JPS5952472B2 - Part identification method - Google Patents

Description

[Detailed description of the invention] The present invention relates to a component identification method.

In order to automatically assemble parts, it is first necessary to identify the parts.

An example in which part identification is performed as part orientation identification will be described below. 1 to 4 show various postures in which the component 1 is pressed against a wall 2. FIG. Figure posture I, 1
As a method for identifying 1, 111, and IV, first, a method for performing identification using a plurality of photoelectric switches is shown in FIG. Regarding postures I, 11, 111, and IV, photoelectric switches 6 and 7 (actually as many as necessary for identification) are provided at positions where each posture can be identified. Now, if transmitted illumination is being used and photoelectric switches 6 and 7 are turned on when light enters and turned off when light is blocked, photoelectric switch 6 is turned on in postures I and IIL.
The photoelectric switch 7 is turned ON in postures 11 and IV. Therefore, by checking the combination of ON and OFF of the photoelectric switches 6 and 7, it is possible to identify the postures I, 11, 111, and IV. However, with this method, after a theoretical study of the position and number of photoelectric switches, the positions had to be finely adjusted using the actual device. Therefore, as the number of postures increases and the number of ON-OFF combination patterns of each photoelectric switch increases, more time will be required for desk study and adjustment. Furthermore, each switch is placed in the appropriate position and in the required number, so if you are handling different types of parts, you will have to adjust the position of each switch each time you change the type of part, and you will have to adjust the position of each switch each time you change the type of part. cannot be identified. In order to solve the above problems, ITV
There is a method using an imaging device such as a camera or a solid-state imaging camera.

According to this method, as shown in FIG. 6, the imaging device divides the image into a finite number of fine pixels 8 and converts the amount of light of each pixel into an amount of electricity, so an appropriate threshold value is set. If they are installed and each pixel is binarized, it will have the same effect as if fine photoelectric switches were lined up across the screen without any gaps. Conventionally, a simple pattern matching method has been used as a component identification method using such an imaging device.

This is done by memorizing the information of all the pixels of the entire screen or of a specific continuous part for each of the possible postures of the part, and preparing the same number of dictionary patterns as the number of possible postures. In this method, all pixels corresponding to each of all dictionary patterns are found to match the part pattern with a certain orientation obtained in the database, and the orientation of the dictionary pattern with the highest degree of matching is determined to be the same as the orientation at that time. Therefore, in this method, by preparing a number of dictionary patterns corresponding to types and orientations, it is possible to identify various types of parts continuously without any setup time. However, it has the disadvantage that it requires a large storage capacity to store information on a very large number of pixels as a dictionary pattern, and it takes a lot of processing time to determine the degree of matching. An object of the present invention is to provide a component identification method that is versatile and quick in calculation processing.

The gist of the present invention is to extract data for posture identification based on video information of parts. This data is called a reference point.
A dictionary pattern is determined based on the reference points thus obtained. Parts are identified by comparing this dictionary pattern with real reference points obtained from actual part video information. At this time, in addition to the orientation of the part, it is also possible to identify specific parts, such as the type of the part. Hereinafter, the present invention will be explained in detail with reference to the drawings. FIG. 7 is a diagram showing an embodiment of the present invention.

This embodiment includes an imaging device 9, an identification device 10, a display device, and so on.
11. It consists of a control device 12. The control device 12 controls the posture of the parts, etc. according to the identification results of the identification device 10. The identification device 10 includes a dictionary pattern creation section 13,
It consists of a pattern comparison and identification section 14. The key point of the invention in the above embodiments lies in the processing content of the identification device 10.

The operation of the identification device 10 will be explained below. First, the creation of a reference point pattern will be described. Figure 8 shows postures 1 and 11 of part 1.
, 111, shows the superimposed video information for IV. This superimposed video information 15 forms a partial region 16 closed by the boundary line of each posture. In the case of component 1 shown in the figure, 19 partial regions 16 are formed. For each of the plurality of partial regions 16 described above, one pixel within the region is set as a reference point. An example of setting the reference point is shown in the 9th section.
As shown in the figure. In Figure 9, ten reference points 1 are used as reference points.
An example is shown in which 7, 18..., 25, 26 are set. However, the above reference points 17, 18, ・
It is assumed that each of 25 and 26 occupies one complete pixel (square) within the partial area 16. Therefore, reference points are not set for partial areas for which a reference point of one pixel cannot be set. Furthermore, in a partial area where there are many pixels that can be used as reference points in the same area, as shown in FIG. 10, set Y=YO to X. Scan from to X, then set Y=YO+1 to X. Scan from to XF. Thereafter, similar scanning is repeated up to YF to scan the entire screen. In this scanning process, the reference point earliest in the reference order for the corresponding partial area is set as the reference point for the corresponding partial area.Of course, this scanning is performed in the same way for other partial areas.In any case, Reference points are set according to the above scanning.Next, reference points are selected.

FIG. 11 shows the selection process using a tri-structure. In the case of FIG. 9, the reference point that is referred to earliest in the scanning process is the reference point 17, so in FIG. 11, postures 1 to 1V are first classified according to the pixel information of the reference point 17. Pixel 8 in FIG. 9 is binarized, and if it is set to "1" when there is a part and "0" when there is no part, the posture 1-1V is determined by "17" and "" of the reference point 17. It is then divided into two parts as shown in Figure 11. The next reference point will be reference point 18. Reference point 18
Since the pixel information of is the same as that of reference point 17, reference point 18 is not adopted as a reference point. The criteria for selecting and discarding reference points is that if it provides information that separates even one pose from a subgroup of poses, it will be included, and if it does not, it will be rejected and not included. . for example,
The reference point 17 and the reference point 23 are used to change the orientations 1 to IV to orientations 1, 11, 111, and 111, even though the pixel information "RO" is opposite.
Since they bring about the same information that is divided into two parts, if one is adopted, the other is not adopted.

By examining the reference points one by one until they are separated from all postures using the above criteria, posture 1-1 can be identified by reference points 17, 19, and 20.

At this time, the dictionary pattern that combines the reference points 17, 19, and 20 corresponding to each posture is as shown in Table 1. At this time,
The problem is the posture 1V.

In posture 1, reference points 17, 19, and 20 are all 4'', so the judgment is made as if there were no parts. In order to distinguish between this no-parts state and posture 1V, information about the presence or absence of parts for posture 1 must be obtained from another source. At this time, if we consider that even if there is no part, it is one posture, all postures including the reference point 22 will be identified, and the dictionary pattern will contain information about the reference point 22 as shown in Table 1. will be added. In this way, dictionary patterns are stored in the memory within the dictionary pattern creation section 13 and are created. The operation then moves on to the actual comparison of the patterns.

In this comparison, the pattern comparison section 14 compares the dictionary pattern and the actual video information. At this time, a search is performed to see if there is actual video information regarding the reference point of the dictionary pattern, and if there is video information at the relevant reference point, identification can be made according to that reference point. The identification device may generally be dedicated hardware, but may also be a general-purpose computer. A more detailed explanation will be given below based on an example of a general-purpose computer. Determination of reference points and creation of dictionary patterns will be described in detail below.

Stage 1.

To find pixels that cannot be referenced. Non-referenceable pixels are pixels where the outline of the part touches the threshold value (or the amount of reflected light changes)
If there is, it is a pixel for which it is unclear whether the binarized value will be '' or 'r'.

FIG. 12 shows non-referenceable pixels for posture 1.
In the figure, pixel parts 30, 31, 32, 33 indicated by diagonal lines
is a non-referenceable pixel. Next, the calculation of the non-referenceable pixels will be described. Hereinafter, the fluctuations in the signal and the threshold value will be expressed as relative amounts, and only the fluctuations in the threshold value will be expressed, and the amount of fluctuation in the threshold value will be referred to as ΔT.

In order to repeatedly obtain the same information ("or n") despite the amount of variation ΔT, the level difference between the signal and the threshold must be
must be larger than To this end, pixels that provide uncertain information are searched for in the following procedure and are not referenced. Let TM be the threshold value used for work.

This has a variation of ΔT. During learning, a higher threshold TH and a lower threshold TL than TM are used. However, TH〉
TM+ΔT. T, <TM−ΔT. First, the threshold T,
The image is binarized by At this time, it must be considered that TL is also changing, but it can be said that a pixel whose binarized value is "O" will always become "0" when binarized using the threshold TM. Next, the image is binarized using a threshold value TH. At this time, when the pixel that is n'' is binarized with TM, it will always be ``1''.
become. In other words, it is unreliable whether it is '07 or N7.' (Well, it is a pixel that is 1 by T and '' by TH.

This is assumed to be a non-referenceable pixel. Stage Ding. To memorize pixel information. The actual procedure for storing this pixel information is as follows:
Simultaneous with stage 1.

That is, the information of each pixel ("" or "1") binarized by the threshold T, (or TH) is stored as a binary image in a certain posture.Next, the above steps 1. Let's specifically explain the processing procedure at Ding.

First, one screen worth of binary image based on the threshold value TL is stored in an image capture memory. FIG. 13 shows the data state of the image capture memory 50. Pixel positions are set in a matrix in the horizontal and vertical directions. The pixel information in the memory 50 is copied to the pattern storage memory 51 as shown in FIG. In the pattern storage memory 51, pixel numbers (positions) are set in the vertical direction, and data of the non-reference mark RUM, postures 1, 11, 111, and IV are set in the horizontal direction. Next, the binary image based on the threshold value TH is captured and stored in the image capture memory 50 in the format shown in FIG. 13 (however, at this time, the binary image based on the threshold value TL is cleared). This is compared with the binary image based on the threshold value T in the pattern storage memory 51 for each pixel, and if there is a difference, the non-reference mark bit is set to "1". However, in the above processing procedure, either TL or TH may be performed first. Stage 2.

Step 1: Repeat the operation of the knife 5 times for each position. As a result, the contents of the pattern storage memory 51 are, for example,
The result will be as shown in Figure 15.

Figure 16 shows this result using an actual example.
The shaded areas are pixels that cannot be referenced. Stage 3.

Consider each pixel in reference order. Now, the reference order is (1, 1), ...... (1, 25),
(2, 1), ...... (2, 25), (3
,1)・・・・・・(3,25),・・・・・・
It shall be something like...

However, this five reference order is an example. The rules (algorithm) to be considered are (a) Unreferenceable pixels RUM
will not be considered.

(b) For all postures, the pixel that takes ``, for example, the first
Pixel (1,1) in Figure 5 and all pixels with n'' are not considered.

(c) The same pattern as the pixel used once, for example, the first pixel.
The pixels having pixels (1, 10) and (1, 11) in FIG. 5 will not be considered hereinafter.

shall be.

Next, the detection procedure will be specifically explained.

Check whether there is a mark that cannot be referenced. If it is, move on to the next pixel. If there is no non-referenceable mark, check whether all bits in the pattern storage memory 51 (other than the non-referenceable mark) are "O" or 1. If so, move on to the next pixel. Yes. If not, check whether it is the same as the pixel previously adopted as a reference point. If so, move on to the next pixel. If not, move on to the next process. The above processing process is repeated in step 15. According to the example shown in Fig. 16, pixels (1, 1) ~
(1, 9) is skipped and processed, and the pixel (1, 10
).

Also, in this case, there are no pixels adopted as reference points yet,
Pixel (1,10) is the first reference. A memory that stores the contents of a pixel that has been referenced once is referred to as a reference storage memory.

It is assumed that all bits of this memory are set to 07 in the initial state. A posture classification memory is required to know how this pixel (1, 10) classifies posture 1-1. In the initial state, all bits of this posture classification memory are 11'. That is, when all bits are 17, this indicates that postures 1-1V are not classified at all.
Next, A between the posture classification memory and the contents of pixel (1, 10)
Take the ND and write it into working memory.

This process applies pixels (
This is a process to find out what kind of information the contents of 1 and 10) bring. Specifically, 1 of pixel (1, 10)
100 and 1111 in the posture classification memory are ANDed and the result is written into the working memory. The data written to the working memory is 1100. Next, a check is made to see if there is any difference among the bits in the working memory that correspond to the bits set to "1" in the posture classification memory.

If there is none, move on to the next pixel. If there is something different,
Moving on to the next process, it is adopted as a reference point. In the above example, the pose classification memory contains N bits from bit 1 to bit 4.
7 is set, so bits 1 to 4 of the working memory are compared to see if there are any differences. In this case, bits 3 and 4 are 0 and bits 1 and 2 are 1. When adopted as a reference point, pixels (1, 10
) is written and the pixel (1,
10) Write the contents.

The data structure of the reference storage memory 52 and dictionary pattern memory 53 is shown in FIG. Then, the first reference point is set as (1, 10). Next, pixel (1, 1
0) to find out how the pose groups were classified.

Therefore, as the next posture classification memory, a working memory and a memory in which ``1'' and 0'' of the working memory are reversed are used.
Use one. As a specific example, in the case of pixel (1, 10), there are two posture classification memories, 1100 and 0011. 1100 indicates that postures 1 and 11 have not been classified yet, and 0011 indicates that postures 111 and I have not yet been classified.

Let's move on to consider the next pixel.

The next pixel is (1, 11), but this is (1, 10)
It is not adopted because it is the same as . Hereinafter, the pixel (2, 6) will be considered as the next reference point candidate. Pixel (
The contents of 2 and 6) are 1110. This data 1110
A for each of the above two posture classification memories.
Take the ND and put it in working memory. The result in the working memory is 110 for the data 1100 in the posture classification memory.
0, and data 0011 becomes 0010. Among the two data 1100 and 0010 in this working memory, ゛ and 11100 are indistinguishable from postures 1 and 11, so
It is not useful for classifying postures 1 and 11. Data 0010 allows posture 111 to be classified. Therefore, pixels (2,
6) is adopted as a reference point. As a result, there are two reference storage memories, pixels (1, 10) and (2, 6).
Furthermore, the contents of pixel (2, 6) are written into the dictionary pattern memory. If there is no new information for posture classification, the posture classification memory remains as it is, and if posture classification is done, the working memory and the previous posture classification memory are set. Bit N7 is the inverted version of ``''.

Therefore, in the above example, the content 1100 of the pose classification memory
is carried over as is, and the content 0011 is changed to 0010 and 000.
It becomes 1. If there is only one "17" in the pose classification memory, it means that the pose corresponding to that bit has been classified, so the AND with the pixel contents will not be performed from now on.In the current example, the pose 111??::.I is pixel (1, 1
0) and (2, 6).
Next, the same process is repeated and pixel (3, 3) becomes the next reference point candidate.

The content of pixel (3, 3) is 0110. This pixel (3
, 3) contents 0110 and contents 1 of the three posture classification memories
AN for each of 100, 0010, 0001
Take D. However, among the above data, the two data 0010 and 0001 described later are excluded from the AND condition in the above case. As a result, 0100 is obtained in the working memory by the AND condition of 01110 and 1100. Is this data posture 1? ::． 11, so it is used as a reference point. At that time, pixels (3, 3
). Furthermore, pixels (
3. Write the contents of 3). Then, pixel (3, 3) is selected as the third reference point. The dictionary pattern storage memory 53 at this time is shown in FIG. Furthermore, the posture division memory 54 stores posture 1-1 in accordance with the above steps as shown in FIG.
A dictionary pattern is obtained that indicates which pose corresponds to a pattern in which the pixels necessary for classifying the pixels and their pixel information are arranged.

One problem here is that the dictionary pattern corresponding to posture 1 is 0000.

This is because the binarized information for every pixel is ``'', so it is difficult to distinguish this from the case where the component is not within the imaging field of view. There is no problem if the presence or absence of the component can be detected by other detection means (for example, a micro switch, etc.), but if this is not possible or if confirmation is desired, the following procedure is additionally required. Stage 4.

Checks whether all bits in the dictionary pattern are zero. When there is a pixel in which all bits are N07, one pixel that always has n'' is selected and used as a reference point.

In the above example, the dictionary pattern corresponding to posture 1 is 00.
0, so when the attitude is 1V, select a pixel that always becomes "r". Among these, the earliest in the reference order is pixel (6, 2).
Therefore, this is set as the fourth reference point. An actual example of the resulting dictionary pattern storage memory 53 is shown in FIG. Next, an example of implementing the above processing using a flowchart will be briefly described.

This processing flowchart can also be realized by a computer. 21 to 27 show the overall flowchart. Various variables in this flowchart are listed below. Incidentally, FIGS. 29 to 31 are explanatory diagrams of various variables at that time. ○Memory BCi. j〕: Image capture memory (Buffer
MemOry), row i and column j.

1 bit 1: Maximum number of rows J: Column ″P(1.j): Pixel (1.j) in pattern storage memory.
The pattern taken by j).

N bit P [(1.j), n]: nth bit of P(1.j)
Bit.

1 bit N: Number of postures R(q): qth reference storage memory l).

N bit Q: Number of reference storage memories (=number of reference points)
S(1): 1st posture classification memory, N bits L: Number of posture classification memories W(1): lth work memo l). N
BIT WCl. k]: k-th bit of W(1).

1 bit SS (1), WW (1): S (1), W (1)
Memory used for temporary storage D(n): Dictionary pattern memory corresponding to posture n Q bits DCn. Q] :D
Q-th bit of (n).

1 bit 1 (Q): Row number of Q-th reference point (reference pixel).

J(Q): Number of columns of the Q-th reference point (reference pixel). O flag FCi. j]: Indicates that the pixel (1.j) cannot be referenced.

1 bit FFCl]: Indicates that the l-th posture classification memory is divided.

1 bit.

(Indicates that the P(1.j) referenced at that time brings new information to the pose classification.) By automatically setting a dictionary pattern, a dictionary pattern can be set, and by comparing this dictionary pattern with real values, it has become possible to identify the shape and orientation of a part.

[Brief explanation of the drawing]

1 to 4 are diagrams showing various postures of parts, FIG. 5 is a diagram showing detection examples for various postures, FIG. 6 is a diagram showing pixel examples, and FIG. 7 is a diagram showing an embodiment of the present invention. Figures 8 and 9 are explanatory diagrams of reference points, Figure 10 is an image scanning example diagram, and Figure 11 is an illustration of reference points.
The figure is an explanatory diagram of the dictionary pattern, Figure 12 is a diagram showing an example of boundary area exclusion, Figures 13, 14, 15, and 17.
Figures 8, 9, and 20 are data configuration diagrams of various memories, and Figure 16 is a diagram of specific pattern examples.
Figures 21 to 27 are process flowcharts, Figure 28
31 to 31 are explanatory diagrams of various variables in the above flowchart. 9... Imaging device, 10... Identification device,
11...Display device, 12...Control device, 13...Dictionary pattern creation section, 14...
...Pattern comparison section.

Claims (1)

[Claims] 1 In a parts identification method in which the type and orientation of the positioned part is imaged and the video signal obtained by the imaging is used to identify the type and orientation of the part, all of the parts to be identified are Reference point candidates are set within a plurality of closed partial regions formed by the outline of the component in a superimposed image obtained by superimposing all video signals of the component and the background for the posture, and the reference point candidates are Among them, those that do not provide the same information regarding the identification of parts are selected as reference points, the video information of the selected reference points is combined to create a dictionary pattern, and the obtained dictionary pattern and the above reference points are combined. A component identification method that identifies the type and orientation of a component by comparing a combination pattern of real video information with the dictionary pattern.