JP2022012668A - Inspection system - Google Patents

Abstract

To provide an inspection system capable of improving inspection accuracy of an inspection object.SOLUTION: An inspection system 10 includes a camera 11 and an inspection device 100. The inspection device 100 generates difference image data GD from inspection image data GK acquired from the camera 11 and output image data GS generated by a learned model that is an automatic encoder. Further, the inspection device 100 generates weight image data GW in which weight is set based on distribution image data GV reflecting a variance value of a pixel value of a pixel using a large amount of normal product image data GN. Then, the inspection device 100 generates determination image data GJ by multiplying the difference image data GD and the weight image data GW to calculate a determination value based on the determination image data GJ and determine whether an inspection object is good or bad.SELECTED DRAWING: Figure 4

Description

The present invention relates to an inspection system.

Generally, a product produced through a production line of a factory is inspected for appearance and the like in an inspection process before being shipped. In recent years, automation using image processing has been promoted in the inspection process of production lines. When promoting automation, inspection systems that utilize machine learning are being actively researched and developed.

For example, an inspection system disclosed in Patent Document 1 has been conventionally known. The conventional inspection system reads the image data of the normal product and performs machine learning using the autoencoder to generate a trained model that accurately reconstructs only the image data of the normal product. Then, in the conventional inspection system, the inspection image data of the inspection object and the output image data reconstructed by inputting the inspection image data into the trained model can be compared and regarded as the same. When the range is exceeded, it is determined that the inspection target is a defective product.

International Publication No. 2020/031984

By the way, in order to accurately generate a trained model using an autoencoder, it is generally necessary to read and train a large amount of trained image data groups consisting of a large amount of normal product image data. In this case, although it is preferable that all the parts for imaging the normal products are the same for the learning image data group, a slight positional deviation (posture deviation) usually occurs for each normal product. Then, in an image (image data) in which such a position shift (posture shift) occurs, for example, a peripheral portion of a normal product may be reflected.

Here, since the minute position shift (posture shift) is not constant, the position of the peripheral portion changes for each image. For minute misalignment (posture misalignment), it is conceivable to perform position correction using template matching or the like, but it is not realistic to make the correction error zero. As a result, when learning is performed using the learning image data in which a minute positional deviation (posture deviation) occurs, it becomes extremely difficult for the trained model to accurately and clearly reconstruct the peripheral portion of the normal product. Therefore, when the inspection image data is compared with the output image data reconstructed by the trained model, the peripheral portion of the output image data is not accurately reconstructed, so that the peripheral edge between the inspection image data and the output image data is not accurately reconstructed. The difference between the parts becomes large. For this reason, it becomes difficult to accurately determine the quality of the product, which may reduce the inspection accuracy of the inspection object.

An object of the present invention is to provide an inspection system capable of improving the inspection accuracy of an inspection object.

The inspection system includes an imaging device that captures an image of the inspection target and an inspection device that determines the quality of the inspection target based on the image of the inspection target captured by the imaging device, and the inspection device is normal. A model storage unit that stores a trained model, which is an auto encoder that outputs output image data that represents the output image of a normal product, using a group of image data for learning that has a plurality of normal product image data that represent the image of the product as input data. Inspection image data using the inspection image data acquisition unit that acquires inspection image data representing the image of the inspection target, and the output image data output by inputting the inspection image data and inspection image data to the trained model. The basic statistics are calculated for the characteristic values of all the pixels of each normal product image data in the learning image data group and the difference image generation unit that generates the difference image showing the difference between the output image data and the output image data. Based on the distribution image generation unit that generates a distribution image that represents the distribution of characteristic values based on, and the distribution image data that represents the distribution image generated by the distribution image generation unit, weights are set according to the basic statistics, and the weights are set. A weight image generation unit that generates a weight image representing the above, a judgment value calculation unit that calculates a judgment value using the difference image data representing the difference image and the weight image data representing the weight image, and a judgment value calculation unit based on the judgment value. It is provided with a quality determination unit for determining the quality of the inspection target.

According to this, when inspecting an inspection object using a trained model that is an autoencoder, a weighted image based on the difference image data generated by using the trained model and the distributed image data representing the distributed image. The determination value can be calculated using the data. As a result, the judgment value can be calculated in consideration of the region where the basic statistic is large, that is, the region where the accuracy of the quality judgment may be lowered. Therefore, the quality of the inspection object is good or bad based on the calculated judgment value. Judgment, that is, inspection accuracy can be improved.

It is a figure which shows the structure of the electric power steering apparatus. It is a figure which shows the capsule which is an example of the inspection object constituting the support device provided in the column device of FIG. It is a figure which shows the structure of an inspection system. It is a figure which shows the functional block composition of a control device. It is a figure which shows the functional block composition of the difference image generation part of FIG. It is a figure which represented the trained model schematically. It is a figure which shows the image of a defective product among the images taken by a camera. It is a figure which shows the output image reconstructed by a trained model. It is a figure which shows the difference image of a defective product. It is a figure which shows the difference image of a normal product. It is a figure which shows the distribution image. It is a figure which shows the weight image. It is a figure which shows the determination image of a defective product. It is a figure which shows the judgment image of a normal product.

(1. Outline of inspection system)
The inspection system inputs the inspection image data of the inspection target to the trained model, which is an autoencoder generated based on the image data of the normal product of the inspection target, and the output image data obtained by inputting the inspection image data and the inspection image. It is a system that inspects using data. Specifically, the inspection system uses a trained model, which is an autoencoder that outputs output image data when the inspection image data is used as input data, and evaluates the quality of the inspection object in the inspection process.

Here, the inspection system sets the product produced through each processing process of the production line as the inspection target. Then, the inspection system of this example determines the quality by determining the presence or absence of scratches, cracks, etc. that are unintentionally generated in the processing process based on the image data. In this example, a capsule constituting the support mechanism of the electric power steering device is exemplified as an inspection target, and the presence or absence of scratches on the capsule is determined and the quality is inspected.

(2. Outline of electric power steering device T)
As shown in FIG. 1, the electric power steering device T (hereinafter referred to as “EPS device T”) includes a steering mechanism 2 that steers the steering wheel W based on the operation of the steering wheel 1 by the driver. .. Further, the EPS device T includes an assist mechanism 3 that assists the driver in operating the steering wheel 1, and a column device 4 that supports a part of the steering mechanism 2 on a mounting stay K provided on a vehicle body (not shown). .. Here, the capsule, which is the inspection target of this example, is provided in the support mechanism 5 of the column device 4. Since the steering wheel 1, the steering mechanism 2, the assist mechanism 3, the column device 4, and the support mechanism 5 can use well-known configurations, detailed configurations and operation thereof will be omitted.

The column device 4 is supported by the mounting stay K of the vehicle body via the support mechanism 5. The support mechanism 5 includes an upper bracket 5A connected to the mounting stay K and a column side bracket 5B fixed to the column device 4. The support mechanism 5 includes a capsule 6 arranged between the upper bracket 5A and the column side bracket 5B. In this example, the capsule 6 is an object to be inspected.

As shown in FIG. 2, the capsule 6 is formed in a flat plate shape and includes an insertion hole 6a through which a bolt 7 for assembling the support mechanism 5 to the mounting stay K is inserted. Further, the capsule 6 is fixed to the upper bracket 5A and the column side bracket 5B by a resin pin or the like (not shown).

Here, the support mechanism 5 exhibits a function of absorbing the load (impact load) input to the steering wheel 1 by fixing the capsule 6 to the upper bracket 5A and the column side bracket 5B by a resin pin or the like. Can be done. That is, when an impact load of a predetermined load or more acts on the support mechanism 5 via the steering wheel 1, for example, a resin pin or the like for fixing the upper bracket 5A and the capsule 6 is sheared along the axial direction of the column device 4. Ru. As a result, the upper bracket 5A can move relative to the vehicle body in the front direction of the vehicle. Therefore, the support mechanism 5 allows the column device 4 to move in the vehicle front direction, and can reduce the impact load input to the steering wheel 1.

As described above, the support mechanism 5 is provided with the capsule 6 so that the support mechanism 5 can be operated so as to reduce the impact load. In this case, the capsule 6 needs to have enough strength to surely shear the resin pin or the like under an impact load of a predetermined load or more. In order for the capsule 6 to have the required strength, it is important that defects such as scratches and cracks have not occurred. The capsule 6 is fixed to the column side bracket 5B, for example, and is lined by the inspection system 10. The above is inspected for defects.

(3. Configuration of inspection system 10)
Next, the configuration of the inspection system 10 will be described. In the inspection system 10 in this example, the capsule 6 which is a component of the EPS device T is the inspection target. Then, in the inspection system 10 of this example, when the capsule 6 is assembled to the column side bracket 5B and transported to the next process in the production line of the EPS device T, that is, before the upper bracket 5A is assembled, the capsule 6 is assembled. The inspection process of imaging and inspecting is set. As shown in FIG. 3, the inspection system 10 includes a camera 11 as an image pickup device and an inspection device 100. The camera 11 mainly captures the inspection target surface S of the capsule 6. Then, the camera 11 outputs the image data G of the imaged inspection target surface S to the inspection device 100.

The inspection device 100 determines the presence or absence of scratches, cracks, or the like on the inspection target surface S of the capsule 6 by using the image data of the inspection target surface S (see FIG. 2) of the capsule 6 captured by the camera 11. .. Then, the inspection device 100 determines that the capsule 6 having no scratches or cracks on the inspection target surface S is a normal product, and determines that the capsule 6 having scratches or cracks on the inspection target surface S is a defective product. ..

By the way, when the inspection target surface S of the capsule 6 is imaged by the camera 11, a slight positional deviation (posture deviation) of the capsule 6 may occur. In this case, for example, in the image captured by the reflection of the illumination light, the edge portion which is a peripheral portion having a pixel value larger than the pixel value which is the characteristic value of the pixel representing the inspection target surface S of the capsule 6. S1 and the column side bracket 5B may be imaged (see FIG. 2). Here, the pixel value is a value representing the shade and brightness of the color of each pixel, and in a black-and-white inspection image, the shade of each pixel is set to a plurality of gradations (for example, 256 using an integer from 0 to 255). It is a value representing (gradation, etc.). The larger the pixel value, the whiter the pixel, and the smaller the pixel value, the blacker the pixel. As the characteristic value of the pixel, for example, a luminance value representing the brightness of the pixel can be exemplified in addition to the pixel value.

(4. Details of inspection device 100)
Therefore, in the inspection step, even though the capsule 6 is a normal product, the reflected edge portion S1 may be erroneously determined as a scratch, a crack, or the like. In order to prevent such an erroneous determination, the inspection device 100 includes a difference image generation unit 110, a distribution image generation unit 120, a weight image generation unit 130, a determination value calculation unit 140, and a determination value calculation unit 140, as shown in FIG. It mainly includes a pass / fail determination unit 150.

The difference image generation unit 110 outputs the difference image data GD representing the difference image by using the autoencoder which is a trained model. The difference image represents the difference between the inspection image obtained by capturing the inspection target surface S of the capsule 6 which is the inspection target and the output image reconstructed by inputting the inspection image data GK representing the inspection image to the auto encoder. It is an image.

Here, the auto encoder used by the difference image generation unit 110 generates only the output image of the normal product as output data when the normal product image data GN of the normal product captured in large quantities by the camera 11 is input as input data. And output the output image data GS. That is, the autoencoder, which is the trained model of this example, is generated by performing machine learning in advance using a large amount of normal product image data GN as learning data.

Therefore, when the capsule 6 containing scratches, cracks, etc., that is, the inspection image data GK of the defective product is input, the difference image generation unit 110 uses the autoencoder to output the capsule 6 containing scratches, cracks, etc., that is, the output image data of the normal product. GS can be generated. Then, the difference image generation unit 110 generates a difference image data GD that represents a difference image that is a difference between the inspection image of the defective product represented by the inspection image data GK and the output image of the normal product represented by the output image data GS. Generate.

That is, when the inspection image data GK representing the inspection image of the defective product is input, the difference image generation unit 110 generates the difference image data GD representing the difference between the inspection image data GK and the output image data GS. Can be done. In this example, as described above, the inspection image of the defective capsule 6 represented by the inspection image data GK includes scratches, cracks, and the like. In this case, in the difference image, the brightness of the pixel corresponding to scratches and cracks is large. Specifically, if the difference image is a black-and-white image, the scratches and cracks are pixel values with respect to the black of the inspection target surface S. Represented by the large white of.

Here, the configuration of the difference image generation unit 110 will be described in more detail with reference to FIG. The difference image generation unit 110 includes a learning image data acquisition unit 111, a learning image data storage unit 112, a model generation unit 113, a model storage unit 114, an inspection image data acquisition unit 115, and an image reconstruction unit 116. And an output unit 117 are mainly provided. The learning image data acquisition unit 111, the learning image data storage unit 112, and the model generation unit 113 function as a learning processing unit (learning phase) of the difference image generation unit 110. Further, the model storage unit 114, the inspection image data acquisition unit 115, the image reconstruction unit 116, and the output unit 117 function as a reproduction image reconstruction unit (inference phase) of the difference image generation unit 110.

The learning image data acquisition unit 111 is the normal product image data representing the image of the inspection target surface S of the capsule 6 of the normal product among the image data G representing the image of the inspection target surface S of the capsule 6 captured by the camera 11. Get GN. That is, the learning image data acquisition unit 111 acquires, among the images represented by the image data G, the normal product image data GN confirmed to be a normal product by prior inspection as teacher data. The learning image data storage unit 112 stores a large amount of normal product image data GN acquired by the learning image data acquisition unit 111 as a learning image data group.

The model generation unit 113 uses the learning image data group stored in the learning image data storage unit 112, that is, a large amount of normal product image data GN as teacher data, and generates a trained model M which is an auto encoder. As a result, the trained model M generated by the model generation unit 113 outputs only the output image data GS representing the output image of the normal product.

Here, the trained model M generated will be described. As shown in FIG. 6, the trained model M is an autoencoder including an input layer M1, an intermediate layer M2, and an output layer M3. Image data G is input to the input layer M1 as input data. In the learning phase, only the normal product image data GN is input to the input layer M1, and in the inference phase, the inspection image data GK is input to the input layer M1. Here, the image data G input as input data is composed of pixel data representing the pixel value of each pixel forming the image.

The intermediate layer M2 has a first intermediate layer M21 that compresses the dimension of the input data and extracts the feature amount contained in the input data. Further, the intermediate layer M2 has a second intermediate layer M22 that expands the dimension of the data compressed by the first intermediate layer M21 and makes the dimension of the data equal to the dimension of the input data. The output layer M3 outputs the output image data GS obtained by extracting the feature amount from the image data G, that is, the normal product image data GN and the inspection image data GK.

As a result, as will be described later, the determination value calculation unit 140 inspects using the inspection image data GK input to the trained model M as input data and the output image data GS output from the trained model M. The quality of the capsule 6 which is the object is determined. Then, the inspection device 100 can reconstruct a high-quality output image in the inference phase by generating a trained model M (autoencoder) that extracts an appropriate feature amount in the learning phase. As a result, the inspection device 100 can provide a high-precision quality determination result.

The model storage unit 114 stores the trained model M generated by the model generation unit 113. Then, the model storage unit 114 outputs the learned model M stored in the image reconstruction unit 116. The inspection image data acquisition unit 115 acquires inspection image data GK representing an image of the inspection target surface S of the capsule 6 which is an inspection target imaged by the camera 11. Then, the inspection image data acquisition unit 115 outputs the acquired inspection image data GK to the image reconstruction unit 116.

The image reconstruction unit 116 inputs the inspection image data GK acquired from the inspection image data acquisition unit 115 as input data to the trained model M acquired from the model storage unit 114. As a result, the image reconstruction unit 116 generates the output image data GS representing the output image of the normal product regardless of whether the inspection image data GK is a normal product or a defective product. Then, the image reconstruction unit 116 generates the difference image data GD represented by the difference between the inspection image data GK and the generated output image data GS.

Here, the difference image data GD generated by the image reconstruction unit 116 will be described. For example, as shown in FIG. 7, the inspection image data GK representing the inspection image of the capsule 6 containing a defect D such as a scratch or a crack on the inspection target surface S, that is, the defective capsule 6, is input to the image reconstruction unit 116. Suppose it was done. In this case, the image reconstruction unit 116 inputs the defective product inspection image data GK as input data to the trained model M, and as shown in FIG. 8, the output image data GS representing the output image of the normal product. Is reconstructed and generated.

Then, the image reconstruction unit 116 calculates the difference between the inspection image data GK of the defective product and the output image data GS of the normal product, and as shown in FIG. 9, the difference representing the difference image in which the common part is omitted. Generate image data GD. Here, in the difference image represented by the difference image data GD, the defect D represented by white is clearly present, and the edge portion S1 of the capsule 6 is also clearly represented by white. When the inspection image data GK of a normal product that does not include the defect D is input to the inspection target surface S, the image reconstruction unit 116 displays the difference image data GD that does not include the defect D, as shown in FIG. To generate. Then, the generated difference image data GD is output from the output unit 117 to the determination value calculation unit 140.

By the way, when a slight positional deviation (posture deviation) occurs in the capsule 6 fixed to the column side bracket 5B, the capsule is particularly shown in the difference image represented by the difference image data GD in FIG. The edge portion S1 of the inspection target surface S of 6 may occur. In FIG. 9, the edge portion S1 in the difference image is larger than the pixel value of the inspection target surface S and is white.

Here, in the difference image, the pixels corresponding to the edge portion S1 have variations in pixel values according to the positional deviation (posture deviation) of the capsule 6 at the time of imaging. When the pixel value of the edge portion S1 is large, it becomes difficult to automatically distinguish the edge portion S1 from the defect D. Therefore, even if the inspection image does not include the defect D, the quality determination unit 150, which will be described later, generates a difference image in which the edge portion S1 is generated due to the positional deviation (posture deviation), so that the inspection image is regarded as a defective product. There is a risk of erroneous judgment.

Therefore, in order to effectively exclude the edge portion S1 generated due to the positional deviation (postural deviation), the distributed image generation unit 120 includes all of the normal product image data GNs constituting the learning image data group. The variance value as a basic statistic is calculated for the pixel value of the pixel. Then, the distribution image generation unit 120 generates a distribution image that reflects the calculated distribution of the dispersion values. In this example, the variance value is exemplified as the basic statistic, but as the basic statistic, in addition to the variance value, the average value of the pixel values (characteristic values), the standard deviation, the average deviation (absolute deviation), etc. Can be exemplified.

Specifically, the distribution image generation unit 120 first acquires a learning image data group stored in the learning image data storage unit 112, that is, a large amount of normal product image data GN, and the acquired normal product image data GN. The pixel value for each and every pixel of is extracted. Next, the distribution image generation unit 120 uses the pixel values extracted for the corresponding pixels in the normal product image data GNs to calculate the dispersion value representing the variation in the pixel values for the entire learning image data group.

Then, as shown in FIG. 11, the distribution image generation unit 120 generates a distribution image reproduced by the pixel values corresponding to the dispersion values. Here, the distribution image generation unit 120 finally sets a region in which the calculated variance value is larger than the preset reference dispersion value as the edge region R1 (the region shown in white in FIG. 11), and the calculated variance value. The region where is equal to or less than the reference dispersion value is defined as the inspection determination region R2 (the region shown in black in FIG. 11). When the distribution image generation unit 120 generates a distribution image, the distribution image data GV representing the generated distribution image is output to the weight image generation unit 130.

The weighted image generation unit 130 acquires the distribution image data GV output from the distribution image generation unit 120. Then, the weight image generation unit 130 sets a small weight for a pixel having a large dispersion value and a large weight for a pixel having a small dispersion value based on the acquired distribution image data GV. That is, the weight image generation unit 130 sets the weight of the edge region R1 to be small and the weight of the inspection determination region R2 to be large.

For example, in the weighted image generation unit 130, as shown in FIG. 12, the edge region R1 is represented by black with a small pixel value, and the inspection determination region R2 is represented by white with a large pixel value. Generate image data GW. Then, the weight image generation unit 130 outputs the generated weight image data GW to the determination value calculation unit 140.

As shown in FIGS. 13 and 14, the determination value calculation unit 140 multiplies the difference image data GD generated by the difference image generation unit 110 and the weight image data GW generated by the weight image generation unit 130. , Judgment image data Generates a judgment image represented by GJ. That is, by multiplying the difference image data GD by the weighted image data GW, the pixel value of the edge region R1 becomes smaller. On the other hand, for the defect D, the weight for the pixel value is not multiplied. Therefore, as shown in FIG. 13, the determination image data GJ representing the determination image of the defective product in which the defect D is represented in white is obtained, and as shown in FIG. 14, when the defect D is not present, the defect D is white. Judgment image data GJ representing a judgment image of a normal product not included is obtained.

Then, the determination value calculation unit 140 calculates the average value obtained by averaging the pixel values (characteristic values) of all the pixels of the determination image as the determination value based on the generated determination image data GJ. The determination value calculation unit 140 outputs the calculated average value (determination value) to the quality determination unit 150.

The quality determination unit 150 determines the quality of the capsule 6 based on the average value (determination value) output from the determination value calculation unit 140. That is, if the average value (judgment value) of the quality determination unit 150 is equal to or higher than the preset determination reference value, the determination image shown in FIG. 13 has a region having a large pixel value, that is, a defect D. , Capsule 6 is determined as a defective product. On the other hand, if the average value (judgment value) is less than the judgment reference value, the pass / fail judgment unit 150 determines that the capsule 6 is a normal product because the defect D having a large pixel value does not exist in the judgment image shown in FIG. do.

As can be understood from the above description, according to the inspection system 10 of this example, the inspection device 100 uses the trained model M when inspecting the capsule 6 using the trained model M which is an autoencoder. The determination image data GJ can be generated by using the difference image data GD generated thereby and the weighted image data GW based on the distributed image data GV. Then, the inspection device 100 can calculate the determination value using the determination image data GJ.

As a result, the inspection device 100 can calculate the determination value in consideration of the region where the dispersion value, which is the basic statistic, is large, that is, the edge portion S1 which may reduce the accuracy of the quality determination. Therefore, the inspection device 100 can improve the quality determination of the capsule 6, that is, the inspection accuracy based on the calculated determination value. In this example, in the production line of the EPS device T, the inspection process is set when the capsule 6 is assembled to the column side bracket 5B and transported to the next process, and the quality of the capsule 6 is judged. However, the quality determination of the capsule 6 can also be performed in the final product inspection after the assembly of the EPS device T is completed.

(5. Others)
In this example described above, the trained model M, which is an autoencoder generated by the model generation unit 113 of the difference image generation unit 110, is stored in the model storage unit 114. By the way, the trained model M is generated by performing machine learning using a learning image data group (a large amount of normal product image data GN) stored in the learning image data storage unit 112. Therefore, as the number of normal product image data GN increases, in other words, as the learning by machine learning progresses, the trained model M with high accuracy can be generated. Therefore, the model generation unit 113 sequentially uses the trained model. M (autoencoder) can be updated.

Further, in the above-mentioned example, as an image pickup device, a camera 11 for capturing an image of the inspection target surface S (surface) of the capsule 6 which is an inspection target is provided, and the image is based on the inspection image data GK of the imaged inspection target surface S. The presence or absence of the defect D existing on the surface of the capsule 6 was inspected. By the way, defects in the inspection target, especially the inspection target produced on the production line, may exist not only on the outside (surface) but also on the inside, and the presence or absence of such internal defects is also the inspection target. obtain.

Therefore, as the image pickup device, for example, it is possible to use a device that acquires the inspection image data GK that captures the internal state of the inspection object by using electromagnetic waves, sound, or the like. Also in this case, the inspection image data GK representing the internal defect and the output image data GS output from the trained model M (autoencoder) using the inspection image data GK as input data are used as in the above-mentioned example. It is possible to make a pass / fail judgment. Therefore, in this case as well, the same effect as that of this example described above can be obtained.

1 ... Steering wheel, 2 ... Steering mechanism, 3 ... Assist mechanism, 4 ... Column device, 5 ... Support mechanism, 5A ... Upper bracket, 5B ... Column side bracket, 6 ... Capsule, 6a ... Insertion hole, 7 ... Bolt, K ... Mounting stay, T ... Electric power steering device, W ... Steering wheel, 10 ... Inspection system, 11 ... Camera, 100 ... Inspection device, 110 ... Difference image generation unit, 111 ... Learning image data acquisition unit, 112 ... Learning Image data storage unit, 113 ... model generation unit, 114 ... model storage unit, 115 ... inspection image data acquisition unit, 116 ... image reconstruction unit, 117 ... output unit, 120 ... distribution image generation unit, 130 ... weighted image generation unit , 140 ... Judgment value calculation unit, 150 ... Good / bad judgment unit, D ... Defect, G ... Image data, GD ... Difference image data, GJ ... Judgment image data, GK ... Inspection image data, GN ... Normal product image data, GS ... Output image data, GV ... distribution image data, GW ... weighted image data, M ... trained model, M1 ... input layer, M2 ... intermediate layer, M21 ... first intermediate layer, M22 ... second intermediate layer, M3 ... output layer , R1 ... Edge area, R2 ... Inspection judgment area, S ... Inspection target surface, S1 ... Edge portion

Claims (11)

An image pickup device that captures an image of the object to be inspected,
An inspection device for determining the quality of the inspection object based on the image of the inspection object captured by the image pickup device is provided.
The inspection device is
A model storage unit that stores a trained model, which is an autoencoder that outputs output image data that represents the output image of a normal product, using a group of image data for learning that has multiple image data of a normal product that represents an image of a normal product as input data. ,
An inspection image data acquisition unit that acquires inspection image data representing an image of the inspection object, and an inspection image data acquisition unit.
Using the inspection image data and the output image data output by inputting the inspection image data into the trained model, a difference image representing the difference between the inspection image data and the output image data is generated. Difference image generator and
Distribution image generation that calculates basic statistics for the characteristic values of all pixels of each normal product image data in the training image data group and generates a distribution image showing the distribution of the characteristic values based on the basic statistics. Department and
A weight image generation unit that sets weights according to the basic statistic based on the distribution image data representing the distribution image generated by the distribution image generation unit and generates a weight image representing the weights.
A judgment value calculation unit that calculates a judgment value using the difference image data representing the difference image and the weight image data representing the weight image, and a judgment value calculation unit.
A quality determination unit that determines the quality of the inspection target based on the determination value,
Equipped with an inspection system. The distribution image generation unit calculates, as the basic statistic, one of the dispersion value of the characteristic value, the average value of the characteristic value, the standard deviation of the characteristic value, and the average deviation of the characteristic value. , The inspection system according to claim 1. The region including the pixel having a large basic statistic in the distribution image is a peripheral portion of the inspection target surface imaged due to the positional shift of the inspection object caused by the image pickup by the imaging device, claim 1 or The inspection system according to 2. The inspection according to any one of claims 1-3, wherein the weight image generation unit sets a small weight for a pixel having a large basic statistic and a large weight for a pixel having a small basic statistic. system. Any of claims 1-4, wherein the determination value calculation unit generates determination image data by multiplying the difference image data and the weighted image data, and calculates the determination value based on the determination image data. The inspection system described in item 1. The inspection system according to claim 5, wherein the determination value calculation unit calculates an average value obtained by averaging the characteristic values of each pixel in the determination image represented by the determination image data as the determination value. The inspection according to any one of claims 1-6, wherein the quality determination unit determines that the inspection target is a defective product in which a defect exists when the determination value is equal to or higher than a preset determination reference value. system. The inspection system according to any one of claims 1-7, wherein the characteristic value is a pixel value representing the shade and brightness of a pixel color or a luminance value representing the brightness of a pixel. The inspection system according to any one of claims 1-8, wherein the inspection object is a product produced via a production line. The inspection system according to claim 9, wherein the inspection object is inspected when it is transported to the next process on the production line. The inspection system according to any one of claims 1-10, wherein the inspection object is a capsule forming a support mechanism for supporting a column device in a vehicle steering device on a vehicle body.

Images