Information processing method, information processing device, inspection method, and inspection device

Abstract

This invention provides an information processing method, information processing device, inspection method, and inspection device that improve inspection accuracy by using a method unaffected by variations in image brightness and by reducing the amount of machine learning data required. [Solution] The information processing device (10) includes a first machine learning model creation unit (11b) that creates a first machine learning model (13a) that generates a contour image (oi) which creates the contour of a virtual product from a virtual product image (ii) which depicts a virtual product that imitates a product (110), and a second machine learning model creation unit (11d) that creates a second machine learning model (13b) that generates a contour estimation image (ei) which restores the contour of the virtual product for each mask position using a mask image (mi) which is a part of the contour image.

Description

Technical Field

[0001] The present invention relates to an information processing method and an information processing apparatus for creating a learned model for inspecting products using X-rays, and an inspection method and an inspection apparatus using the learned model.

Background Art

[0002] There is known an X-ray inspection apparatus that irradiates a product with X-rays to obtain an X-ray transmission image and inspects the product based on the light and shade levels of the image. For example, in Patent Document 1, there is disclosed an inspection apparatus that detects a defective seal bite by comparing a sample image subjected to X-ray impermeable processing with an actual X-ray transmission image on one of the packaging area and the seal area of a bagged product. In Patent Document 2, there is disclosed an inspection apparatus that inspects the presence or absence of foreign matter and the number of products by machine learning the coordinates and dimensions of the overlapping portions of products in an X-ray transmission image for products such as bagged winners in which a plurality of products are packaged.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

[0004]

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] When photographing a product such as a bagged winner with an X-ray inspection apparatus, the overlapping portion becomes darker than other portions, making it difficult to determine which part is a single winner. When performing an inspection using the area area binarized with a brightness threshold, the inspection may be affected by variations in brightness and may be difficult. To machine-learn such an inspection target, it is necessary to prepare a large number of actual X-ray transmission images in order to improve inspection accuracy.

[0006] This invention was made in view of these circumstances, and aims to provide an information processing method, information processing device, inspection method, and inspection device that improve inspection accuracy using a method unaffected by variations in image brightness and that incorporates features to reduce the amount of machine learning material. [Means for solving the problem]

[0007] To solve the above problems, the first aspect of the present invention is an information processing method executed by a computer, characterized by comprising: a first machine learning step of creating a first machine learning model that generates a contour image created from a virtual product image in which a virtual product that imitates a real product is depicted; and a second machine learning step of creating a second machine learning model that generates a contour estimation image by moving a mask image obtained by masking a part of the contour image and restoring the contour of the virtual product for each mask position.

[0008] A second aspect of the present invention provides an inspection method comprising: a first inspection step of inputting an X-ray transmission image of the product into the first machine learning model and the second machine learning model created by the information processing method described in the first aspect, in order to create a contour image of the product; a second inspection step of inputting the contour image of the product into the second machine learning model, using a mask image that masks a part of the contour image, moving the mask, and creating a contour estimation image that restores the contour of the product for each mask position; and a third inspection step of comparing the contour image created in the first inspection step with the contour estimation image created in the second inspection step to perform an inspection of the product.

[0009] In the third embodiment of the information processing method, it is also preferable to further include an additional learning step in which, in the first embodiment, an X-ray transmission image of a sample of the product is input to the first machine learning model to create a contour image of the sample, the contour image of the sample is input to the second machine learning model, a mask image is created by masking a part of the contour image, the mask is moved, and a contour estimation image is created by restoring the contour of the sample for each mask position.

[0010] In the inspection method of the fourth embodiment, in the second embodiment, it is also preferable to take the difference between the contour image and the contour estimation image in the third inspection step to inspect for any or all of the following: the shape of the product, the number of products, and the presence or absence of foreign matter.

[0011] In the fifth embodiment of the inspection method, it is also preferable to fix the movement of the mask in accordance with the characteristics of the product, as in the second or fourth embodiment.

[0012] In the sixth embodiment of the inspection method, in any of the second, fourth, or fifth embodiments, it is also preferable that the movement of the mask be performed by gradually changing the size and / or sliding width of the mask from large to small.

[0013] An information processing apparatus according to a seventh aspect of the present invention is characterized by comprising: a first machine learning model creation unit that creates a first machine learning model that generates a contour image created from a virtual product image in which a virtual product that imitates a product is depicted; and a second machine learning model creation unit that creates a second machine learning model that generates a contour estimation image by moving a mask image which is a part of the contour image, and restoring the contour of the virtual product for each mask position.

[0014] An inspection apparatus according to the eighth aspect of the present invention is characterized by comprising: an X-ray irradiation unit for irradiating a product with X-rays; an X-ray detection unit for acquiring an X-ray transmission image of the product; a first inspection unit for inputting the X-ray transmission image of the product into a first machine learning model created by an information processing device according to the seventh aspect to create a contour image of the product; a second inspection unit for inputting the contour image of the product into a second machine learning model created by an information processing device according to the seventh aspect, and for moving a mask image that masks a part of the contour image to restore the contour of the product to create a contour estimation image; and a third inspection unit for comparing the contour image created by the first inspection unit with the contour estimation image created by the second inspection unit to perform an inspection of the product. [Effects of the Invention]

[0015] According to the present invention, it is possible to provide an information processing method, an information processing device, an inspection method, and an inspection device that improve inspection accuracy using a method unaffected by variations in image brightness and that incorporates features to reduce the amount of machine learning material. [Brief explanation of the drawing]

[0016] [Figure 1] This is a block diagram of the configuration of an information processing device 10 according to an embodiment of the present invention. [Figure 2] This figure shows an example of training for the first machine learning model 13a. [Figure 3] This figure shows an example of training for the second machine learning model 13b. [Figure 4] This figure shows another training example of the second machine learning model 13b. [Figure 5] This is a flowchart of the information processing method by the information processing device 10. [Figure 6] This diagram illustrates an example of additional learning for the information processing device 10. [Figure 7] This is a block diagram of the configuration of the inspection device 100 according to an embodiment of the present invention. [Figure 8]It is a flowchart of an inspection method by the inspection device 100. [Figure 9] It is a diagram showing an example of inspection by the inspection device 100. [Figure 10] It is a diagram showing another example of inspection by the inspection device 100.

Embodiments for Carrying Out the Invention

[0017] Next, preferred embodiments of the present invention will be described based on the drawings.

[0018] (Information Processing Device) FIG. 1 is a configuration block diagram of an information processing device 10 according to an embodiment of the present invention. The information processing device 10 is a device that creates a learned model used in an inspection device 100 described later. The information processing device 10 includes an arithmetic unit 11, a main storage device 12, an auxiliary storage device 13, and various user interfaces 14.

[0019] The arithmetic unit 11 is composed of one or more CPUs (Central Processing Unit), multi-core CPUs, or GPUs (Graphics Processing Unit), etc. The arithmetic unit 11 is connected to each hardware part constituting the information processing device 10 via a bus. The main storage device 12 and the auxiliary storage device 13 are composed of storage media such as ROM (Read Only Memory), RAM (Random Access Memory), and hard disks. The arithmetic unit 11 executes a program stored in the auxiliary storage device 13 using the main storage device 12 as a work area. The user interface 14 is a part where an operator gives instructions to the information processing device 10. The user interface 14 includes, for example, operation buttons, operation switches, or a keyboard, and has a display unit and its graphical user interface.

[0020] The calculation unit 11 includes, as functional units, a virtual image acquisition unit 11a, a first machine learning model creation unit 11b, a mask image creation unit 11c, and a second machine learning model creation unit 11d. Each functional unit may be configured with hardware such as an ASIC (Application-Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array).

[0021] Figure 2 shows an example of training for the first machine learning model 13a, where group (A) is an example of a virtual product image ii, and group (B) is an example of a contour image oi generated from the virtual product image ii in (A).

[0022] The virtual image acquisition unit 11a acquires a virtual product image ii (see Figure 2(A)) which depicts a simulated product 110 to be inspected by the inspection device 100 described later (hereinafter referred to as "virtual product"). One or more virtual products are depicted, and it is sufficient that they are drawn to mimic the shape of the product 110. The virtual product image ii may be created using application software such as graphic software or drawing software, or it may be photographic data of the product 110. The virtual product image ii may be created by the information processing device 10, acquired by reading data created on another device, or acquired via communication if the information processing device 10 is equipped with a communication unit.

[0023] The first machine learning model creation unit 11b trains the first machine learning model 13a to generate a contour image oi (see Figure 2(B)) by binarizing the virtual product image ii acquired by the virtual image acquisition unit 11a, setting the contour to "1" and the background to "0," resulting in a white contour image of the virtual product against a black background. The first machine learning model creation unit 11b also trains the first machine learning model 13a to output the number of virtual products from the contour image oi, based on the basic shape of the virtual products. In the example in Figure 2, the basic shape of the virtual product is a long, narrow closed ellipse, so in example (1) of Figure 2, the number of products is "2," and in example (2) of Figure 2, the number of products is "3." The first machine learning model 13a generates the contour image oi using, or in combination with, known estimation models such as support vector regression, multiple regression analysis, neural networks, and deep learning, or edge detection processes such as differential filtering. The first machine learning model 13a functions as a classifier created as a result of machine learning. The first machine learning model 13a is stored in the auxiliary memory 13, and the generated contour image oi and the number of virtual products are stored in the main memory 12.

[0024] Figure 3 shows an example of training for the second machine learning model 13b, where group (A) is an example of a mask image mi, and group (B) is an example of a contour estimation image ei generated from the mask image mi of (A).

[0025] The mask image creation unit 11c reads the contour image oi generated by the first machine learning model 13a and creates a mask image mi by masking a part of the contour image oi. As shown in Figure 3(A), the mask image creation unit 11c creates multiple mask images mi(1), mi(2), ... mi(x) from a single contour image oi by sliding a rectangular mask M, for example. The mask M is preferably set to a black rectangle the same color as the background, but is not limited to this as long as it is a shape and color suitable for contour reconstruction. The size of the mask M is preferably set to a size that roughly matches the height × width of the basic shape of the virtual product, for example, but may be suitably changed depending on the inspection application. The movement (slide width) of the mask M is, for example, scanning horizontally or vertically while shifting it at a pitch of about half the size of the mask M, but may be suitably changed. The learning accuracy and estimation accuracy can be controlled by changing the size and / or slide width of the mask M during the learning phase.

[0026] The second machine learning model creation unit 11d trains the second machine learning model 13b to generate contour estimation images ei (see Figure 3(B)) that reconstruct the contours of virtual products from the mask image mi created by the mask image creation unit 11c. The second machine learning model 13b generates contour estimation images ei using or in combination with known estimation models such as support vector regression, multiple regression analysis, neural networks, and deep learning. The second machine learning model creation unit 11d trains the second machine learning model 13b to generate contour estimation images ei(1), ei(2), ... ei(x) for each mask position, i.e., for each mask image mi(1), mi(2), ... mi(x). The second machine learning model creation unit 11d also trains the second machine learning model 13b to output the number of virtual products from the contour estimation images ei, based on the basic shape of the virtual products. The second machine learning model 13b functions as a classifier created as a result of machine learning. The second machine learning model 13b is stored in the auxiliary memory 13, and the contour estimation image ei and the number of virtual products are stored in the main memory 12.

[0027] The second machine learning model creation unit 11d trains the second machine learning model 13b on various patterns. Figure 4 shows another example of training the second machine learning model 13b, where group (A) is an example of a mask image mi, and group (B) is an example of a contour estimation image ei generated from the mask image mi in (A), showing an example of training on more complex shapes. Example (1) in Figure 4 is a pattern where two virtual products overlap, and example (2) in Figure 4 is a pattern where three virtual products overlap. The mask image creation unit 11c also moves the mask to create multiple mask images mi(1), mi(2), ... mi(x) for these, and the second machine learning model 13b learns to estimate contour estimation images ei(1), ei(2), ... ei(x) for each mask position.

[0028] (Information processing methods) Figure 5 is a flowchart of the information processing method by the information processing device 10, and is a learning flowchart by the information processing device 10. When information processing starts, the process first moves to step S1, where the virtual image acquisition unit 11a acquires a virtual product image ii.

[0029] Next, the process moves to step S2, where the first machine learning model creation unit 11b inputs the virtual product image ii to the first machine learning model 13a, and the first machine learning model 13a generates a contour image oi. The first machine learning model creation unit 11b receives this output and stores it in the main memory 13.

[0030] Next, the process moves to step S3, where the mask image creation unit 11c creates multiple mask images mi(1), mi(2), ... mi(x) by moving the mask from a single contour image oi.

[0031] Next, the process moves to step S4, where the second machine learning model creation unit 11d moves the mask, that is, inputs the mask images mi(1), mi(2), ... mi(x) with the moved mask position into the second machine learning model 13b.

[0032] Next, the process moves to step S5, where the second machine learning model 13b generates contour estimation images ei(2), ei(3), ..., ei(x), and the second machine learning model creation unit 11d receives these outputs and stores them in the main memory 13.

[0033] Next, in step S6, the second machine learning model creation unit 11d determines whether the mask movement is complete, that is, whether estimation is complete for all mask images mi(1), mi(2), ... mi(x). If it is not complete (NO), it returns to step S4; if it is complete (YES), it terminates the information processing.

[0034] According to the above information processing device 10 and information processing method, firstly, pre-training for products (objects to be inspected) is performed by identifying contours (difference in brightness from neighbors) rather than image brightness, so the influence of areas with small brightness differences can be reduced. Also, since contour images are handled (binarization where contours are "1" and backgrounds are "0"), processing time can be shortened. Secondly, the first machine learning model 13a for contour generation is trained on virtual product image ii, so there is no need to prepare a large number of actual X-ray transmission images as training material, and sufficient pre-training can be performed with virtual product image ii. In addition, by applying a mask to the training material and moving the mask, many training images can be obtained from a single training material, thus reducing the amount of training material required.

[0035] As a modification (1) of the information processing method in Figure 5, it is preferable to further train the first machine learning model 13a and the second machine learning model 13b with the X-ray transmission image pti of the sample product 110. Figure 6 shows an example of additional training, where (A) is an example of the sample X-ray transmission image pti, (B) is an example of the contour image oi generated from the X-ray transmission image pti of (A), and (C) is an example of the mask image mi of the contour image oi of (B). By adopting this modification, the accuracy of the contour estimation image ei can be further improved. Since the information processing device 10 performs a sufficient number of pre-trainings with the virtual product image ii, training on the sample X-ray transmission image pti is sufficiently effective if it is done for about one-tenth of the number of virtual product images.

[0036] As a preferred modification (2) of the information processing method in Figure 5, in step S4, it is also preferable to move the mask M by gradually changing the size and / or sliding width of the mask M from large to small. This allows for detailed checking of the area containing the virtual product after narrowing it down, thereby improving learning accuracy and learning speed.

[0037] Furthermore, with this embodiment of the information processing device 10 and information processing method, learning is possible using only OK products without learning from NG products that do not meet quality standards, thus reducing the amount of learning material. In addition, the performance of the trained models, namely the first machine learning model 13a and the second machine learning model 13b, can be sufficiently verified not only with actual X-ray transmission images but also with virtual product images ii as input.

[0038] (Inspection device) Figure 7 is a block diagram of the configuration of an inspection apparatus 100 according to an embodiment of the present invention. The inspection apparatus 100 is an X-ray inspection apparatus that utilizes the first machine learning model 13a and the second machine learning model 13b created by the information processing apparatus 10 described above. The inspection apparatus 100 comprises a transport unit 120, an X-ray irradiation unit 122, an X-ray detection unit 124, an operation unit 126, a display unit 128, a control unit 130, and a storage unit 140.

[0039] The transport unit 120 is a belt conveyor that transports the product 110 to be inspected, and the transport speed and other parameters are controlled by the control unit 130. An X-ray irradiation unit 122 is provided above the transport unit 120, and X-rays are irradiated downward from this X-ray irradiation unit 122. An X-ray detection unit 124 (line sensor, etc.) is provided below the belt of the transport unit 120, and this X-ray detection unit 124 detects the X-rays that have passed through the product 110. The X-ray detection unit 124 is connected to the control unit 130, and the detection signal from the X-ray detection unit 124 is input to the control unit 130. In the control unit 130, an X-ray transmission image Xti of the product 110 is created from the detection signal using known image processing and stored in the storage unit 140. The operation unit 126 is equipped with, for example, operation buttons, operation switches, or a keyboard, and the display unit 128 is equipped with a liquid crystal or organic EL display, and the display unit 128 displays a graphical user interface for inspection.

[0040] The control unit 130 is composed of one or more CPUs (Central Processing Units), multi-core CPUs, or GPUs (Graphics Processing Units), and is connected to the various hardware components of the inspection device 100 via a bus. The storage unit 140 is composed of storage media such as ROM (Read Only Memory), RAM (Random Access Memory), and hard disks, and stores the programs of the first machine learning model 13a and the second machine learning model 13b created by the information processing device 10.

[0041] The control unit 130 controls the transport unit 120 and the X-ray irradiation unit 122, and also includes a first inspection unit 131, a second inspection unit 132, and a third inspection unit 133 as functional units. Each functional unit may be composed of hardware such as an ASIC (Application-Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array).

[0042] The first inspection unit 131 acquires an X-ray transmission image Xti of product 110 from the X-ray detection unit 124, inputs it to the first machine learning model 13a, and receives an output of a contour image Xoi of product 110 from the first machine learning model 13a. The contour image Xoi is stored in the work area of ​​the storage unit 140.

[0043] The second inspection unit 132, similar to the mask image creation unit 11c of the information processing device 10, reads the contour image oi generated by the first machine learning model 13a, masks a portion of the contour image Xoi, and creates mask images Xmi(1), Xmi(2), ... Xmi(x) by moving the mask M, which are then input to the second machine learning model 13b. The second machine learning model 13b then outputs contour estimation images Xei(1), Xei(2), ... Xei(x) which restore the contour of the product 110. Each contour estimation image Xei is stored in the work area of ​​the storage unit 140.

[0044] The third inspection unit 133 compares the contour image Xoi of product 110 created by the first inspection unit 131 with the contour estimation image Xei of product 110 created by the second inspection unit 132 to perform an inspection of product 110. The inspection performed by the third inspection unit 133 will be described later.

[0045] (Testing method) Figure 8 is a flowchart of the inspection method using the inspection device 100. When the inspection is started, the process first proceeds to step S101, where an X-ray transmission image Xti of the product 110 is acquired. Specifically, the product 110 is transported by the transport unit 120, the X-ray irradiation unit 122 irradiates it with X-rays, the X-ray detection unit 124 detects the X-rays, and the control unit 130 creates the X-ray transmission image Xti.

[0046] Next, the process moves to step S102, where the first inspection unit 131 inputs the X-ray transmission image Xti of product 110 into the first machine learning model 13a, and the first machine learning model 13a generates a contour image Xoi of product 110. The first inspection unit 131 receives this output and stores it in the storage unit 140.

[0047] Next, the process moves to step S103, where the second inspection unit 132 creates multiple mask images Xmi(1), Xmi(2), ... Xmi(x) by moving the mask from the contour image Xoi of product 110.

[0048] Next, the process moves to step S104, where the second inspection unit 132 moves the mask, that is, inputs the mask images Xmi(1), Xmi(2), ... Xmi(x) with the moved mask position into the second machine learning model 13b.

[0049] Next, the process moves to step S105, where the second machine learning model 13b generates contour estimation images Xei(2), Xei(3), ... Xei(x) of the product 110, and the second inspection unit 132 receives these outputs and stores them in the storage unit 140.

[0050] Next, in step S106, the second inspection unit 132 determines whether the mask movement is complete, that is, whether estimation is complete for all mask images Xmi(1), Xmi(2), ... Xmi(x). If it is not complete (NO), it returns to step S104; if it is complete (YES), it proceeds to step S107.

[0051] When the process moves to step S107, the third inspection unit 133 performs an inspection by taking the image difference between the contour image Xoi created in step 102 and the contour estimation images Xei(1), Xei(2), ... Xei(x) created in step S105. The third inspection unit 133 performs one or all of the following inspections (i) to (iii). (i) Inspect the quantity of product 110. The quantity will be measured by the number of basic shapes. The contour image Xoi and the contour estimation image Xei are each checked to see if they meet the specified number and if the numbers match. If the number is not the specified number, or if the two do not match, it is detected as an anomaly. . (ii) Detect the presence or absence of a foreign object. If a foreign object is present, the contour of the foreign object is created in the contour image Xoi in step 102, but since the shape of the foreign object has not been learned by machine learning, the contour of the foreign object is not created in the contour estimation image Xei in step S104. The presence or absence of a foreign object is detected by taking the difference between these images. (iii) Detect any deformities in product 110. If there is a deformity, the contour image Xoi in step 102 will have a contour outline of the deformed product 110, but in the contour estimation image Xei in step S105, product 110 will be restored to its normal shape by machine learning. The presence or absence of a deformity is detected by taking the difference between these images. In step S108, the third inspection unit 133 provides inspection reports for inspections (i) to (iii) via the display unit 128 and terminates the inspection.

[0052] Figure 9 shows an example of inspection by the inspection device 100, where (A) is an example of an X-ray transmission image Xti of product 110, (B) is an example of a contour image Xoi generated from the X-ray transmission image Xti of (A), (C) is an example of a mask image Xmi of the contour image Xoi of (B), and (D) is an example of a contour estimation image Xei of product 110 estimated from the mask image Xmi of (C).

[0053] The inspected product 110 is a bag of sausages, with multiple sausages overlapping. As shown in (A), one of the sausages has a missing BP (point of birth). As shown in (B), the contour of the missing BP is created in the contour image Xoi. (C) is a mask image Xmi with a mask M applied to the area of ​​the missing BP, and (D) is the contour estimation image Xei created from this mask image Xmi. By taking the image difference between (B) and (D), the third inspection unit 133 detects the missing BP of product 110.

[0054] Figure 10 shows another example of inspection by the inspection device 100, where (A) is an example of an X-ray transmission image Xti of product 110, (B) is an example of a contour image Xoi generated from the X-ray transmission image Xti of (A), (C) is an example of a mask image Xmi of the contour image Xoi of (B), and (D) is an example of a contour estimation image Xei of product 110 estimated from the mask image Xmi of (C). As shown in (A), a foreign object C is mixed into the bag. As shown in (B), the contour of the foreign object C is created in the contour image Xoi. (C) is a mask image Xmi with a mask M applied to a part of the foreign object C, and (D) is the contour estimation image Xei created from this mask image Xmi. By taking the image difference between (B) and (D), the third inspection unit 133 detects the foreign object C.

[0055] As a variation (1) of the inspection method in Figure 8, similar to the variation (2) of the information processing method, in step S104, it is also preferable to move the mask M by gradually changing the size and / or sliding width of the mask M from large to small. This allows for a detailed check of a specific area of ​​the product 110 after narrowing it down, thereby improving inspection accuracy and speed.

[0056] As a modification (2) of the inspection method in Figure 8, it is also preferable to fix the movement of the mask M depending on the characteristics of the product 110. For example, for products such as boxed cosmetics, where the position of the product is unlikely to change, the inspection can be further sped up by fixing the position of the mask M and inspecting only specific positions.

[0057] Furthermore, for inspections that detect general anomalies, the mask M can be set larger than the basic shape of product 110, while for inspections that detect minor anomalies, the mask M can be set smaller than the basic shape of product 110. In this way, inspection accuracy and estimation accuracy can be controlled by changing the size and / or slide width of the mask M during inspection.

[0058] While preferred embodiments and variations of the present invention have been described above, these can be modified and combined based on the knowledge of those skilled in the art, and such forms are also included within the scope of the present invention. [Explanation of symbols]

[0059] 10…Information processing device, 11…Calculation unit, 11b…First machine learning model creation unit, 11d…Second machine learning model creation unit, 13a…First machine learning model, 13b…Second machine learning model, 100…Inspection device, 110…Product, 122…X-ray irradiation unit, 124…X-ray detection unit, 130…Control unit, 131…First inspection unit, 132…Second inspection unit, 133…Third inspection unit, ii…Virtual product image, oi…Contour image, mi…Mask image, ei…Contour estimation image, pti…X-ray transmission image of sample, Xti…X-ray transmission image of product, Xoi…Contour image of product, Xmi…Mask image of product, Xei…Contour estimation image of product

Claims

1. A method of information processing performed by a computer, A first machine learning step involves creating a first machine learning model that generates a contour image from a virtual product image depicting a virtual product that mimics the product, and creating a contour image that creates the outline of the said virtual product. A second machine learning step involves creating a second machine learning model that generates a contour estimation image by moving a mask image obtained by masking a portion of the contour image, and then reconstructing the contour of the virtual product for each mask position. An information processing method characterized by having the following features.

2. Using the first machine learning model and the second machine learning model created by the information processing method described in claim 1, A first inspection step involves inputting an X-ray transmission image of the product into the first machine learning model to create a contour image of the product. A second inspection step involves inputting the contour image of the product into the second machine learning model, using a mask image obtained by masking a portion of the contour image, moving the mask, and creating a contour estimation image by restoring the contour of the product for each mask position. An inspection method characterized by comprising: a third inspection step, which compares a contour image created in the first inspection step with a contour estimation image created in the second inspection step to perform an inspection of the product.

3. moreover, The information processing method according to claim 1, further comprising an additional learning step in which an X-ray transmission image of a sample of the product is input to the first machine learning model to create a contour image of the sample, the contour image of the sample is input to the second machine learning model, a mask image is created by masking a part of the contour image, the mask is moved, and a contour estimation image is created by restoring the contour of the sample for each mask position.

4. The inspection method according to claim 2, wherein in the third inspection step, the difference between the contour image and the contour estimation image is taken to inspect for any or all of the following: the shape of the product, the number of products, and the presence or absence of foreign matter.

5. The inspection method according to claim 2, characterized in that the movement of the mask is fixed according to the characteristics of the product.

6. The inspection method according to claim 2, characterized in that the movement of the mask is performed by gradually changing the size and / or sliding width of the mask from large to small.

7. A first machine learning model creation unit creates a first machine learning model that generates a contour image from a virtual product image depicting a virtual product that imitates the product, and A second machine learning model creation unit creates a second machine learning model that generates a contour estimation image by moving a mask image obtained by masking a part of the contour image and restoring the contour of the virtual product for each mask position. An information processing device characterized by comprising:

8. An X-ray irradiation unit that irradiates the product with X-rays, An X-ray detection unit for acquiring an X-ray transmission image of the aforementioned product, A first inspection unit that inputs the X-ray transmission image of the product into the first machine learning model created by the information processing device described in claim 7 to create a contour image of the product, A second inspection unit inputs the contour image of the product into the second machine learning model created by the information processing device described in claim 7, moves a mask image obtained by masking a part of the contour image, and creates a contour estimation image by restoring the contour of the product. The third inspection unit compares the contour image created by the first inspection unit with the contour estimation image created by the second inspection unit to perform an inspection of the product. An inspection device characterized by being equipped with the following features.

Images