JP2022087930A - Method for performing inspection on image to be inspected, information processing apparatus, and computer program - Google Patents

Abstract

To provide technique capable of satisfying both high determination accuracy and flexibility of determination method.SOLUTION: A method for performing an inspection on an image to be inspected comprises: (a) a process of creating a converted image by performing image conversion for removing noise of the image to be inspected using a machine learning model including a convolutional neural network that has an encoder unit and a decoder unit; (b) a process of generating measurement results for one or more inspection items by performing image processing on the converted image; and (c) a process of generating determination results regarding the inspection item by comparing the measurement results with a predetermined determination standard.SELECTED DRAWING: Figure 4

Description

The present disclosure relates to a method of performing an inspection of an image to be inspected, an information processing apparatus, and a computer program.

As an inspection system for inspecting and determining an image, there are an inspection system using a machine learning model as disclosed in Patent Document 1 and a rule-based inspection system using image processing. Inspection systems that use machine learning models are basically difficult to intervene in judgment. On the other hand, a rule-based inspection system using image processing is realized by a person designing processing parameters and judgment rules, so that it is basically possible to intervene in the judgment. However, from the viewpoint of determination accuracy, an inspection system using a machine learning model is often superior to a rule-based inspection system using image processing.

Japanese Unexamined Patent Publication No. 2020-119159

As described above, in the prior art, there is a problem that it is difficult to satisfy both the demands of high determination accuracy and flexibility of the determination method. The inventor of the present disclosure has found that by successfully combining a machine learning model and image processing, it is possible to satisfy two demands of high judgment accuracy and flexibility of judgment method.

According to the first aspect of the present disclosure, there is provided a method of performing an inspection of an image to be inspected. This method is a step of creating a converted image by (a) performing image conversion for removing noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit. And (b) a step of generating a measurement result for one or more inspection items by performing image processing on the converted image, and (c) comparing the measurement result with a predetermined determination criterion. This includes a step of generating a determination result regarding the inspection item.

According to the second aspect of the present disclosure, an information processing apparatus for performing an inspection of an image to be inspected is provided. This information processing device includes a memory for storing a machine learning model including a convolutional neural network having an encoder unit and a decoder unit, and a processor for executing an operation using the machine learning model. The processor creates a converted image by (a) performing image conversion for removing noise of the image to be inspected using the machine learning model, and (b) an image for the converted image. By executing the process, a process of generating a measurement result for one or more inspection items, and (c) by comparing the measurement result with a predetermined determination criterion, a determination result for the inspection item is generated. It is configured to perform processing and.

According to the third aspect of the present disclosure, a computer program for causing a processor to inspect an image to be inspected is provided. The computer program creates a converted image by (a) performing an image conversion that removes noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit. Processing, (b) processing to generate measurement results for one or more inspection items by executing image processing on the converted image, and (c) comparison of the measurement results with predetermined determination criteria. By doing so, the processor is made to execute the process of generating the determination result regarding the inspection item.

A block diagram of an information processing device that functions as an image inspection system. An explanatory diagram showing an example of a machine learning model. A flowchart showing the processing procedure of the inspection process. Explanatory drawing which shows the flow of inspection process. A flowchart showing the processing procedure of the learning process. Explanatory diagram showing a learning process. The flowchart which shows the processing procedure of 2nd Embodiment of a learning process.

FIG. 1 is a block diagram of an information processing apparatus 100 that functions as an image inspection system. The information processing device 100 includes a processor 110, a memory 120, an interface circuit 130, an input device 140 connected to the interface circuit 130, and a display unit 150. The processor 110 functions as an image inspection unit 112 that executes an inspection of the image to be inspected, and a learning execution unit 114 that executes learning of a machine learning model and optimization of image processing. The image inspection unit 112 and the learning execution unit 114 are each realized by the processor 110 executing a computer program stored in the memory 120. However, at least one of the image inspection unit 112 and the learning execution unit 114 may be realized by a hardware circuit. Further, the processor that executes image inspection and learning may be a processor included in a remote computer connected to the information processing unit 100 via a network. The machine learning model 200, the image processing program 300, and the image to be inspected Di are stored in the memory 120. The machine learning model 200 and the image processing program 300 are used for processing by the image inspection unit 112.

FIG. 2 is an explanatory diagram showing an example of the machine learning model 200. The machine learning model 200 has a function of converting the image to be inspected Di into a converted image Dic, and is configured as an autoencoder. The autoencoder has a substantially symmetrical encoder / decoder structure. Specifically, the machine learning model 200 is configured as a convolutional neural network having an encoder unit 210 between the input layer 201 and the connection layer 202 and a decoder unit 220 between the connection layer 202 and the output layer 203. .. The encoder unit 210 has a function of extracting the features of the image to be inspected Di. The decoder unit 220 has a function of restoring an image from the extracted features to generate a converted image Dic.

The machine learning model 200 is not limited to the autoencoder, and other types of convolutional neural networks can be used. Specifically, it can be configured as U-Net or CycleGAN. The autoencoder, U-Net, and CycleGAN are all convolutional neural networks having an encoder unit and a decoder unit. The U-Net has an encoder / decoder structure similar to that of an autoencoder. However, U-Net differs from the autoencoder in that it has a path for skipping and connecting the feature map output in each layer of the encoder unit to the corresponding layer of the decoder unit. The learning of the autoencoder and U-Net can be performed as unsupervised learning based on image restoration of the input image, or can be performed as supervised learning using a desirable converted image. In autoencoder and U-Net learning, cross entropy error, MSE (Mean Square Error), PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural SIMilarity), etc. are used as loss functions. The CycleGAN has a generator and a discriminator, and the generator has an encoder unit, a transformer unit, and a decoder unit. When an image is input to the trained CycleGAN, the image converted by the generator is output. A U-Net structure is adopted for a normal CycleGAN generator, but it is also established even if the neural network structure is changed to an autoencoder. However, the learning method and loss function of CycleGAN are different from those of the autoencoder and U-Net. CycleGAN is learned based on so-called domain conversion, and Cycle-Consistency Loss or the like is used as a loss function. As described above, the autoencoder, U-Net, and CycleGAN all correspond to a convolutional neural network having an encoder unit and a decoder unit. As the convolutional neural network having an encoder unit and a decoder unit, it is also possible to use another type of GAN (Generative Adversarial Network) such as pix2pix.

When using any of the above-mentioned types of models, the machine learning model 200 has a function of performing image conversion for removing noise of the image to be inspected Di to generate a converted image Dic. preferable. In the present disclosure, the "noise of the image to be inspected Di" means a pixel value distribution such as shading, brightness, and shading that is irrelevant to the quality determination of the image to be inspected Di. In the machine learning model 200 shown in FIG. 2, since the noise of the image to be inspected Di is not recognized as a feature by the encoder unit 210, the decoder unit 220 can generate the converted image Dic in which the noise is reduced.

FIG. 3 is a flowchart showing a processing procedure of the inspection process, and FIG. 4 is an explanatory diagram showing a flow of the inspection process.

In step S110, the image inspection unit 112 generates the converted image Dic by performing image conversion for removing noise of the image to be inspected Di by using the trained machine learning model 200. In the example shown in FIG. 4, the converted image Dic contains the foreign matter FM, the scratch DF, and the crack CR contained in the image Di to be inspected, and other noises are removed. In the present embodiment, the foreign matter FM, the scratch DF, and the crack CR are defects set in advance as inspection items related to the image to be inspected Di. In this embodiment, three inspection items are set, but one or more inspection items may be set. In the present disclosure, the phrase "defect" includes a defect that is small in size and does not prevent the image Di to be inspected from being good, and a defect that is large in size and is determined to be defective in the image Di to be inspected. Including both.

In step S120, the image inspection unit 112 uses the image processing program 300 to perform image processing on the converted image Dic to generate measurement results for each inspection item. In the example of FIG. 4, as the measurement results by image processing, the size of the foreign matter FM is 0.5 mm ² , the size of the scratch DF is 4 mm, and the size of the crack CR is 2 mm.

The image processing program 300 can generate measurement results for each inspection item by, for example, combining various processes such as the following as necessary.
・ Edge detection processing ・ Circle detection processing ・ Straight line detection processing ・ Frequency separation processing ・ Color space separation processing ・ Binarization processing ・ Pattern matching processing ・ Feature point extraction processing

The above-mentioned individual processing is called "individual image processing". The image processing program 300 is set to execute a plurality of individual image processes in order. The image processing parameters set in the image processing program 300 include a plurality of individual image processing processing sequences and processing parameters for each individual image processing. The processing parameters of the individual image processing are, for example, the filter size and the individual weights in the filter when the filter is used in the individual image processing.

In step S130, the image inspection unit 112 uses the measurement result obtained in step S120 to execute the determination according to a predetermined determination criterion. In the example of FIG. 4, the criteria are defined that the foreign matter FM is good when it is less than 1 mm ² , the scratch DF is good when it is less than 3 mm, and the crack CR is good when it is less than 5 mm. As a result, it is determined that the foreign matter FM and the crack CR are good, but the scratch DF is bad. The judgment criteria can be arbitrarily set by the user of the inspection system. For example, a determination criterion for the following items may be provided.
(1) Detection point of determination item: For example, the threshold value of the size of foreign matter or the like may change depending on whether or not foreign matter or the like is on the wiring.
(2) Color: A change in color due to burning, charring, stains, etc. can be used as a criterion for determining defects.
(3) Misalignment: Device misalignment, printing misalignment, etc. can be used as judgment criteria.
(4) Shape defect: Judgment can be made by using an index of shape defect such as an area, a deviation amount from a reference, and a position of the center of gravity.

As described above, in the present embodiment, after the image conversion using the machine learning model 200, the image processing using the image processing program 300 is executed to acquire the measurement results for one or more inspection items, so that the disturbance Even for the image Di to be inspected, which has a lot of noise such as variations and variations, the inspection can be performed with high accuracy. In addition, since the pass / fail judgment for the inspection item is performed using the measurement result obtained by the image processing, it is possible to flexibly change the pass / fail judgment method as compared with the case where the pass / fail judgment is executed by the machine learning model 200. Is.

FIG. 5 is a flowchart showing a processing procedure of a first learning step S200 for learning the machine learning model 200 and a second learning step S300 for optimizing the image processing program 300, and FIG. 6 is a flowchart showing these learning steps. It is explanatory drawing which shows. The learning step is performed before the inspection step described above.

The learning data used at the time of learning the machine learning model 200 differs depending on the configuration of the machine learning model 200, and is, for example, as follows.

<When learning using reconstruction error>
With an autoencoder or U-Net, unsupervised learning can be performed based on the reconstruction error due to image restoration. As a loss function, MSE (Mean Square Error), PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural SIMilarity), etc. are used. The learning data is configured to include, for example, the following two.
(A) Learning image including defects and noise of inspection items (b) True measurement result in learning image Since noise is not recognized as a feature by the encoder unit 210, a teacher using a learning image containing noise When none learning is performed, the machine learning model 200 can be trained so as to remove noise in the input image.

<When learning using segmentation information>
With an autoencoder or U-Net, supervised learning can be performed using segmented images. A cross entropy error is used as the loss function. The learning data is configured to include, for example, the following three.
(A) Learning image containing defects and noise of inspection items (b) Correct image containing the same defects as the learning image and not containing noise (c) True measurement result in the learning image As a "correct image" , A segmentation image obtained by segmenting a learning image can be used. When creating a segmented correct image, it is preferable to remove the noise contained in the learning image by image processing or manually. Alternatively, first, two images, a semantic map and an output image, are prepared as simulation images such as CAD images, the semantic map is used as the correct image, and the learning image is created by adding noise to the output image. You may create it. If supervised learning is performed on a machine learning model 200 composed of an autoencoder or U-Net, machine learning is performed so that an image in which noise is removed in an actual test image can be obtained as compared with unsupervised learning. The learning of the model 200 can be performed. The first learning data for executing the learning of the machine learning model 200 and the second learning data for executing the optimization of the image processing parameters may be separately prepared. In this case, the first learning data does not need a true measurement result, and the second learning data does not need a correct image.

<When learning using images from another domain>
CycleGAN is learned based on so-called domain conversion, and Cycle-Consistency Loss or the like is used as a loss function. The learning data is configured to include, for example, the following three.
(A) First image group including defects and noise of inspection items (b) Second image group including defects of inspection items and not containing noise (c) True measurement results in images of the first image group First The image group is an image group of the first domain, and is, for example, a set of actual images of an object to be inspected. The second image group is an image group of the second domain, and is, for example, a set of simulation images represented by a CAD image. The second image group does not need to be associated with the first image group on a one-to-one basis. Even when CycleGAN is used, the first learning data for executing the learning of the machine learning model 200 and the second learning data for executing the optimization of the image processing parameters are separately separated. You may prepare. In this case, the first learning data does not need a true measurement result, and the second learning data does not need a second image group.

In the learning process shown in FIGS. 5 and 6, a case where the machine learning model 200 is configured by an autoencoder or U-Net and supervised learning is performed will be described. As shown in FIG. 6, as the learning data, a learning image Dt and a corresponding correct answer image Dto are prepared in advance. The correct image Dto is an image containing foreign matter FM, scratch DF, and crack CR, and does not contain noise. In the present embodiment, the learning image Dt is an image in which noise is added to the correct answer image Dto.

In the first step S210 of the first learning step S200, the learning execution unit 114 generates a plurality of inferred images Dtc by image-converting a plurality of learning image Dt using the machine learning model 200. In step S220, the learning execution unit 114 calculates the loss La of the inferred image Dtc. In the present embodiment, the loss La is the value of the loss function calculated according to the difference between the inferred image Dtc and the correct image Dto. As the loss function, various loss functions used for learning the convolutional neural network can be used. In the case of unsupervised learning, the loss La is calculated according to the difference between the inferred image Dtc and the learning image Dt. In step S230, the learning execution unit 114 executes learning of the machine learning model 200 using the loss La calculated in step S220. In step S240, the learning execution unit 114 determines whether or not the learning of the machine learning model 200 is completed, and if not, returns to step S210 and repeats steps S210 to S230. Whether or not the training of the machine learning model 200 is completed is, for example, whether or not the loss La is equal to or less than a predetermined threshold value, or whether or not the update of the loss value (loss) of the verification data has converged. Judged by. The learning method of the machine learning model 200 according to these steps S210 to S240 is the same as the normal learning method.

In the first step S310 of the second learning step S300, the learning execution unit 114 generates a plurality of inferred images Dtc by image-converting a plurality of learning image Dt using the trained machine learning model 200.

In step S320, the learning execution unit 114 generates a measurement result by performing image processing of a plurality of inferred images Dtc using the image processing program 300. In the example of FIG. 6, the measurement results for the inferred image Dtc are 0.9 mm ² for the foreign matter FM, 5 mm for the scratch DF, and 2 mm for the crack CR. The judgment result of comparing these measurement results with the judgment criteria is that the foreign matter FM is good, the scratch DF is bad, and the crack CR is good. In FIG. 6, the true values for the same inferred image Dtc are 1 mm ² for foreign matter FM, 3 mm for scratch DF, and 0.8 mm for crack CR, and the true judgment result is that foreign matter FM is defective and scratch DF is defective. , Crack CR has been shown to be good.

In step S330, the learning execution unit 114 calculates the difference Eb between the measurement result and the true value. When there are a plurality of inspection items as shown in FIG. 6, the difference Eb can be calculated as a mean square error or a mean absolute error regarding all the differences between the measurement results and the true values in the plurality of inspection items. At this time, it is preferable to calculate the mean square error and the mean absolute error by using the difference normalized by dividing the difference between the measurement result and the true value by the true value.

In step S340, the learning execution unit 114 executes the optimization of the parameters of the image processing program 300 using the difference Eb. The parameters of the image processing program 300 are synonymous with "parameters of image processing". As described above, the image processing parameters include a plurality of individual image processing processing sequences and processing parameters for each individual image processing. Optimizing the image processing parameters means a process of adjusting the image processing parameters so that the difference Eb between the measurement result and the true value becomes as small as possible. For example, as the processing order of a plurality of individual image processes used for image processing, a plurality of processing orders are prepared in advance, and the one having the smallest difference Eb between the measurement result and the true value is selected as the optimum processing order. be able to. Further, as a plurality of processing parameters for individual image processing, a plurality of processing parameter groups can be prepared in advance, and the one having the smallest difference Eb between the measurement result and the true value can be selected as the optimum processing parameter group. Either the processing order optimization or the processing parameter optimization may be executed first. In addition, multiple sets of combinations of multiple processing sequences and multiple processing parameter groups are prepared, and the one with the smallest difference Eb between the measurement result and the true value is selected as the optimum parameter for image processing. You may try to do it.

In step S350, the learning execution unit 114 determines whether or not the optimization of image processing is completed, and if not, returns to step S310 and repeats steps S310 to S340. When returning to step S310, it is preferable to use a new image as a plurality of learning images Dt. Whether or not the optimization of image processing is completed depends on, for example, whether or not the difference Eb between the measurement result and the true value is equal to or less than a predetermined threshold value, or whether or not the specified number of iterations is searched. Judged.

As described above, in the learning process shown in FIGS. 5 and 6, the optimization of the image processing parameters is executed after the learning of the machine learning model 200 is executed, so that the optimization of the machine learning model 200 and the image processing is successful. It is possible to do it.

In an inspection system that has already been learned and once established, if a new disturbance caused by measurement or manufacturing occurs and the inspection accuracy deteriorates, two countermeasures can be considered. The first corresponding method is a method of executing the second learning step S300 in which the machine learning model 200 is maintained as it is and the parameters of the image processing are optimized. The second corresponding method is a method of executing the first learning step S200 for learning the machine learning model 200 while maintaining the parameters of the image processing as they are. In the latter case, when the learning image Dt is created by adding noise to the correct image Dto, it is preferable to add noise corresponding to the new disturbance to the correct image Dto. Alternatively, when the correct image Dto is created by removing the noise from the learning image Dt, it is preferable to manually remove the noise due to the new disturbance contained in the learning image Dt by image processing or manual operation.

FIG. 7 is a flowchart showing the processing procedure of the second embodiment of the learning process. Of these learning steps, the first learning step S200 is the same as in FIG. 5, and the second learning step S301 is different from FIG. That is, in the second learning step S301 of FIG. 7, steps S312 and S314 are added between steps S310 and S320, and the other steps are the same as those of FIG.

In step S312, the learning execution unit 114 calculates the loss La of the inferred image Dtc obtained in step S310. In step S314, the learning execution unit 114 executes additional learning of the machine learning model 200 using the loss La calculated in step S312. These steps S312 and S314 are substantially the same as steps S220 and S230 of the first learning step S200. In this way, if additional learning of the machine learning model 200 is executed in the second learning step, the accuracy of the machine learning model 200 can be improved. It should be noted that steps S312 and S314 may be executed after step S310, and may be executed at a timing different from the timing shown in FIG.

-Other embodiments:
The present disclosure is not limited to the above-described embodiment, and can be realized in various forms without departing from the spirit thereof. For example, the present disclosure can also be realized by the following aspect. The technical features in each of the embodiments described below correspond to the technical features in the above embodiments in order to solve some or all of the problems of the present disclosure, or some or all of the effects of the present disclosure. It is possible to replace or combine as appropriate to achieve the above. Further, if the technical feature is not described as essential in the present specification, it can be appropriately deleted.

(1) According to the first aspect of the present disclosure, a method for performing an inspection of an image to be inspected is provided. This method is a step of creating a converted image by (a) performing image conversion for removing noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit. And (b) a step of generating a measurement result for one or more inspection items by performing image processing on the converted image, and (c) comparing the measurement result with a predetermined determination criterion. This includes a step of generating a determination result regarding the inspection item.
According to this method, after image conversion using a machine learning model, measurement results related to inspection items are acquired by executing image processing, so even for images to be inspected that have a lot of noise such as disturbance and variation. The inspection can be performed with high accuracy. Further, since the judgment is made for the inspection item using the measurement result obtained by the image processing, it is possible to flexibly change the judgment method as compared with the case where the judgment is executed by the machine learning model.

(2) In the above method, the machine learning model may be configured as an autoencoder, U-Net, or CycleGAN.
According to this method, it is possible to easily construct a machine learning model that performs image conversion that removes noise from the image to be inspected.

(3) The method further optimizes the parameters of the image processing after the learning step of executing the learning of the machine learning model before the step (a) and the first learning step. It may have a learning process. The second learning step includes (i) inputting a learning image into the trained machine learning model to generate an inferred image, and (ii) executing the image processing on the inferred image. (Iii) Optimizing the parameters of the image processing by using the difference between the true value for the inspection item and the learning measurement result in the learning image and the step of generating the learning measurement result for the inspection item. It may include a step of executing the conversion.
According to this method, the optimization of the image processing parameters is executed after the training of the machine learning model is executed, so that the optimization of the machine learning model and the image processing can be executed well.

(4) In the above method, the second learning step may include a step of executing additional learning of the machine learning model using the inference image after the step (i).
According to this method, the accuracy of the machine learning model can be improved because the additional learning of the machine learning model is executed in the second learning step of executing the optimization of the image processing parameters.

(5) According to the second aspect of the present disclosure, an information processing apparatus for executing an inspection of an image to be inspected is provided. This information processing device includes a memory for storing a machine learning model including a convolutional neural network having an encoder unit and a decoder unit, and a processor for executing an operation using the machine learning model. The processor creates a converted image by (a) performing image conversion for removing noise of the image to be inspected using the machine learning model, and (b) an image for the converted image. By executing the process, a process of generating a measurement result for one or more inspection items, and (c) by comparing the measurement result with a predetermined determination criterion, a determination result for the inspection item is generated. It is configured to perform processing and.
According to this information processing device, after image conversion using a machine learning model, measurement results related to inspection items are acquired by executing image processing, so that the image to be inspected has a lot of noise such as disturbance and variation. Can also perform accurate inspections. Further, since the judgment is made for the inspection item using the measurement result obtained by the image processing, it is possible to flexibly change the judgment method as compared with the case where the judgment is executed by the machine learning model.

(6) According to the third aspect of the present disclosure, a computer program for causing a processor to inspect an image to be inspected is provided. The computer program creates a converted image by (a) performing an image conversion that removes noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit. Processing, (b) processing to generate measurement results for one or more inspection items by executing image processing on the converted image, and (c) comparison of the measurement results with predetermined determination criteria. By doing so, the processor is made to execute the process of generating the determination result regarding the inspection item.
According to this computer program, after image conversion using a machine learning model, measurement results related to inspection items are acquired by executing image processing, so even for images to be inspected that have a lot of noise such as disturbance and variation. , Can perform accurate inspection. Further, since the judgment is made for the inspection item using the measurement result obtained by the image processing, it is possible to flexibly change the judgment method as compared with the case where the judgment is executed by the machine learning model.

The present disclosure can also be realized in various forms other than the above. For example, it can be realized in the form of a non-transitory storage medium or the like in which a computer program for realizing the image inspection function is recorded.

100 ... Information processing device, 110 ... Processor, 112 ... Image inspection unit, 114 ... Learning execution unit, 120 ... Memory, 130 ... Interface circuit, 150 ... Display unit, 200 ... Machine learning model, 201 ... Input layer, 202 ... Connection Layer, 203 ... Output layer, 210 ... Encoder section, 220 ... Decoder section, 300 ... Image processing program

Claims (6)

It is a method of performing an inspection of the image to be inspected.
(A) A step of creating a converted image by performing image conversion for removing noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit.
(B) A step of generating measurement results for one or more inspection items by executing image processing on the converted image, and
(C) A step of generating a judgment result regarding the inspection item by comparing the measurement result with a predetermined judgment standard, and a step of generating the judgment result.
Including, how. The method according to claim 1.
The method, wherein the machine learning model is configured as an autoencoder, U-Net, or CycleGAN. The method according to claim 1 or 2, further
Before the step (a),
The first learning process for executing the learning of the machine learning model and
After the first learning step, a second learning step of optimizing the parameters of the image processing and
Have,
The second learning step is
(I) A step of inputting a learning image into the trained machine learning model to generate an inference image, and
(Ii) A step of generating a learning measurement result regarding the inspection item by executing the image processing on the inferred image, and
(Iii) A step of performing optimization of the parameters of the image processing by using the difference between the true value for the inspection item and the measurement result at the time of learning in the learning image.
Including, how. The method according to claim 3.
The second learning step is performed after the step (i).
A step of performing additional learning of the machine learning model using the inference image,
Including, how. An information processing device that inspects the image to be inspected.
A memory that stores a machine learning model including a convolutional neural network having an encoder part and a decoder part,
A processor that executes operations using the machine learning model, and
Equipped with
The processor
(A) A process of creating a converted image by executing an image conversion for removing noise of the image to be inspected using the machine learning model.
(B) A process of generating measurement results for one or more inspection items by executing image processing on the converted image, and a process of generating measurement results.
(C) A process of generating a judgment result regarding the inspection item by comparing the measurement result with a predetermined judgment standard.
An information processing device that is configured to run. A computer program that causes the processor to inspect the image to be inspected.
The computer program is
(A) A process of creating a converted image by performing image conversion for removing noise of the image to be inspected using a machine learning model including a convolutional neural network having an encoder unit and a decoder unit.
(B) A process of generating measurement results for one or more inspection items by executing image processing on the converted image, and a process of generating measurement results.
(C) A process of generating a judgment result regarding the inspection item by comparing the measurement result with a predetermined judgment standard.
A computer program that causes the processor to execute.

Images