JPS6346583A - Pattern recognizing method - Google Patents

Abstract

PURPOSE:To improve the position detecting accuracy of a body to be inspected by optimizing the area ratio between the low-level and high-level image parts of a reference pattern when this reference pattern is previously decided. CONSTITUTION:A TV camera 11 photographs a detecting pattern and sends it to an image processing circuit 12. The circuit 12 receives a reference binarizing level registered on a CPU 13 via a D/A converting circuit 14 and binarized the photographed image to store it in an image memory 15. When the reference binarizing level is decided, two intermediate value set when the area ratio between high and low levels of the image obtained by changing slightly the two binarizing levels having a difference has the minimum change. A correlation circuit 16 reads out coordinate data A on the stored binarized image to compare it with a reference pattern B registered to a pattern setting circuit 17 for calculation and storage of the degree of coincidence between both patterns. The circuit 16 obtains the variance of position between the position where the degree of coincidence is maximum and a reference position. Based on this variance of position, the position of the body to be inspected is detected.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a pattern recognition method suitable for use in semiconductor manufacturing equipment and the like.

[Prior Art] When wire bonding the joint of a semiconductor chip and the joint of a frame using a wire bonding device, in order to maintain the optimal positional relationship between the chip and the frame,
It is necessary to detect the misalignment of the chip. This positional shift is detected by a pattern recognition device.

The conventional pattern recognition method consists of: ■ A captured image of a chi, pu-shaped detection pattern taken by a television camera is converted into a binarized image by an image processing circuit at a standard binarization level. The binarized image is scanned, and each part of the 2-1 image is compared with a reference pattern that has been registered in advance in the pattern setting circuit. The degree of matching is calculated by a correlation circuit, and the position that most matches the reference pattern is found in both fi images, and the positional deviation from the Ii?ll! position is determined.

[Problems to be solved by the invention] However, in the conventional pattern recognition method, when the reference pattern is registered in the pattern setting circuit in advance, no
The setting of the area ratio between the low-level image portion and the high-level image portion in the quasi-pattern is left to the judgment of the operator. In such a case, the area ratio of the low-level image part and the high-level image part in the reference pattern deviates from the optimal state as described later, and the operator's This includes a degree of variation (individual differences).

The present invention optimizes the area ratio of the low-level image part and the high-level image part in the reference pattern when determining the reference pattern in advance12. The purpose is to improve the accuracy of detecting the position of a subject.

[-F step to solve the problem] The unreleased I51 converts the captured image of the subject into a binary image that has been binarized using I/Bell, and The image is compared with a reference butter] / determined by η, and by this comparison, 2 (ti images) (in a pattern recognition method that detects the position d of the subject based on the degree of matching with the pattern) , (Attempting to import Vero pattern), (Raw image), (After binarizing at the quasi-2-digitalization level to obtain the reference binarized image), (Checking the degree of matching of the quasi-2-digitalized image) 1Ii within the outer frame range where the ratio of the low-level image portion and the high-level image portion each accounts for approximately 50% within the image area surrounded by the outer frame indicated by l- by changing the outer frame range to be calculated.
The binarized image is made into an optimal J, (quasi-pattern).

[Operation] In the detection accuracy of the position of the subject based on the degree of coincidence between the binary image of the subject and the reference pattern, the obtained degree of coincidence is compared with the degree of reference coincidence to determine whether or not the detection is successful. The degree of conformance to this standard is determined by the detection environment and the condition of the subject! Due to the change in sensitivity, it is not possible to take a very large value, and from experience, for example, 70~
It is 80%. Therefore, the area ratio of the low-level image part and the high-level image part in the t, g * pattern is 20% to 80%.
%, a poem in which all the original binarized images of the subject compared to the reference pattern were high-level image parts has a matching degree of 80% with Ih 5. There is a risk of false detection.

Therefore, as in the present invention, by changing the outer frame range that determines the matching degree of the reference binarized image by 'r1, it is possible to easily increase the area ratio of the low-level image part and the high-level image part in the reference pattern by 50. % to 50%, optimization of the area ratio can be realized.

Therefore, according to the present invention, when determining the reference pattern, the area ratio of the low-level image part and the high-level image part in the basic pattern is set to h^, thereby improving the accuracy of detecting the position of the object. can do.

[Example] No. 114 is a block diagram showing the butter used for dispensing unreleased IIl]/recognition device 5/i, and FIG. 2 is the base? Listen to the method for determining the tI8 binarization level (schematic diagram, Figure 3). (Schemes 1 and 4 showing the state of changing the outer frame range of the quasi-pattern
It is.

In Figure 1, 11 is a TV camera, 12 is III
1 image processing + I'il path, 13 is a computer, 14 is [
)/A conversion circuit, 15 is a poor image seventy circuit, 16 is a correlation circuit,
Reference numeral 17 represents a pattern, an input circuit, and 18 a mask setting circuit.

First, a method for detecting the (& position) of a subject using this pattern recognition device will be described.

The television camera 11 photographs the detection pattern of the subject 1-, such as a half- and nine-body chip, and transmits the captured image (video signal) to the image processing circuit 12. Tano, (semi-24dti
The conversion level is applied via the D/A conversion circuit 14,
2 captured images are converted into 2 images at the standard binarization level (2
Make it a converted image. A binarized image is expressed as O (high-level image part: self-portrait part) or l (low-level image part: black image part) depending on the density of the captured image.For example, 128 bits in the X direction and 128 bits in the Y direction. The bit is specified and stored in the image memory 15.

1-The 21A image is scanned as follows.

That is, the correlation circuit 16 includes a memory in the image memory 15;
Has it been done? Data A at coordinates (0,0) of the valued image is
This thickness to read? Data AI day (0,0), for example, 32 bits, x32 bits, is binarized data at each point within the data. The correlation circuit 16 connects each point of data A manually entered as coordinates (0, 0) and the pattern setting circuit 1.
For example, 32 bits x 32 bits of reference pattern B are registered in advance in step 7, and the corresponding points are compared with each other, and whether or not they match, that is, the degree of matching, is calculated and stored in memory. At this time, if the obtained matching degree is lower than the predetermined reference matching degree, it is determined that the detection has failed, and the calculation result of the matching degree is ignored and not stored.

(The calculation of this degree of matching will be explained in detail later)
& To optimize the quasi-pattern, the level value (0 or 1) of each point in data A and the level value (0 or 1) of each point in reference pattern B
0 or 1).

By taking the logical product of the negative form of this exclusive OR and the level value (0 or 1) of each point of the reference mask degree C2 set in the mask setting circuit 18, data A and the optimal reference pattern are The degree of agreement will be calculated.

A concrete illustration of the above-mentioned matching degree calculation procedure is shown in Figure 4 (A
) to (E), and FIG. 1 is an example of 4 bits×4 bits. FIG. 4(A) shows data A, and FIG. 4(B) shows ノ, Ii? F! 4(C) shows the reference mask degree C2, and FIG. 4(D) shows the data A), 'l;? The negative form of the exclusive OR of tR pattern B is shown, and FIG. 4(E) shows the logical product of the negative form of the exclusive OR and the reference mask degree C2. The shaded areas in FIGS. 4(A), (B), (D), and (E) represent 1's as low-level image areas, and the non-shaded areas represent 0's as high-level image areas. Further, the shaded area in FIG. 4(C) represents 1 as an effective area, which will be described later, and the non-shaded area represents 0 as an invalid area. Here, no! As will be described in detail later, in the quasi-pattern B, the area ratio of the low-level image portion and the high-level image portion is set to 50%:50% by the semi-masking degree C2.

When the correlation circuit 16 completes the comparison of the data A at the coordinates (Q, 0), it reads out the data A at the coordinate 15 (1, 0) consisting of 32 bits x 32 bits shifted by 1 bit in the X direction, for example. Compare data A of (1, 0) with ノ, (monk pattern B. In this way, the correlation circuit 16
sequentially scans each piece of data from coordinates (0, 0) to coordinates (θ5, 95) and calculates their matching degree. When the correlation circuit 16 finishes calculating the degree of coincidence over the entire range of the binarized images stored in the image memory 15, it selects the position with the highest degree of coincidence stored in its own memory, and calculates the degree of coincidence from the reference position. The positional deviation (!, ?) is determined, and the position of the object is detected based on this.

By the way, the reference binarization level registered in the computer 13 is set when the reference pattern is predetermined.
When a captured image of a subject is binarized, it is necessary to be able to always stably convert the captured image into a z-value, which may change depending on the shooting conditions and the like.

Therefore, in this embodiment, when the LIi image processing circuit 12 binarizes the reference image 20 as shown in FIG. 2 into which the reference pattern is to be captured, the computer
As the binarization level for 3, use the binarization level a and the binarization level b which have a level difference of Δ or more, and move the levels a and b little by little within the range of l - intersection 2. and change it. This gives the group 7 for level a
1.142 (for example, the area ratio of the high level image part is 40%, the area ratio of the low level image part is BO%) and the reference binarized image obtained for level b (for example, the area of the high level image R part) (30% area ratio, 70% area ratio of low-level image area)
The intermediate value between a and b at which the change in the area ratio of the high-level image part and the low-level image part: for example, 10% is the minimum is set as the appropriate semi-binarization level.

That is, according to this embodiment, the basic pattern is determined from 1 to 1], (the quasi-binarization level is converted to Iik A, so the captured image of the subject is always Stable binarization can improve the accuracy of detecting the position of the object.

In addition, when calculating the matching degree between the data A read from the binarized image of the subject and ), '; The degree of matching obtained is compared with the standard matching degree to determine whether the detection is successful or not. Since the degree of conformity with the standard is subject to changes in the detection environment and the state of the subject, it is not possible to obtain a very large degree of accuracy, and empirically it is, for example, 70 to 80%. Therefore, the area ratio of the high-level image part and the low-level image part in the reference pattern is 80% to 2.
Under the condition that 1 is set to 0%, if all the original binarized images of the subject compared with the reference pattern are high-level image parts, there will be a matching degree of 80% at that point. There is a risk of false detection.

Therefore, in this embodiment, the group ? F! The image is binarized in the image processing circuit 12 at the standard binarization level 1-.

After obtaining the reference binarized image, the outer frame range for calculating the matching degree of the reference binarized image is changed from the initial outer frame indicated by Zl to the range indicated by 22 at the 31st ffl. As a result, the high-level image portion (
I)) and the low-level image portion (L) can each account for approximately 50%, and the correlation circuit 16 uses the basic image surrounded by the outer frame Z2 as the optimal pattern.
shall be used to calculate the degree of match.

Note that the specific way to change the outer frame range of the quasi-binarized image is determined by the mask setting circuit 18, for example k, 32, etc.)
Set the outer periphery of the initial outer frame Zl of X32 bits to 0 (invalid area)
, the central part is set to 1 (A1 effective area), and the mask degree C of the outer frame Z is used. That is, the correlation circuit 16 calculates ('l)
The level value (0 or 1) of each point of the data A read from the reference z-valued image obtained by the image processing circuit 12 and the pattern setting circuit 171. For example, 321-
After taking the exclusive OR of the level value (1) of each point of the overall high level pattern BO in which all points of y x 32 bits are high level image parts, ■ the negative form of this exclusive OR, The level value of each point of the above mask degree C (0 for the outer part of the outer frame Z,
Take the logical AND of the inner part. As a result, correlation circuit 16
It is possible to calculate the respective occupancy ratios of the high-level image portion and the low-level image portion within the image area surrounded by Z with a certain range of outside: 1 set to 0 (invalid area). Therefore, if we gradually expand or contract the outer range of the outer frame Z to be changed to 0 (invalid area), -
The masking degree of the outer frame Z2 with a 1-foot area ratio of approximately 50%, that is, the reference masking degree C2 can be determined. That is, in this embodiment, the reference pattern set in the pattern setting circuit 17 is optimized by adding the reference mask degree C2 set in the mask setting circuit 18.

However, in the implementation of 12 and 11, it is also possible to attach the reference mask 11 C2 mentioned above, fix the reference pattern as the optimum reference pattern, and set this optimum reference pattern in the pattern setting circuit 17. .

That is, according to this embodiment, when the reference pattern is predetermined, the area ratio of the low-level image part and the high-level image part in the reference pattern can be easily set to 50% to 50% by adjusting the j constant 1.
- Since the reference pattern is optimized, the pass/fail f4 determination of the detection is ensured, and the accuracy of the position detection of the object;
T:Kill.

[Effects of the Invention] As described above, according to the present invention, it is possible to easily optimize the area ratio of the low-level image portion and the high-level image portion in the reference pattern by predetermining the reference pattern. , it is possible to improve the accuracy of detecting the position of the subject.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a pattern recognition device used to implement the present invention, FIG. 2 is a schematic diagram showing a method for determining a reference binarization level, and FIG. 3 is a diagram showing a method for determining a standard binarization level. Schematic diagram showing J-shaped islands, Figures 4 (A) to (E)
FIG. 2 is a schematic diagram showing a procedure for calculating the degree of matching. 12... Image processing circuit, 15... Image memory, 16
... Correlation circuit, 17... Pattern setting circuit, 18.
...Mask setting circuit, 20...Reference image, A...2
Valued image data, B... reference pattern, C2...
Standard mask degree, L...low level image part, H...high level image part, 21.22-...outer frame, a, b-2 image 1
Chemical 1/Bell. Agent patent attorney + 11! Shuji Kawa 2nd Figure □ Words 3 Figure 4tii2] (A) (B) Data A
Standard pattern date (C)
(D) Reference mask MC2AXOR day (E) AXORB AND C2

Claims (1)

[Claims] (1) A captured image of the subject is converted into a binarized image at a reference binarization level, and this binarized image is compared with a predetermined reference pattern, and the binarized image and the reference pattern are obtained by this comparison. In a pattern recognition method that detects the position of a subject based on the degree of matching with a reference pattern, a reference image in which a reference pattern is to be captured is binarized at a reference binarization level to obtain a reference binarized image, and then The outer frame range for calculating the matching degree of the binarized image is changed, and within the outer frame range where the ratio of the low-level image part and the high-level image part each accounts for approximately 50% within the image area surrounded by the above outer frame. A pattern recognition method characterized in that a reference binarized image of is used as an optimal reference pattern.