Surface inspection method, manufacturing method of object, quality management method of object, generation method of learning model, surface inspection device, manufacturing facility of object, and learned model

Abstract

To provide a surface inspection method, a manufacturing method of an object, a quality management method of an object, a generation method of a learning model, a surface inspection device, a manufacturing facility of an object, and a learned model capable of precisely detecting a defect in which the detection has been difficult by a conventional surface inspection method.SOLUTION: A surface inspection method includes an image generation step in which image generation means of a computer inputs a first image in which an inspection object has been imaged to a learned model and outputs a second image from the learned model, and a defect candidate detection step in which defect candidate detection means of the computer detects a defect candidate section of an inspection object on the basis of the first image and the second image, and the learned model is a model subjected to machine learning with a first image captured according to a first optical condition as an input value and a second image captured according to the first optical condition, or a second optical condition different from the first optical condition as an output value.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a surface inspection method, an object manufacturing method, an object quality control method, a learning model generation method, a surface inspection device, an object manufacturing facility, and a learned model. [Background technology]

[0002] Generally, in steel strip production lines, inspection of surface defects is carried out for quality assurance and quality control of product surfaces. In recent years, the introduction of surface inspection devices has progressed for the inspection of surface defects, aiming to automate inspection and reduce labor. In general surface inspection devices, defects present on the non-inspection surface are optically detected by irradiating the non-inspection surface with light from an illumination source and capturing an image of the reflected light. In such surface inspection methods, a line light source and a line sensor are often used to capture images of both specularly reflected light and diffusely reflected light (see Patent Document 1).

[0003] In such surface inspection devices, in order to detect target surface defects, measures are taken to increase the contrast between surface defects and normal areas in the acquired image, i.e., the S / N ratio (signal-to-noise ratio). Examples of such measures include adjusting the lighting and camera angle, and using optical filters (see Patent Document 2).

[0004] On the other hand, there are many different types of surface defects on steel strips, and there are few optical systems that can obtain a good S / N ratio for all of them.In addition to the optical system using lighting and imaging devices, there are also cases where innovations are made in the image correction section.

[0005] For example, Patent Document 3 discloses the following surface inspection method. In this method, first, a plurality of texture feature images are generated by filtering an original image using a plurality of spatial filters. Next, for each position on the original image, values ​​at corresponding positions in the plurality of texture feature images are extracted to generate a feature vector at each position on the original image. Next, for each feature vector, the degree of anomaly in the multidimensional distribution formed by the feature vector is calculated, and an anomaly image indicating the degree of anomaly for each position on the original image is generated. Next, in this anomaly image, portions where the degree of anomaly exceeds a predetermined value are detected as defective parts or defect candidate parts.

[0006] Recently, several methods have been proposed for using images generated using neural networks in surface inspection. For example, Patent Document 4 discloses a method for generating artificial images of non-defective products and using them for inspection. Furthermore, Patent Document 5 discloses a method for generating multiple images with defects from an original image and using them for inspection. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2008-145373 [Patent Document 2] Patent Publication No. 2021-169949 [Patent Document 3] Japanese Patent Application Laid-Open No. 2018-155690 [Patent Document 4] Patent Publication No. 2021-89219 [Patent Document 5] Japanese Patent Application Publication No. 2023-76347 Summary of the Invention [Problem to be solved by the invention]

[0008] However, since there are a wide variety of forms of surface defects that occur on steel strips, it is difficult for the surface inspection methods disclosed in Patent Documents 1 to 5 to achieve good detection performance for all of them.

[0009] For example, as in Patent Documents 1 and 2, it is common to use techniques such as adjusting the angle of the lighting or imaging device or using optical filters to detect defects on a specific surface or with a specific shape, but this does not result in good detection performance for defects of other shapes.

[0010] Furthermore, Patent Document 3 discloses a technology that enables the detection of surface defects with a low S / N ratio that exist in an image captured by a predetermined optical system using an image processing system. The technology disclosed in Patent Document 3 has the problem that it is not possible to detect defective parts unless the captured image is captured so that normal parts and defective parts have different textures.

[0011] The present invention has been made in consideration of the above, and aims to provide a surface inspection method, an object manufacturing method, an object quality control method, a learning model generation method, a surface inspection device, an object manufacturing facility, and a learned model that can accurately detect defects that were difficult to detect using conventional surface inspection methods. [Means for solving the problem]

[0012] (1) A surface inspection method according to the present invention comprises: An image generation step in which an image generation means provided in a computer inputs a first image of an object to be inspected into a trained model and outputs a second image from the trained model; a defect candidate detection step in which a defect candidate detection means included in the computer detects a defect candidate portion of the inspection object based on the first image and the second image; Including, The trained model is a model that has been subjected to machine learning so that the first image captured under first optical conditions is used as an input value, and the second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value.

[0013] (2) A surface inspection method according to the present invention is the surface inspection method according to (1) above, the first optical condition and the second optical condition include any one of an angle of an illumination device that irradiates light onto the inspection object, an angle of an imaging device that images the inspection object, whether or not an optical filter is used, and a type of the optical filter; In the image generation step, the type of the first optical condition of the first image input to the trained model is the same as the type of the second optical condition of the second image output from the trained model.

[0014] (3) A method for manufacturing an object according to the present invention inspects the surface of an object by the surface inspection method described in (1) or (2) above, and manufactures the object based on the inspection results.

[0015] (4) A quality control method for an object according to the present invention inspects the surface of an object by the above (1) or (2), and controls the quality of the object based on the inspection results.

[0016] (5) The method for generating a learning model according to the present invention involves a model generation means provided in a computer generating a trained model that has undergone machine learning, using a first image captured under first optical conditions as an input value and a second image captured under the first optical conditions or second optical conditions different from the first optical conditions as an output value.

[0017] (6) A surface inspection device according to the present invention includes: an image generation unit that inputs a first image of an object to be inspected into a trained model and outputs a second image from the trained model; a defect candidate detection unit that detects a defect candidate portion of the inspection object based on the first image and the second image; Equipped with The trained model is a model that has been subjected to machine learning so that the first image captured under first optical conditions is used as an input value, and the second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value.

[0018] (7) The surface inspection device according to the present invention is the surface inspection device described in (6) above, the first optical condition and the second optical condition include any one of an angle of an illumination device that irradiates light onto the inspection object, an angle of an imaging device that images the inspection object, whether or not an optical filter is used, and a type of the optical filter; In the image generation unit, the type of the first optical condition of the first image input to the trained model is the same as the type of the second optical condition of the second image output from the trained model.

[0019] (8) An object manufacturing facility according to the present invention includes the surface inspection device according to (6) or (7) above.

[0020] (9) The trained model of the present invention is one that has been subjected to machine learning so that a first image captured under first optical conditions is used as an input value, and a second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value. [Effects of the Invention]

[0021] The surface inspection method, surface inspection device, and trained model according to the present invention enable accurate detection of defects that are difficult to detect using conventional surface inspection methods. Furthermore, the trained model generation method according to the present invention enables generation of a trained model that can accurately detect surface defects on an inspection target. Furthermore, the object manufacturing method and object manufacturing equipment according to the present invention enable manufacturing of objects free of surface defects. Furthermore, the object quality control method according to the present invention enables appropriate management of the quality of manufactured objects. [Brief explanation of the drawings]

[0022] [Figure 1] FIG. 1 is a block diagram showing an example of the configuration of a surface inspection apparatus according to an embodiment. [Figure 2] FIG. 2 is a schematic diagram showing an overview of an image generating step performed by the image generating unit of the surface inspecting device according to the embodiment. [Figure 3] FIG. 3 is a diagram showing an example of learning images captured under a plurality of optical conditions in the model generation step by the model generation device. [Figure 4] FIG. 4 is a diagram showing an example of a surface inspection device according to an embodiment in which a specular reflection camera and a diffuse reflection camera are arranged as imaging devices. [Figure 5] FIG. 5 is a diagram showing an example of the operation of the image generating unit in the surface inspecting device according to the embodiment. [Figure 6] FIG. 6 is a flowchart showing the flow of the surface inspection method executed by the surface inspection device according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0023] A surface inspection method, an object manufacturing method, an object quality control method, a learning model generation method, a surface inspection device, an object manufacturing facility, and a learned model according to embodiments of the present invention will be described with reference to the drawings.

[0024] The components in the following embodiments include those that are easily replaceable by those skilled in the art, or those that are substantially identical. In the following description, descriptions of identical or overlapping parts will be omitted or simplified as appropriate. In the drawings referred to below, identical or overlapping parts are denoted by the same reference numerals.

[0025] (Surface inspection equipment) A surface inspection device according to an embodiment will be described with reference to Figs. 1 to 4. The surface inspection device according to the embodiment is for detecting surface defects of an inspection object based on an image of the inspection object. An example of an object to be inspected by the surface inspection device according to the embodiment is a steel strip S as shown in Fig. 1.

[0026] 1, the surface inspection device 1 includes an imaging device 2, an illumination device 3, an image processing device 4, and a display device 5. The surface inspection device 1 may also include a model generation device 6, if necessary.

[0027] The imaging device 2 captures an image of the area on the steel strip S illuminated by the lighting device 3, and transmits the resulting image (original image) data of the surface of the steel strip S to the image processing device 4. The imaging device 2 may be, for example, either a line sensor camera with a one-dimensional imaging element or an area camera with a two-dimensional imaging element, and in either case, it captures images in synchronization with the transport of the steel strip S.

[0028] When the imaging device 2 is a line sensor camera, a continuous lighting device is used as the lighting device 3. When the imaging device 2 is an area camera, a flash lighting device is used as the lighting device 3, which emits a flash every time the steel strip S advances a certain distance.

[0029] The lighting device 3 illuminates the surface of the steel strip S that is the object of inspection by the surface inspection device 1. As the imaging device 2, either continuous lighting or flash lighting is used depending on the type of imaging device 2, as described above.

[0030] The image processing device 4 is realized by, for example, a general-purpose computer such as a workstation or a personal computer, or a server located in the cloud. The image processing device 4 analyzes the image data of the surface of the steel strip S transmitted from the imaging device 2. Then, the image processing device 4 detects surface defects (defect candidates) if any exist on the surface of the steel strip S, determines the type and harmfulness of those surface defects, and outputs that information to the display device 5.

[0031] Specifically, the image processing device 4 includes an image input unit 41, an image correction unit 42, an image generation unit 43, and a defect candidate detection unit 44. The image processing device 4 may also include a defect feature calculation unit 45 and a defect determination unit 46, as necessary.

[0032] The image input unit 41 has an internal temporary storage area, and sequentially buffers the image data of the surface of the steel strip S transmitted from the imaging device 2 in the temporary storage area.

[0033] The image correction unit 42 sequentially reads out image data from the temporary storage area of ​​the image input unit 41, and performs correction processing on the read out image data to generate a corrected image. In this correction processing, the image correction unit 42 performs the following processing.

[0034] First, the image correction unit 42 determines whether or not both or either of the edges of the steel strip S are included in the image. If both or either of the edges of the steel strip S are included in the image, the image correction unit 42 detects the position of the edge and sets the image area corresponding to the outside of the edge of the steel strip S as an area outside the inspection target. The image correction unit 42 also fills the area outside the inspection target with, for example, a mirror image of the internal area of ​​the steel strip S, with the edge position as the boundary. Next, the image correction unit 42 corrects (shading correction) brightness unevenness in the image of the steel strip S caused by uneven illuminance of the lighting device 3, etc., so that the brightness of the entire image is uniform.

[0035] The image generation unit 43 inputs an original image (hereinafter referred to as the "first image") of the steel strip S to be inspected into a trained model, and causes the trained model to output another image (hereinafter referred to as the "second image"). The second image is an image captured under the optical conditions (hereinafter referred to as the "first optical conditions") when the first image was captured, or an image captured under optical conditions (hereinafter referred to as the "second optical conditions") different from the first optical conditions.

[0036] Furthermore, as will be described later, the trained model is a model that has been subjected to machine learning so that a first image captured under first optical conditions is used as an input value, and a second image captured under the first optical conditions or a second image captured under second optical conditions is used as an output value. Note that the "second image captured under first optical conditions" is, for example, an image of another steel strip S captured under the same optical conditions as the first image (the types of optical conditions and the values ​​of each type are the same).

[0037] In addition, the trained model has been subjected to machine learning so that when an image of steel strip S taken under optical conditions that make it difficult to detect surface defects is input, it outputs an image of steel strip S taken under optical conditions that make it easy to detect surface defects.

[0038] In addition, the types of first optical conditions and second optical conditions include, for example, one of the following: the angle of the lighting device 3 that irradiates light onto the steel strip S to be inspected, the angle of the imaging device 2 that images the steel strip S to be inspected, whether or not an optical filter is used, and the type of optical filter.

[0039] Furthermore, the type of the first optical condition of the first image that the image generation unit 43 inputs to the trained model is the same as the type of the second optical condition of the second image that is output from the trained model. For example, if the first image is an image captured at "camera angle: A°", the second image will be an image captured at "camera angle: B°".

[0040] Furthermore, when the image generating unit 43 inputs a first image captured under first optical conditions into the trained model, if a second image captured under the same first optical conditions is output, the types and values ​​of the optical conditions (first optical conditions) of both images are the same. For example, if the first image is an image captured at "camera angle: A°", the second image will also be an image captured at "camera angle: A°". Note that the first image and the second image in this case have the same types and values ​​of the optical conditions, but are images of different steel strips S.

[0041] An overview of the processing by the image generation unit 43 will be described with reference to FIGS. 2 to 4. First, as shown in FIG. 2, images captured in advance under a plurality of different optical conditions (optical systems) are used as training images, and the image generation network is made to learn them. The training images are, for example, images that differ in the angle of the lighting device 3, the angle of the imaging device 2, whether or not an optical filter is used, the type of optical filter, etc. The training images may be captured while changing the optical conditions in actual manufacturing equipment, or may be captured using sample objects in offline experimental equipment.

[0042] By training the image generation network with a sufficient number of images, it operates as a trained model that inputs a first image captured under first optical conditions and outputs a second image captured under either the first optical conditions or the second optical conditions. A generative adversarial network may be used as the image generation network. The generative adversarial network generates an image using a generative model called a "Generator" and determines whether the generated image is correct using a discriminative model called a "Discriminator."

[0043] In this way, the image generation unit 43 is trained with a sufficient number of images in advance, and the image generation network that has become a trained model is used to generate an image equivalent to an image captured under different second optical conditions from an image captured under first optical conditions in an actual manufacturing facility.

[0044] Here, Fig. 3 shows examples of images taken under multiple optical conditions in an experimental facility as actual learning images. In this example, actual surface defects such as bare plating defects and scale defects were collected as samples, and images were taken by setting the angle of the lighting device 3 (hereinafter referred to as the "lighting angle") to 10° and changing the angle of the imaging device 2 (hereinafter referred to as the "camera angle") from 10° to 70° in 5° increments. Note that in Fig. 3, the two images labeled "appearance" are images of surface defects (bare plating defects and scale defects) in the collected samples. Furthermore, for the angle of the lighting device 3 (lighting angle) and the angle of the imaging device 2 (camera angle), the direction perpendicular to the surface of the object to be inspected is set to 0°.

[0045] As shown in Figure 3, for example, unplated defects are imaged differently from normal areas when the camera angle is in the range of 40° to 70°, making it possible to recognize the surface defects. On the other hand, scale defects are imaged differently from normal areas when the camera angle is in the range of 15° to 35°, making it possible to recognize the surface defects. Therefore, it can be seen that images captured at a camera angle of 40° to 70° are effective for detecting unplated defects, and images captured at a camera angle of 15° to 35° are effective for detecting scale defects, and that the effective optical conditions vary depending on the type of surface defect.

[0046] Therefore, the image generation unit 43 uses a trained model that has previously trained training images captured under various optical conditions to correct an image (first image) captured under optical conditions that make it difficult to detect surface defects into an image (second image) captured under optical conditions that make it easy to detect surface defects. This allows, for example, a unclear first image (e.g., an image captured at a camera angle other than 40° to 70°) in which it is difficult to detect non-plating defects to be converted into a clear second image (e.g., an image captured at a camera angle of 40° to 70°) as shown in FIG. 3 in which it is easy to detect non-plating defects. Furthermore, for example, a unclear first image (e.g., an image captured at a camera angle other than 15° to 35°) in which it is difficult to detect scale defects to be converted into a clear second image (e.g., an image captured at a camera angle of 15° to 35°) as shown in FIG. 3 in which it is easy to detect scale defects.

[0047] Fig. 4 shows an example of the installation of the imaging device 2 in an actual manufacturing facility. In an actual manufacturing facility, if the imaging device 2 is installed under multiple optical conditions (for example, multiple angles), the cost of the facility increases and there are also restrictions on the installation location. For this reason, the imaging device 2 is often configured as two optical systems consisting of a specular reflection camera 21 and a diffuse reflection camera 22, as shown in Fig. 4.

[0048] For example, in the surface inspection device 1 shown in Figure 4, the illumination angle is set to 10° relative to the direction perpendicular to the surface of the steel strip S, the camera angle of the specular reflection camera 21 is set to 10°, and the camera angle of the diffuse reflection camera 22 is set to 45°. At this time, assume that a non-plating defect and a scale defect as shown in Figure 3 have occurred. In this case, the diffuse reflection camera 22, which is installed at a camera angle of 45°, can detect the non-plating defect but cannot detect the scale defect.

[0049] Therefore, the image generation unit 43 outputs image data (first image) with "illumination angle: 10°, camera angle: 45°" as an input to the trained model, as shown in Fig. 5. This image with "illumination angle: 10°, camera angle: 25°" has a contrast with normal parts, making it possible to detect scale defects.

[0050] Here, if the first optical conditions of the first image input to the trained model are appropriate and the image is one in which surface defects are easy to detect, the trained model outputs a second image captured under the first optical conditions. On the other hand, if the first optical conditions of the first image input to the trained model are inappropriate and the image is one in which surface defects are difficult to detect, the trained model outputs a second image captured under the second optical conditions.

[0051] Furthermore, as shown in FIG. 5, the image generation unit 43 inputs one first image into a trained model to output one piece of second image data. At this time, multiple trained models may be prepared to generate multiple second images from one first image. In this case, the image generation unit 43 inputs the same first image into each of the multiple trained models, causing each trained model to output multiple second images captured under different optical conditions. Next, from the multiple second images, the second image most suitable for detecting surface defects is selected, and defect candidates are detected by the defect candidate detection unit 44, which will be described later.

[0052] The defect candidate detection unit 44 performs defect detection based on the image generated by the image generation unit 43. That is, the defect candidate detection unit 44 detects defect candidate portions of the steel strip S to be inspected based on the first image and the second image. The defect candidate detection unit 44 may set a brightness (contrast) threshold for normal portions, as is used in general surface inspection devices, and detect portions outside this threshold as defect candidate portions.

[0053] The defect candidate detection unit 44 may detect defect candidate portions using the method disclosed in Patent Document 3. In this case, first, multiple spatial filters are applied to the image, and position values ​​corresponding to each position on the input image or corrected image are extracted from the filtered feature image to extract feature vectors for each position on the image. Next, the distribution of the extracted multiple feature vectors in multidimensional space is analyzed to calculate the degree of abnormality of each texture feature vector, and the generated anomaly image is binarized using a predetermined degree of abnormality as a threshold. Next, image areas where consecutive pixels with an abnormality degree equal to or greater than the predetermined threshold are connected are detected as defects or defect candidates.

[0054] The defect feature calculation unit 45 calculates a defect feature amount for each defect candidate portion detected by the defect candidate detection unit 44. The defect feature calculation unit 45 may calculate the defect feature amount by the method disclosed in Patent Document 3. In this case, the defect feature amount is calculated using a defect portion grayscale image obtained by cutting out the area of ​​the defect candidate portion from the corrected image, and a defect portion abnormality degree image obtained by similarly cutting out the area of ​​the defect candidate portion from the abnormality degree image.

[0055] The defect determination unit 46 determines the defect type and the degree of harmfulness of each defect candidate portion based on the defect feature amount calculated by the defect feature calculation unit 45 .

[0056] The display device 5 is realized by a general-purpose display such as a liquid crystal display, an organic display, etc. The display device 5 displays detection information (images of surface defects and positions of surface defects) detected by the image processing device 4, determination information (type, harmfulness), statistical information (total number of defects by type and harmfulness for the entire steel strip S, occurrence rate, etc.), etc.

[0057] The model generation device 6 is realized by, for example, a general-purpose computer such as a workstation or a personal computer, or a server located in the cloud. The model generation device 6 generates a trained model to be used by the image generation unit 43. Specifically, the model generation device 6 generates a trained model that has been subjected to machine learning so that a first image captured under first optical conditions is used as an input value, and a second image captured under the first optical conditions or second optical conditions is used as an output value. Examples of machine learning techniques that can be used include deep learning using a neural network, logistic regression, and random forest.

[0058] (Surface inspection method) The processing of the surface inspection method executed by the surface inspection device according to the embodiment will be described with reference to Fig. 6. The surface inspection method includes an imaging step, an image generation step, and a defect candidate detection step.

[0059] In the imaging step, the imaging device 2 captures a first image of the steel strip S to be inspected (step S1). Subsequently, in the image generation step, the image generation unit 43 inputs the first image into the trained model and outputs a second image (step S2). Subsequently, in the defect candidate detection step, the defect candidate detection unit 44 detects defect candidate portions of the steel strip S to be inspected based on the first image and the second image (step S3), thereby completing this process.

[0060] The surface inspection method, surface inspection device, and trained model according to the above-described embodiments can accurately detect defects that are difficult to detect using conventional surface inspection methods. Furthermore, the surface inspection method, surface inspection device, and trained model according to the embodiments can unify comparison standards and inspection standards by generating images under the same optical conditions for results and inspection standards from multiple surface inspection devices with different optical conditions installed in actual manufacturing equipment. Furthermore, the trained model generation method according to the embodiments can generate a trained model that can accurately detect surface defects in an inspection target.

[0061] (Object manufacturing equipment) The surface inspection device according to the embodiment can also be applied to an object manufacturing facility. In this case, the object manufacturing facility inspects the surface of the object using the surface inspection device 1 and manufactures the object based on the inspection results. Such an object manufacturing facility can manufacture objects without surface defects.

[0062] (Manufacturing method of objects) The surface inspection method according to the embodiment can also be applied to a manufacturing method of an object. In this case, the manufacturing method of the object inspects the surface of the object using the surface inspection method described above, and manufactures the object based on the inspection results. According to such a manufacturing method of the object, it is possible to manufacture an object without surface defects.

[0063] (Quality control method for objects) The surface inspection method according to the embodiment can also be applied to a quality control method for an object. In this case, the quality control method for an object inspects the surface of the object using the surface inspection method described above, and controls the quality of the object based on the inspection results. According to such a quality control method for an object, the quality of a manufactured object can be appropriately controlled.

[0064] The surface inspection method, object manufacturing method, object quality control method, learning model generation method, surface inspection device, object manufacturing equipment, and learned model according to the present invention have been specifically described above using the detailed description and examples for carrying out the invention, but the spirit of the present invention is not limited to these descriptions and should be broadly interpreted based on the claims. Furthermore, it goes without saying that various changes, modifications, etc. based on these descriptions are also included in the spirit of the present invention. [Explanation of symbols]

[0065] 1. Surface inspection equipment 2. Imaging device 21 Specular reflection camera 22 Diffuse Reflection Camera 3. Lighting equipment 4. Image processing device 41 Image input unit 42 Image correction section 43 Image generation unit 44 Defect candidate detection unit 45 Defect feature calculation unit 46 Defect Judgment Section 5 Display device 6 Model generation device S steel strip

Claims

1. An image generation step in which an image generation means provided in a computer inputs a first image of an object to be inspected into a trained model and outputs a second image from the trained model; a defect candidate detection step in which a defect candidate detection means included in the computer detects a defect candidate portion of the inspection object based on the first image and the second image; Including, The trained model is a model that has been subjected to machine learning so that the first image captured under first optical conditions is used as an input value, and the second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value. Surface inspection methods.

2. the first optical condition and the second optical condition include any one of an angle of an illumination device that irradiates light onto the inspection object, an angle of an imaging device that images the inspection object, whether or not an optical filter is used, and a type of the optical filter; In the image generation step, the type of the first optical condition of the first image input to the trained model and the type of the second optical condition of the second image output from the trained model are the same. The surface inspection method according to claim 1 .

3. 3. A method for manufacturing an object, comprising inspecting a surface of the object by the surface inspection method according to claim 1 or 2, and manufacturing the object based on the inspection results.

4. 3. A quality control method for an object, which comprises inspecting the surface of the object by the surface inspection method according to claim 1 or 2, and controlling the quality of the object based on the inspection results.

5. A method for generating a learning model in which a model generation means provided in a computer generates a trained model through machine learning, using a first image captured under first optical conditions as an input value and a second image captured under the first optical conditions or second optical conditions different from the first optical conditions as an output value.

6. an image generation unit that inputs a first image of an object to be inspected into a trained model and outputs a second image from the trained model; a defect candidate detection unit that detects a defect candidate portion of the inspection object based on the first image and the second image; Equipped with The trained model is a model that has been subjected to machine learning so that the first image captured under first optical conditions is used as an input value, and the second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value. Surface inspection equipment.

7. the first optical condition and the second optical condition include any one of an angle of an illumination device that irradiates light onto the inspection object, an angle of an imaging device that images the inspection object, whether or not an optical filter is used, and a type of the optical filter; In the image generation unit, the type of the first optical condition of the first image input to the trained model and the type of the second optical condition of the second image output from the trained model are the same. The surface inspection device according to claim 6.

8. An object manufacturing facility comprising the surface inspection apparatus according to claim 6 or 7.

9. A trained model that has been subjected to machine learning so that a first image captured under first optical conditions is used as an input value, and a second image captured under the first optical conditions or second optical conditions different from the first optical conditions is used as an output value.

Images