures 6 a to e, respectively, and Figure 8 A graph showing patterns for the standard model, Figure 9 shows pattern similarity φ and pattern strength ψ for each bright and dark pattern.
FIG. 10 is a block diagram showing an embodiment of an apparatus for carrying out the pattern recognition method according to the present invention. 1...Round bar, 10...ITV camera, 11...
A/D converter, 12...shift register, 13,
19...Subtractor, 14,20...Average value calculation circuit, 15,21...Normalization circuit, 16,17...
Arithmetic circuit, 18...Standard pattern setter, 22,
23...Judgment circuit.

Claims (1)

[Claims] 1. Extracting a brightness pattern A (C _a1 , C _a2 , ..., C _ao ) consisting of n pieces of image data from image data of a predetermined field of view in which an object to be detected is present; The average value _a is obtained from A using the following formula, _a = _o 〓 ⁱ⁼¹ C _ai /n, and the brightness pattern A is converted around the average value _a to form a pattern A' (C _a1 ', C _a2 , '..., C _ao ) is calculated using the following formula, C _ai ′=C _ai − _a , and further, a pattern that normalizes the pattern A′ is
A″(C _a1 ″, C _a2 ″,…, C _ao ″) is expressed as On the other hand, a standard pattern B (C _b1 , C _b2 , ..., C _bo ) consisting of image data showing n brightness levels related to the shape of the object to be detected is prepared in advance, and from this standard pattern B, the Find the pattern B′ transformed around the average value, and further normalize the pattern B′.
B'' is calculated, the inner product of the patterns A'' and B'' and the inner product of the patterns A' and B'' is calculated, and based on the two inner product values, the brightness pattern A is determined to be the shape of the object to be detected. A pattern recognition method characterized by recognizing whether or not there are related light and dark patterns. 2 Extract a brightness pattern A (C _a1 , C _a2 , ..., C _ao ) consisting of n pieces of image data from the image data of a predetermined field of view in which the object to be detected exists, and calculate the average value _a from the brightness pattern A. The pattern _{A ′} (C _a1 ′, C _a2 , ′..., C _ao ) obtained by converting the brightness pattern A around the average value C _a is obtained by the following formula, a = o 〓 ⁱ =1 C _ai /n. A pattern obtained by the formula, C _ai ′=C _ai − _a , and further normalized from the above pattern A′
A″(C _a1 ″, C _a2 ″,…, C _ao ″) is expressed as On the other hand, a standard pattern B (C _b1 , C _b2 , ..., C _bo ) consisting of image data showing n brightness levels related to the shape of the object to be detected is prepared in advance, and from this standard pattern B, the Find a pattern B' converted around the average value, find the inner product of the pattern A'' and B', and the inner product of the patterns A' and B', and calculate the brightness pattern A based on the two inner product values. A pattern recognition method characterized by recognizing whether or not the brightness pattern is related to the shape of the object to be detected.