Inspection device, inspection method, and inspection program

Abstract

To efficiently detect such a failure that a color and a shape of an optical material become indefinite.SOLUTION: An inspection device (1) includes: an image acquisition section (102) for acquiring an inspection image which has been cut out in a manner to include an image region extracted from a taken image (111) of an optical material on the basis of a variation in a luminance value; and a quality determination section (106) for determining quality on the basis of an output value obtained by inputting the inspection image into a quality determination model (113) structured by machine learning.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an inspection device or the like that judges whether an object to be inspected is good or bad. [Background technology]

[0002] Conventionally, there are known techniques for automatically determining whether an object to be inspected is a non-defective product by using an image of the object. For example, Patent Document 1 listed below describes an inspection device for inspecting defects in a phosphor-containing glass member. This inspection device includes an illumination unit that irradiates the phosphor-containing glass member with inspection light that causes the phosphor to emit light, and an imaging unit that images the phosphor-containing glass member irradiated with the inspection light. Furthermore, this inspection device includes an image processing unit that processes the image of the phosphor-containing glass member imaged by the imaging unit, and various defects in the phosphor-containing glass member are detected by this image processing. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 5854370 Summary of the Invention [Problem to be solved by the invention]

[0004] However, it has been difficult to detect surface contamination and special chippings on optical materials using conventional techniques such as those described above, which rely on image processing for defect detection. This is because such defects often have irregular colors and shapes. Therefore, it is possible to use a pass / fail judgment model constructed using machine learning. Even defects with irregular colors and shapes can be detected with high accuracy using a pass / fail judgment model constructed using machine learning with a large amount of training data.

[0005] However, calculations using models constructed by machine learning generally have the problem of high computational costs. Also, in defect inspections, the majority of products are usually non-defective. Therefore, it is not efficient to inspect the entire image area of ​​all inspection targets at the expense of high computational costs in order to detect only a small number of defective products.

[0006] An object of one aspect of the present invention is to provide an inspection device or the like that can efficiently detect defects in optical materials that result in irregular colors or shapes. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, an inspection device according to one aspect of the present invention includes an image acquisition unit that acquires an inspection image cut out from an image of an optical material, which is an object to be inspected, so as to include an image region extracted based on variations in brightness values, and a pass / fail judgment unit that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning.

[0008] In addition, in order to solve the above-mentioned problems, an inspection method according to one aspect of the present invention is an inspection method executed by an inspection device, and includes an image acquisition step of acquiring an inspection image cut out from an image of an optical material that is an inspection object so as to include an image region extracted based on variations in brightness values, and a pass / fail judgment step of inputting the inspection image into a pass / fail judgment model constructed by machine learning and judging the pass / fail of the optical material based on an output value obtained.

[0009] In order to solve the above problems, an inspection device according to one aspect of the present invention comprises an image acquisition unit that acquires an inspection image cut out from an image of an optical material, which is an object to be inspected, so as to include an image region extracted based on an average brightness value, and a pass / fail judgment unit that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning, and the image acquisition unit extracts, as the image region, a region from among the regions constituting the captured image, where the average brightness value of the region falls outside a range determined based on a reference brightness value.

[0010] In order to solve the above-mentioned problems, an inspection device according to one embodiment of the present invention includes an image acquisition unit that acquires an inspection image cut out to include an image region extracted based on a comparison result of brightness values ​​between an image of an optical material that is an object to be inspected and one or more reference images that are images obtained by imaging the optical material in a good condition, and a pass / fail judgment unit that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning.

[0011] In order to solve the above-mentioned problems, an inspection method according to one aspect of the present invention is an inspection method executed by an inspection device, and includes: an image acquisition step of acquiring an inspection image cut out from an image of an optical material, which is an object to be inspected, so as to include an image region extracted based on an average value of brightness values; and a pass / fail judgment step of judging the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning, wherein in the image acquisition step, a region of the image constituting the captured image, whose average brightness value falls outside a range determined based on a reference brightness value, is extracted as the image region.

[0012] In order to solve the above-mentioned problems, an inspection method according to one aspect of the present invention is an inspection method executed by an inspection device, and includes an image acquisition step of acquiring an inspection image cut out to include an image region extracted based on a comparison result of brightness values ​​between an image of an optical material that is an inspection object and one or more reference images that are images obtained by imaging the optical material in a good condition, and a pass / fail judgment step of inputting the inspection image into a pass / fail judgment model constructed by machine learning and judging the pass / fail of the optical material based on an output value obtained.

[0013] The inspection device according to each aspect of the present invention may be realized by a computer. In this case, the inspection program that causes the computer to operate as each part (software element) of the inspection device to realize the inspection device, and the computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention. [Effects of the Invention]

[0014] According to one aspect of the present invention, defects in optical materials that result in irregular colors or shapes can be efficiently detected. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a block diagram showing an example of a configuration of a main part of an inspection device according to an embodiment of the present invention; [Figure 2] FIG. 1 is a diagram showing an overview of an inspection system including the inspection device. [Figure 3] 1A and 1B are diagrams showing an example of a non-defective workpiece image and an example of a defective workpiece image; [Figure 4] 10 is a flowchart illustrating an example of a process for generating teacher data. [Figure 5] FIG. 10 is a diagram illustrating an example of acquiring a learning image from a workpiece image. [Figure 6] FIG. 10 is a diagram illustrating an example of rotating a learning image. [Figure 7]FIG. 10 is a diagram showing a specific example of association of correct answer data in training data of a pass / fail determination model for determining the presence or absence of dirt. [Figure 8] FIG. 10 is a diagram showing a specific example of correspondence between correct answer data in training data of a pass / fail determination model for determining the presence or absence of special chipping. [Figure 9] 4 is a flowchart showing an example of a workpiece inspection process performed by the inspection system. [Figure 10] FIG. 10 is a block diagram showing another example of the configuration of the main parts of the inspection device according to the embodiment of the present invention. [Figure 11] FIG. 10 is a block diagram showing yet another example of the configuration of the main parts of an inspection device according to an embodiment of the present invention. [Figure 12] 10A to 10C are diagrams showing some examples of how to determine an area in a workpiece image for which an average brightness value is to be calculated. [Figure 13] FIG. 10 is a block diagram showing yet another example of the configuration of the main parts of an inspection device according to an embodiment of the present invention. [Figure 14] 3A to 3C are diagrams showing examples of a reference image, a workpiece image, and a difference image processed by an inspection device according to an embodiment of the present invention. [Figure 15] FIG. 10 is a diagram showing an example of a captured image obtained by coaxial epi-illumination photography. [Figure 16] 10A to 10C are diagrams showing other examples of a reference image, a workpiece image, and a difference image processed by the inspection device according to one embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] First Embodiment [System Overview] An overview of an inspection system 100 according to one embodiment of the present invention will be described with reference to FIG. 2. FIG. 2 is a diagram showing an overview of the inspection system 100. The inspection system 100 is a system that automatically determines whether an inspection object is a non-defective product based on an image of the inspection object, and includes an inspection device 1 that makes the determination and an imaging device 2 that captures the image used for this determination. As will be described in detail below, the inspection system 100 can efficiently detect defects in which the color or shape of the inspection object becomes irregular. Furthermore, the inspection system 100 can also determine the type of defect.

[0017] The inspection system 100 is suitable for inspecting optical materials. This optical material may be any material that transmits light and may also be referred to as a light-transmitting member. Typical optical materials include inorganic materials such as glass. The optical material may also contain particles, such as phosphors, that cause the wavelength of transmitted or reflected light to differ from that of incident light. In this case, the object being inspected may also be a wavelength conversion member. Wavelength conversion members may include inorganic materials, such as glass or alumina, that contain phosphors, or materials composed solely of phosphors. Glass itself may also be provided with wavelength conversion properties. Generally, optical materials containing particles are difficult to inspect for quality, making high-precision automated inspection difficult. However, the inspection system 100 enables accurate quality determination of optical materials containing particles.

[0018] FIG. 2 shows an example in which a workpiece W made of phosphor-containing glass is used as the inspection object. The workpiece W is formed by cutting a plate of phosphor-containing glass to a predetermined size. In a captured image of the cut phosphor-containing glass plate, multiple workpieces W are shown lined up, and grooves G are visible between adjacent workpieces W in the captured image. Below, an example of determining whether each of the workpieces W is good or bad will be described.

[0019] In the example of Figure 2, the imaging device 2 images the workpiece W using coaxial epi-illumination. Coaxial epi-illumination is an illumination method in which the optical axis of the light irradiating the object is aligned with the optical axis of the camera to obtain reflected light from the surface of the object. The image captured by the imaging device 2 is input to the inspection device 1 and used to determine whether the workpiece W is good or bad. When inspecting the quality of the workpiece W, defects have the characteristic of appearing in the image due to differences in reflection, regardless of color, so the captured image may be monochrome. The captured image may also be color.

[0020] [Regarding pass / fail judgment] The criteria for determining pass / fail in the inspection system 100 will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of a pass / fail workpiece image and an example of a failing workpiece image. The workpiece image is an image of a workpiece area (hereinafter referred to as "work area") extracted from a captured image captured by the imaging device 2. When multiple workpiece areas are included in one captured image, as in the example of Fig. 2, a workpiece image is generated for each workpiece area.

[0021] The workpieces W1 to W3 shown in Figure 3 are all made of phosphor-containing glass and therefore contain phosphor particles. Therefore, the surface has glass and phosphor particle surfaces, each with different reflectances, so the image of each workpiece does not have a uniform brightness value, and the entire surface appears mottled due to the phosphor particles. Of these, workpiece W1 is an example of a non-defective workpiece. On the other hand, workpieces W2 and W3 are examples of defective workpieces, particularly those containing defects in optical materials that result in irregular colors or shapes.

[0022] Workpiece W2 is an example of a workpiece with dirt on its surface. In Figure 3, workpiece W2 has dirt attached to the area surrounded by the dashed line. When dirt is present on the surface of an optical material, the dirt can blacken due to heat or light, preventing light from passing through. For this reason, when workpiece W2 with a dirty surface is used in combination with, for example, an LED (Light Emitting Diode), problems such as poor light-emitting characteristics may occur. Therefore, workpiece W2 with dirt attached must be determined to be a defective product.

[0023] Dirt on workpieces is often transparent or white. In this case, oblique illumination makes it difficult for the dirt to appear in captured images due to reflections from optical materials and the effects of excitation light. Therefore, imaging with coaxial incident illumination, as shown in the example of Figure 2, is preferable. However, even when imaging with coaxial incident illumination, the surface of a phosphor-containing glass workpiece can have indentations caused by, for example, phosphor particle detachment, and the shape and density of the dirt are not uniform, making it difficult to accurately determine the presence or absence of dirt using conventional image processing. In light of this, the inspection device 1 is configured to perform pass / fail judgment using a pass / fail judgment model constructed through machine learning using images of the workpiece with dirt attached as training data. Using such a pass / fail judgment model makes it possible to accurately determine the presence or absence of dirt.

[0024] Workpiece W3 is an example of a workpiece containing special chipping C. Special chipping is different from normal chipping in that the broken pieces do not peel off (see cross section A-A' in Figure 3). If special chipping occurs in a workpiece, for example, when the workpiece is assembled into an LED package, there is a risk that a piece will peel off and become caught between the LED and the workpiece, causing the LED to be mounted at an angle, or other problems. Therefore, workpiece W3 containing special chipping C must also be determined to be a defective product.

[0025] When imaging a normal chipping with coaxial epi-illumination, the chipped portion appears completely black. However, because special chipping has not peeled off from the optical material, it often appears light black or the same color as the workpiece, depending on the degree of peeling. Furthermore, similar to the dirt mentioned above, the area around the special chipping may contain depressions caused by fallen phosphor particles or voids (air gaps) that appear on the surface. Therefore, attempting to detect special chipping simply by color differences can result in oversight or overdetection. Thus, it is difficult to accurately determine the presence or absence of special chipping using conventional image processing. In consideration of this issue, the inspection device 1 is configured to perform pass / fail judgment using a pass / fail judgment model constructed through machine learning using images of workpieces with special chipping as training data. Using such a pass / fail judgment model, it is possible to accurately determine the presence or absence of special chipping.

[0026] [Configuration of inspection device] The configuration of the inspection device 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the configuration of the main parts of the inspection device 1. As shown in Fig. 1, the inspection device 1 includes a control unit 10 that controls each part of the inspection device 1, and a storage unit 11 that stores various data used by the inspection device 1. The inspection device 1 also includes a communication unit 12 for communicating with other devices. Furthermore, the inspection device 1 includes an input unit 13 that accepts input operations for the inspection device 1, and an output unit 14 that allows the inspection device 1 to output data.

[0027] The control unit 10 includes a work area extraction unit 101, an image acquisition unit 102, an image rotation unit 103, a teacher data generation unit 104, a learning unit 105, and a pass / fail determination unit 106. The memory unit 11 stores a captured image 111, teacher data 112, and a pass / fail determination model 113.

[0028] The work area extraction unit 101 extracts a work area from the captured image 111. The extraction of the work area will be described later.

[0029] The image acquisition unit 102 acquires an inspection image or a learning image cut out from the captured image 111 so as to include an image area extracted based on the variation in brightness values. The inspection image is an image used to determine the quality of an object to be inspected, and the learning image is an image used as training data 112. In this embodiment, as described above, the work area extraction unit 101 extracts a work area from the captured image 111, and the image acquisition unit 102 cuts out an inspection image or a learning image from this work area.

[0030] Image rotation unit 103 rotates the training image so that the peripheral portion of the optical material shown in the training image is positioned on a predetermined one of the four sides (top, bottom, left, and right) of the training image. Similarly, image rotation unit 103 rotates the test image so that the peripheral portion of the optical material shown in the test image is positioned on a predetermined one of the four sides (top, bottom, left, and right) of the test image. These image rotations will be described later with reference to FIG. 6.

[0031] The teacher data generation unit 104 generates teacher data 112 used when constructing the pass / fail determination model 113. Then, the learning unit 105 constructs the pass / fail determination model 113 by supervised learning using the teacher data 112. The generation of the teacher data 112 and the construction of the pass / fail determination model 113 will be described later.

[0032] The pass / fail determination unit 106 determines the pass / fail of the object to be inspected based on the output value obtained by inputting the inspection image to the pass / fail determination model 113. At this time, if the image rotation unit 103 has rotated the inspection image, the pass / fail determination unit 106 determines the pass / fail of the object to be inspected based on the output value obtained by inputting the inspection image rotated by the image rotation unit 103 to the pass / fail determination model 113.

[0033] As described above, the captured image 111 is an image of an object to be inspected, and is generated by the imaging device 2. For example, a learning image is generated from the captured image 111 of an object to be inspected whose pass / fail is known. Also, an inspection image is generated from the captured image 111 of an object to be inspected whose pass / fail is unknown.

[0034] As described above, the teacher data 112 is data used when constructing the pass / fail determination model 113, and corresponds, as correct answer data, to learning images generated from captured images 111 of an object to be inspected. Details of the teacher data 112 will be described later.

[0035] The pass / fail determination model 113 is a trained model that determines whether an object to be inspected is pass / fail. In response to an input of an inspection image cut out from a captured image 111 capturing an object to be inspected, the pass / fail status of which is unknown, the pass / fail determination model 113 outputs a value indicating whether the object to be inspected is a pass / fail product. Details of the pass / fail determination model 113 will be described later.

[0036] As described above, the inspection device 1 includes an image acquisition unit 102 that acquires an inspection image cut out from a captured image of an optical material (workpiece W in the example of FIG. 2) that is an inspection target, so as to include an image region extracted based on variations in brightness values. The inspection device 1 also includes a pass / fail judgment unit 106 that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model 113 constructed by machine learning.

[0037] Here, research by the inventors of the present application has revealed that in an image captured of an optical material, which is the object to be inspected, image areas in which defects have occurred, such as irregular color or shape, have larger variations in brightness values ​​than image areas in which no defects have occurred.

[0038] Therefore, the inspection device 1 is configured to acquire an inspection image cut out from the captured image so as to include an image region extracted based on the variation in brightness values.The inspection device 1 then judges the quality of the optical material based on the output value obtained by inputting the inspection image cut out so as to include a partial image region of the captured image, rather than the entire captured image, into a quality judgment model.As a result, the inspection device 1 can efficiently detect defects in the optical material that result in irregular colors or shapes.

[0039] [Generation of training data] The flow of the process of generating the teacher data 112 in the inspection system 100 will be described with reference to Fig. 4. Fig. 4 is a flowchart showing an example of the process of generating the teacher data 112. In the following, an example will be described in which the object to be inspected is a workpiece made of phosphor-containing glass.

[0040] In S1, the imaging device 2 captures an image of a workpiece whose quality is known, and generates a captured image. The inspection device 1 acquires the generated captured image through communication via the communication unit 12 or input via the input unit 13, and stores it as a captured image 111 in the storage unit 11. Note that the quality of the workpiece does not necessarily need to be known at the stage of S1, and it is sufficient to know it at the latest at the time of processing in S7.

[0041] In S2, the work area extraction unit 101 extracts a work area from the captured image 111. For example, the work area extraction unit 101 may identify the position of a groove between two workpieces shown in the captured image 111 by image analysis, and extract the area surrounded by the identified groove as the work area. In this case, it is preferable that the work area extraction unit 101 extracts the work area so as not to include the groove between two workpieces. This is because if the work area is extracted so as to include the groove portion, the accuracy of the determination may be reduced.

[0042] In S3, the image acquisition unit 102 identifies an image area with a relatively large variation in brightness values ​​from among multiple image areas that show one workpiece within the workpiece area extracted in S2. Then, in S4, the image acquisition unit 102 acquires a learning image that includes the image area identified in S3 from the workpiece image extracted in S2.

[0043] A specific example of the processes in S3 and S4 will now be described with reference to FIG. 5. FIG. 5 is a diagram showing an example of acquiring learning images from workpiece images. In FIG. 5, workpieces W4 and W5 are examples of workpieces that contain defects that result in irregular color or shape. Specifically, workpiece W4 contains special chipping near the left end of its upper end. In addition, workpiece W5 has dirt attached to it.

[0044] When constructing a pass / fail judgment model 113 for judging the presence or absence of defects occurring at the edge of a workpiece, such as special chipping, the image acquisition unit 102 preferably acquires a learning image such as image P1 in Fig. 5. That is, the image acquisition unit 102 preferably acquires a learning image cut out from the captured image so as to include each image region set along the outer edge of the workpiece W4, based on the variation in brightness values ​​in the image region. This makes it possible to efficiently acquire learning images by narrowing down the image region to one in which defects occurring at the edge of the workpiece W4 can be detected.

[0045] In this case, in S3, the image acquisition unit 102 first calculates an index value indicating the variation in brightness values ​​in each image region set along the outer edge of the workpiece W4 shown in the captured image. For example, the image acquisition unit 102 may calculate the variance of brightness values ​​in each image region as the index value.

[0046] It should be noted that the way in which each image area is set may be determined in advance. For example, multiple image areas may be set in advance. As a specific example, a strip-shaped area along the top side of the rectangular workpiece W4 may be equally divided to set multiple image areas. In this case, the strip-shaped area along the other three sides of the workpiece W4 is similarly equally divided to set multiple image areas. It should be noted that the width of the strip-shaped area may be set to a width that is sufficient to accommodate the special chipping.

[0047] Alternatively, for example, an image area may be set when calculating an index value. In this case, the image acquisition unit 102 calculates the index value of each image area along the outer edge of the workpiece W4 while moving the image area for which the index value is to be calculated along the outer edge of the workpiece W4. Note that whether multiple image areas are set in advance or the image area is set when calculating the index value, the size of the image area is the same as or smaller than the size of the learning image.

[0048] On the other hand, when constructing a pass / fail judgment model 113 for judging the presence or absence of defects whose appearance position is irregular, such as the stains shown on the workpiece W5, in S1, an image of a workpiece that is known to have stains is captured. Then, in the process of S3, the image acquisition unit 102 sets image areas so that the entire surface of the workpiece image is covered (in other words, so that each pixel of the workpiece image is included in at least one of the multiple image areas), and identifies image areas from the image areas that have large variations in brightness values.

[0049] Next, the image acquisition unit 102 identifies an image region where the calculated index value is relatively large. Multiple image regions may be identified. Note that, instead of identifying an image region where the index value is relatively large, the image acquisition unit 102 may identify an image region where the calculated index value is equal to or greater than a predetermined threshold. In other words, the image region may be identified based on a threshold that is an absolute standard. However, because variations in brightness values ​​may vary due to changes in imaging conditions and individual differences between inspection targets, it is preferable to identify an image region where the index value is relatively large. By identifying an image region where the index value is relatively large, it is possible to stably extract an image region where a defect is likely to occur and use it as a learning image.

[0050] Then, in S4, the image acquisition unit 102 acquires a learning image including the image area identified in S3 from the workpiece image. For example, if image area A1 shown in FIG. 5 is identified in S3, the image acquisition unit 102 may use image P1, which is cropped to include image area A1, as the learning image. Similarly, if image area A2 shown in FIG. 5 is identified in S3, the image acquisition unit 102 may use image P2, which is cropped to include image area A2, as the learning image. Note that if the image acquisition unit 102 identifies multiple image areas, it may acquire a learning image for each of the image areas.

[0051] The size of the learning image may be determined in advance. If the predetermined size of the learning image and the size of the image area identified in S3 are the same, the image acquisition unit 102 may simply cut out the image area identified in S3 and use it as the learning image. On the other hand, if the size of the image area identified in S3 is smaller than the predetermined size of the learning image, the image acquisition unit 102 may cut out a learning image of a predetermined size that includes the image area identified in S3 and the area surrounding that image area. Depending on the use of the workpiece, the image acquisition unit 102 may cut out a learning image of a rectangular size, for example, with one side measuring 0.05 mm to 200 mm.

[0052] Furthermore, when the object to be inspected is a workpiece as in this example, it is preferable that in S3 and S4, the image acquisition unit 102 acquires learning images of a size smaller than the smallest workpiece among multiple types of workpieces of different sizes. This is because learning images of a size smaller than the smallest workpiece can be extracted from any type of workpiece. In other words, with the above configuration, learning images of the same size can be acquired from captured images of any type of workpiece, and these can be used as training data for constructing the pass / fail determination model 113.

[0053] In S5, the image rotation unit 103 determines whether or not to rotate the learning image. Specifically, the image rotation unit 103 determines not to rotate the learning image (NO in S5) if the periphery of the optical material (specifically, the workpiece) shown in the learning image is located on a predetermined side of the four sides (top, bottom, left, and right) of the learning image. In this case, the process proceeds to S7. Note that when constructing a pass / fail determination model 113 that determines the presence or absence of defects whose appearance positions are irregular, such as dirt, the processes of S5 and S6, which will be described later, are omitted.

[0054] On the other hand, if the peripheral portion is not located on a predetermined side of the learning image, the image rotation unit 103 determines to rotate it (YES in S5). In this case, the process proceeds to S6. In S6, the image rotation unit 103 rotates the learning image so that the peripheral portion of the workpiece is located on a predetermined side of the four sides (top, bottom, left, and right) of the learning image.

[0055] A specific example of the processes in S5 and S6 will now be described with reference to Fig. 6. Fig. 6 is a diagram showing an example of rotating a learning image. Learning image P3 in Fig. 6 is an image cut out from the outer edge on the lower end side of the workpiece image, and learning image P4 is an image extracted from the outer edge on the left end side of the workpiece image.

[0056] More specifically, in the learning image P3, the bottom side corresponds to the periphery of the workpiece, and special chipping C1 occurs on the bottom side, while in the learning image P4, the left side corresponds to the periphery of the workpiece, and special chipping C2 occurs on the left side.

[0057] Here, the special chipping C1 in training image P3 is a semicircular shape that convexes upward, while the special chipping C2 in training image P4 is a semicircular shape that convexes to the right, and their appearances are significantly different. For this reason, if machine learning is performed using training data in which the same correct answer data (data indicating the presence of a special chipping) is associated with these training images, there is a possibility that efficient learning will not be achieved.

[0058] Therefore, as described above, the image rotation unit 103 rotates the training images so that the peripheral portion of the workpiece depicted in the training images is positioned on a predetermined side of the four sides (top, bottom, left, and right) of the training images. In the example of Fig. 6, the training images P3 and P4 are rotated so that the peripheral portion of the workpiece is positioned on the upper side of the four sides (top, bottom, left, and right) of the training images.

[0059] Specifically, the image rotation unit 103 rotates the learning image P3 by 180 degrees to generate a learning image P3'. In the learning image P3', the special chipping C1 is located on the top side. The image rotation unit 103 also rotates the learning image P4 by 90 degrees to the right to generate a learning image P4'. In the learning image P4', the special chipping C2 is located on the top side.

[0060] In this way, by rotating the learning images with the image rotation unit 103, the position of the peripheral portion shown in the learning images can be unified. This allows the appearance of chippings and other defects that occur in a predetermined shape along the peripheral portion of the workpiece to be made similar between the learning images, making it possible to efficiently learn the characteristics of that appearance. Of course, this configuration is also effective for optical materials other than workpieces, as special chippings tend to occur on the peripheral portions. Note that the rotation of the inspection image is performed in the same way.

[0061] In S7, the teacher data generation unit 104 generates teacher data by associating the training images with the correct answer data. Then, the teacher data generation unit 104 stores the generated teacher data in the storage unit 11 as teacher data 112, thereby completing the processing in Fig. 4. The processing in Fig. 4 is repeated until the number of teacher data 112 required to construct the pass / fail judgment model 113 is obtained.

[0062] The above-mentioned correct answer data is data indicating whether the object to be inspected is good or bad. When a workpiece having special chipping is imaged in S1, the correct answer data may be data indicating that the workpiece is defective or that special chipping is present. Similarly, when a workpiece having dirt is imaged in S1, the correct answer data may be data indicating that the workpiece is defective or that dirt is present. Furthermore, when a workpiece known to be good in S1 is imaged, the correct answer data may be data indicating that the workpiece is good. Such correct answer data may be input by user operation via the input unit 13, for example. In this case, the teacher data generation unit 104 generates teacher data 112 by associating the input correct answer data with the learning image.

[0063] A specific example of correct data association will now be described with reference to Fig. 7 and Fig. 8. Fig. 7 is a diagram showing a specific example of correct data association in the training data of the pass / fail determination model 113 that determines the presence or absence of dirt. Fig. 8 is a diagram showing a specific example of correct data association in the training data of the pass / fail determination model that determines the presence or absence of special chipping.

[0064] 7 are examples of learning images cut out from the captured image 111 of a non-defective workpiece. These learning images are associated with correct answer data indicating that the workpiece is a non-defective product and are used as training data.

[0065] On the other hand, images P7 and P8 are examples of learning images cut out from captured image 111 of a workpiece with dirt attached. These learning images are associated with correct answer data indicating that the product is defective or that it is dirty, and are used as training data.

[0066] Note that the stains captured in image P7 are soot-like, while the stains captured in image P8 are water droplet-like. Although they are both the same stain, their appearances are different. It is preferable to detect stains with different appearances as different types of defects. This improves the accuracy of defect detection. In this case, the correct answer data can be used as data indicating the type of defect. For example, the correct answer data for "soot-like stains" can be associated with image P7 as training data, and the correct answer data for "water droplet-like stains" can be associated with image P8 as training data.

[0067] In this way, by classifying defects into multiple types and performing machine learning using training data for each type of defect, it is possible to construct a pass / fail determination model 113 that distinguishes between and detects each of the multiple types of defects. In this case, the pass / fail determination unit 106 determines the pass / fail of the optical material based on the output value obtained by inputting the inspection image into the pass / fail determination model 113 that distinguishes between and detects each of the multiple types of defects. This configuration can improve the determination accuracy compared to using the pass / fail determination model 113 that determines the presence or absence of defects without distinguishing between types.

[0068] Furthermore, images P9 to P16 in FIG. 8 are examples of learning images cut out from captured images of good-quality workpieces. Note that "good" in FIG. 8 means that it does not contain special chipping. For example, even if an image contains normal chipping or scratches that are not special chipping, it is still considered a "good" image if it does not contain special chipping. In addition, images that contain areas with dents or areas where phosphor particles are concentrated are also considered to be "good." These learning images are associated with correct answer data that indicates that they are good products and are used as training data.

[0069] Image P9 has relatively little variation in the brightness values ​​of the pixels that make up this image, but images P10 to P16 have areas where pixels with low brightness values ​​(close to black) are concentrated. Areas where normal chipping has occurred are areas where pixels with low brightness values ​​are concentrated. Note that such areas caused by normal chipping can be easily detected by image processing, so they can be detected by image processing in a separate process from the judgment using the pass / fail judgment model 113.

[0070] Since an image region including pixels with low brightness values ​​has a large variation in brightness values ​​overall, images such as P10 to P16 are likely to be acquired as learning images by the processes of S3 and S4 in Fig. 4. In other words, by the processes of S3 and S4 in Fig. 4, images including regions that may be confused with special chipping, such as P10 to P16, are likely to be acquired as learning images. By using such confusing images as learning images, the judgment accuracy of the pass / fail judgment model 113 can be improved.

[0071] Note that a learning image of a non-defective product can also be generated by cutting out any image area from a captured image of a workpiece that is known to be non-defective. In this case, the process in Figure 4 does not need to be performed.

[0072] On the other hand, images P17 to P20 in FIG. 8 are examples of learning images cut out from captured images of workpieces with special chipping. These learning images show special chippings C3 to C6. These learning images are associated with correct answer data indicating that the product is defective or that it contains special chipping, and are used as training data. Image areas where special chipping has occurred have a large variation in brightness values, although not as large as normal chipping. Therefore, by performing the processes in S3 and S4 in FIG. 4, learning images containing special chipping, such as P17 to P20, can be easily generated.

[0073] [Building a pass / fail judgment model using machine learning] By performing machine learning using the teacher data 112 generated and recorded by the above-described processing, it is possible to construct a pass / fail determination model 113. It is preferable to construct two types of pass / fail determination models 113: one that determines the presence or absence of special chipping, and one that determines the presence or absence of dirt.

[0074] The type of machine learning algorithm and model is not particularly limited as long as it can construct a pass / fail determination model 113 that uses an inspection image as input data and outputs an output value indicating the pass / fail of an object to be inspected. For example, the pass / fail determination model 113 may be a convolutional neural network model.

[0075] [Test procedure] The flow of workpiece inspection (inspection method) by the inspection system 100 will be described with reference to Fig. 9. Fig. 9 is a flowchart showing an example of workpiece inspection processing by the inspection system 100. Note that the processing of S11 to S16 is generally similar to S1 to S6 in Fig. 4, except that the name of the image has changed from "learning image" to "inspection image." The following will focus on the differences from the flowchart in Fig. 4. Also, the following will describe an example of determining pass / fail based on the presence or absence of special chipping.

[0076] In S11, the imaging device 2 captures an image of the workpiece, the quality of which is unknown, as the inspection object, to generate a captured image, and the inspection device 1 acquires the generated captured image. Then, in S12, the work area extraction unit 101 extracts a work area from the captured image captured in S11. At this time, it is preferable that the work area extraction unit 101 extracts the work area so as not to include grooves between the workpieces. Note that the extraction of the work area can be performed, for example, by image recognition software.

[0077] In S13, the image acquisition unit 102 identifies an image area with a large variation in brightness values ​​from among multiple image areas in which one workpiece is captured. Then, in S14 (image acquisition step), the image acquisition unit 102 acquires an inspection image cut out from the captured image so as to include the image area identified in S13. It is preferable that the size of the inspection image be the same as that of the learning image used to build the pass / fail determination model 113.

[0078] Even if the inspection target is the same type, the variation in brightness value may vary due to changes in imaging conditions or individual differences between the inspection targets. For this reason, in S13, an image area with a relatively large variation in brightness value may be identified from multiple image areas that capture a single workpiece. This makes it possible to stably extract image areas that are likely to contain defects and use them as inspection images, even if the variation in brightness value varies.

[0079] Furthermore, when performing a pass / fail judgment based on the presence or absence of special chipping, it is preferable that in S14, the image acquisition unit 102 acquires an inspection image set along the outer edge of the workpiece shown in the captured image, such as image P1 in FIG. 5. That is, it is preferable that in S13, the image acquisition unit 102 identifies an image area to be cut out based on the variation in brightness values ​​in each image area set along the outer edge of the workpiece shown in the captured image. Then, it is preferable that in S14, the image acquisition unit 102 acquires an inspection image cut out from the captured image so as to include that image area. Note that, in identifying an image area with a large variation in brightness values, for example, the control unit 10 may import data for multiple image areas, calculate the variation in brightness values ​​in each image area, and then determine which image area has a large variation in brightness values.

[0080] With this configuration, an inspection image is obtained that is cut out based on the variation in brightness values ​​in each image area set along the outer edge of the workpiece, so that special chipping that often occurs on the outer edge of the workpiece can be efficiently detected and the quality of the workpiece can be determined.

[0081] Furthermore, when the object to be inspected is a workpiece made of an optical material, in S13 and S14, it is preferable that the image acquisition unit 102 acquires an inspection image of a size smaller than the smallest workpiece among multiple types of workpieces of different sizes.

[0082] This is because an inspection image of a size smaller than the smallest workpiece can be extracted from any type of workpiece. In other words, with the above configuration, when determining the quality of any type of workpiece, an inspection image of the same size is acquired, and the quality of the workpiece can be determined based on the output value obtained by inputting this inspection image into the quality determination model 113. This makes it possible to use a common quality determination model 113 for multiple types of workpieces. Furthermore, because the inspection image is not enlarged or reduced, detection accuracy can be maintained even if the type of workpiece changes.

[0083] In S15, the image rotation unit 103 determines whether or not to rotate the inspection image. Specifically, if the peripheral portion of the workpiece shown in the inspection image is located on a predetermined side of the four sides (top, bottom, left, and right) of the inspection image, the image rotation unit 103 determines not to rotate the image (NO in S15). In this case, the process proceeds to S17. On the other hand, if the peripheral portion is not located on a predetermined side of the four sides (top, bottom, left, and right) of the inspection image, the image rotation unit 103 determines to rotate the image (YES in S15). In this case, the process proceeds to S16.

[0084] The predetermined side is the same side as the side used as a reference when determining whether to rotate the learning image in S5 of Fig. 4. For example, if the learning image is rotated when the periphery of the workpiece is not located on the top side of the learning image in S5 of Fig. 4, it is also determined in S15 that the inspection image should be rotated when the periphery of the workpiece is not located on the top side of the inspection image.

[0085] In S16, the image rotation unit 103 rotates the inspection image so that the peripheral portion of the workpiece shown in the inspection image is positioned on a predetermined side of the four sides (top, bottom, left, and right) of the inspection image. With this configuration, the positions of the peripheral portions shown in the inspection images input to the quality determination model 113 are unified, so that chipping or the like that occurs in a predetermined shape along the peripheral portion input to the quality determination model 113 can be detected with high accuracy to determine the quality of the workpiece.

[0086] Then, in S17, the quality determination unit 106 inputs the inspection image rotated by the image rotation unit 103 to the quality determination model 113. Then, in S18 (quality determination step), the quality determination unit 106 determines the quality of the workpiece based on the output value from the quality determination model 113 constructed by machine learning.

[0087] For example, when using a quality determination model 113 constructed by learning using two types of training data, one for a non-defective product and one for a product with special chipping, the quality determination model 113 outputs the probability that the workpiece shown in the input inspection image is a non-defective product and the probability that the workpiece has special chipping. In this case, the quality determination unit 106 can determine whether the workpiece is good or bad by, for example, comparing these values ​​with a predetermined threshold. For example, the quality determination unit 106 may determine that the workpiece is good if (1) the probability that the workpiece is a non-defective product is equal to or greater than a threshold and (2) the probability that the workpiece has special chipping is less than a threshold, and may determine that the workpiece has special chipping if at least one of the conditions (1) and (2) is not satisfied.

[0088] In S19, the pass / fail judgment unit 106 outputs the judgment result of S18 to the output unit 14. For example, if the output unit 14 is a display device that displays and outputs an image, the pass / fail judgment unit 106 may cause the output unit 14 to display and output an image showing the judgment result of S18. At this time, the pass / fail judgment unit 106 may also display the inspection image or the like used for the judgment. This allows the user to visually confirm the validity of the pass / fail judgment result.

[0089] The inspection process for determining pass / fail based on the presence or absence of dirt is the same as the above process, except that the image area subjected to the processing in S13 is different (including areas other than the periphery), the processing in S15 and S16 is not performed, and the pass / fail determination model 113 used is different.

[0090] [Modification] In the above embodiment, an example was described in which the inspection device 1 also generates the training data and constructs the pass / fail determination model 113, but the generation of the training data and the construction of the pass / fail determination model 113 may be performed by another information processing device.

[0091] Also, some of the processes in the inspection may be performed by another information processing device. For example, among the processes in FIG. 9, the process of S17 in which the inspection image is input to the quality determination model 113 and calculations are performed may be performed by another information processing device. In this case, the quality determination unit 106 transmits the inspection image to the other information processing device, acquires an output value of the quality determination model 113 from the other information processing device, and performs the quality determination process of S18 based on the acquired output value. Also, for example, the process of S13 in FIG. 9 may be performed by another information processing device. In this case, the image acquisition unit 102 acquires, in S14, the inspection image generated by the other information processing device based on the identification result of S13. Second Embodiment Other embodiments of the present invention will be described below. For ease of explanation, the same reference numerals will be used to designate components having the same functions as those described in the above embodiment, and the description thereof will not be repeated.

[0092] The present embodiment aims to efficiently detect defects in optical materials that result in irregular colors or shapes, and to more efficiently detect defects that are uniformly spread over a slightly wider area than the localized defects assumed in embodiment 1.

[0093] For this purpose, the image acquisition unit 102 of the inspection device 1 according to this embodiment is configured to acquire, as a learning image or an inspection image, an image cut out from the captured image so as to include an image area extracted based on the average brightness value.

[0094] [Configuration of inspection device] In the inspection device 1 of this embodiment, the image acquisition unit 102 is configured to acquire an inspection image cut out from a captured image of an optical material, which is an object to be inspected, so as to include an image region extracted based on an average value of brightness values. As an example, the image acquisition unit 102 extracts, from among the regions constituting the captured image, a region whose average brightness value falls outside a range determined based on a reference brightness value, as the image region.

[0095] The quality determination unit 106 determines the quality of the optical material based on the output value obtained by inputting the inspection image acquired as described above into a quality determination model 113 constructed by machine learning.

[0096] [Testing method] The inspection method executed by the inspection device 1 according to this embodiment includes S14 (image acquisition step) and S18 (determining pass / fail) similarly to the inspection method according to embodiment 1 (see FIG. 9). However, the content of the process in S14 differs from that of the inspection method according to embodiment 1 as follows.

[0097] In this embodiment, in S14, the image acquisition unit 102 acquires an inspection image cut out from a captured image of an optical material, which is an inspection object, so as to include an image region extracted based on the average brightness value. Specifically, the image acquisition unit 102 extracts, as the image region, a region from the captured image whose average brightness value falls outside a range determined based on a reference brightness value.

[0098] In S18, the pass / fail judgment unit 106 judges the pass / fail of the optical material based on the output value obtained by inputting the inspection image acquired as described above into the pass / fail judgment model 113 constructed by machine learning. More specifically, the pass / fail judgment unit 106 inputs the inspection image, which has been enlarged or reduced to the same size as the learning image used as training data for the pass / fail judgment model 113, into the pass / fail judgment model 113. Alternatively, the image acquisition unit 102 may acquire an inspection image whose average brightness value is outside a range determined based on a reference brightness value and has the same size as the learning image. In this case, the pass / fail judgment unit 106 may input the inspection image acquired by the image acquisition unit 102 into the pass / fail judgment model 113 without resizing it.

[0099] According to the above-described configuration and method, the image area of ​​the captured image to be judged as pass / fail can be narrowed down to an area having an average brightness value outside the range determined by the reference brightness value. When a defect occurs in an optical material, resulting in an irregular color or shape, a difference occurs between the average brightness value of the area where such a defect occurs and the average brightness value of the area where no defect occurs. Therefore, according to the above-described configuration, such defects can be detected efficiently. For example, defects that occur uniformly over a certain area on the captured image and are unlikely to exhibit variations in brightness value can also be detected efficiently.

[0100] [Image region extraction] There are several possible methods for the image acquisition unit 102 to extract an image region based on the average brightness value: The image acquisition unit 102 may extract an image region based on the average brightness value of a region in which a non-defective optical material appears in an image obtained by capturing an image of the optical material, as a reference brightness value.

[0101] Alternatively, the image acquisition unit 102 may extract an image region based on an overall average luminance value of the captured image as a reference luminance value. The overall average luminance value as a reference luminance value is the average value of the luminance values ​​of the entire captured image of the optical material that is the inspection target. As the overall average luminance value, the average of the luminance values ​​of all the pixels in the captured image may be applied, or the average of the average luminance values ​​(partial average luminance values) of each partial region obtained by partitioning the captured image may be applied.

[0102] (Extraction of image regions based on the average brightness value of non-defective products) <Extracting areas with high average brightness> The image acquisition unit 102 may use the average value of the luminance values ​​of areas in the captured images obtained by capturing images of multiple sheets of non-defective optical materials as the reference luminance value.The image acquisition unit 102 may then extract, as image areas, areas in the captured images where the average value of the luminance values ​​of the areas exceeds a range determined based on the reference luminance value.

[0103] With the above configuration, the image area to be judged as good or bad can be narrowed down to areas with a higher average brightness value. This allows for efficient detection of defects that appear brighter than normal areas. Such defects may include, for example, aggregation of phosphor particles in optical materials.

[0104] As shown in FIG. 10, in the inspection device 1 according to this embodiment, the memory unit 11 may store a non-defective product image group 114. As an example, the non-defective product image group 114 includes a plurality of captured images of a plurality of phosphor-containing glasses that have been determined to be non-defective in advance with respect to phosphor particle aggregation. More specifically, the non-defective product image group 114 may be a group of workpiece images of non-defective workpieces W contained in phosphor-containing glass. It is preferable that each of the workpiece images in the non-defective product image group 114 does not include a groove G (see FIG. 2 ) portion.

[0105] First, the work area extraction unit 101 extracts a work area from the captured image 111 of the phosphor-containing glass, which is the object to be inspected, and outputs the work image corresponding to the work area to the image acquisition unit 102.

[0106] The image acquisition unit 102 acquires the average value of the luminance values ​​of the non-defective image group 114 as the reference luminance value. The average value of the luminance values ​​may be calculated in advance and stored in the storage unit 11. In this case, it is not necessary to store the non-defective image group 114 in the storage unit 11.

[0107] As one example, the image acquisition unit 102 may extract, as an image area, an area of ​​the workpiece image received from the workpiece area extraction unit 101 that exceeds a range determined based on the average value of the acquired brightness values. The image acquisition unit 102 cuts out an inspection image including the extracted image area from the workpiece image. As one example, the area that exceeds the range determined based on the average value of the acquired brightness values ​​may be an area that exceeds the average value of the acquired brightness values. In another example, the area that exceeds the range determined based on the average value of the acquired brightness values ​​may be an area that exceeds a value obtained by adding or subtracting a predetermined margin to the above-mentioned average value of the brightness values.

[0108] In the inspection device 1 of FIG. 10, the image rotation unit 103 may be omitted.

[0109] The pass / fail judgment unit 106 receives an inspection image from the image acquisition unit 102 that includes an image area that exceeds the average value of a range determined based on the brightness value, and judges the pass / fail of the object to be inspected based on the output value obtained by inputting the inspection image into the pass / fail judgment model 113.

[0110] This allows the inspection device 1 to efficiently detect defects in the aggregation of phosphor particles in the phosphor-containing glass.

[0111] The pass / fail determination model 113 used by the pass / fail determination unit 106 may be a model constructed using correct answer data indicating the presence or absence of aggregation of phosphor particles that should be determined to be defective. The pass / fail determination model 113 is constructed by the teacher data generation unit 104 and the learning unit 105.

[0112] Specifically, in this embodiment, the image acquisition unit 102 may be configured to acquire a learning image cut out from an image of an optical material, which is the object to be inspected, so as to include an image region extracted based on the average brightness value (S3 to S4 in FIG. 4).

[0113] In this embodiment, the pass / fail judgment model 113 used by the pass / fail judgment unit 106 for pass / fail judgment is constructed using the learning images acquired by the image acquisition unit 102. That is, the learning images used in this embodiment are images cut out from the captured image so as to include an image region extracted based on the average brightness value.

[0114] The training data generation unit 104 associates the training images acquired as described above with the correct answer data to generate training data 112 (S7). The learning unit 105 constructs a pass / fail determination model 113 by supervised learning using the training data 112.

[0115] <Extracting areas with low average brightness> In another example, the image acquisition unit 102 may extract, as an image area, an area that constitutes a captured image, where the average brightness value of the area is below a range determined based on the reference brightness value.

[0116] With the above configuration, the area to be judged as pass / fail can be narrowed down to an area with a low average brightness value. This allows for efficient detection of defects that appear darker than normal areas. Such defects may be, for example, dirt adhering to optical materials. It is also possible to detect defects caused by special chipping.

[0117] In this embodiment, as an example, the non-defective product image group 114 includes a plurality of captured images of a plurality of phosphor-containing glasses that have been determined to be non-defective in terms of stains and special chipping. In this embodiment, the captured images in the non-defective product image group 114 may be divided into workpiece images so that the grooves G are not captured.

[0118] First, the work area extraction unit 101 extracts a work area from the captured image 111 of the phosphor-containing glass, which is the object to be inspected, and outputs the work image corresponding to the work area to the image acquisition unit 102.

[0119] The image acquisition unit 102 acquires the average value of the luminance values ​​of the non-defective image group 114 as the reference luminance value. The average value of the luminance values ​​may be calculated in advance and stored in the storage unit 11. In this case, it is not necessary to store the non-defective image group 114 in the storage unit 11.

[0120] As one example, the image acquisition unit 102 may extract, as an image area, an area of ​​the workpiece image received from the workpiece area extraction unit 101 that falls below a range determined based on the average value of the acquired brightness values. The image acquisition unit 102 cuts out an inspection image including the extracted image area from the workpiece image. As one example, the area that falls below the range determined based on the average value of the acquired brightness values ​​may be an area that falls below the average value of the acquired brightness values. In another example, the area that falls below the range determined based on the average value of the acquired brightness values ​​may be an area that falls below a value obtained by adding or subtracting a predetermined margin to the above-mentioned average value of the brightness values.

[0121] Even when extracting a region with a low average brightness value, the image rotation unit 103 may be omitted.

[0122] The pass / fail judgment unit 106 receives an inspection image from the image acquisition unit 102 that includes an image area that is below the average value of a range determined based on the brightness value, and judges the pass / fail of the object to be inspected based on the output value obtained by inputting the inspection image into the pass / fail judgment model 113.

[0123] This allows the inspection device 1 to efficiently detect stains and special chipping defects in the phosphor-containing glass.

[0124] The pass / fail determination model 113 used by the pass / fail determination unit 106 may be a model constructed using correct answer data indicating the presence or absence of stains and special chipping that should be determined to be defective. The pass / fail determination model 113 is constructed by the teacher data generation unit 104 and the learning unit 105.

[0125] Specifically, in this embodiment, the image acquisition unit 102 may be configured to acquire a learning image cut out from an image of an optical material, which is the object to be inspected, so as to include an image region extracted based on the average brightness value (S3 to S4 in FIG. 4).

[0126] In this embodiment, the pass / fail judgment model 113 used by the pass / fail judgment unit 106 for pass / fail judgment is constructed using the learning images acquired by the image acquisition unit 102. That is, the learning images used in this embodiment are images cut out from the captured image so as to include an image region extracted based on the average brightness value.

[0127] The teacher data generation unit 104 associates the correct answer data with the learning images acquired as described above to generate teacher data 112 (S7). The learning unit 105 constructs a pass / fail determination model 113 by supervised learning using the teacher data 112. Note that the pass / fail determination model 113 may be constructed for each type of defect.

[0128] (Image region extraction based on overall average brightness value) <Extracting areas with relatively high brightness values> The image acquiring unit 102 may use an overall average luminance value, which is the average value of the luminance values ​​of the entire captured image of the object to be inspected, as the reference luminance value.The image acquiring unit 102 may then extract, as an image region, a region from the captured image whose average luminance value exceeds a range determined based on the overall average luminance value.

[0129] According to the above-described configuration, areas with relatively high brightness values ​​are extracted as image areas from the captured image of the object under inspection. Therefore, it is possible to narrow down the image area to be judged pass / fail without being affected by changes in imaging conditions or brightness fluctuations due to individual differences in the object under inspection. This allows for efficient detection of defects that appear brighter than normal areas. Examples of such defects include aggregation of phosphor particles in optical materials.

[0130] 11 , in the inspection device 1 according to this embodiment, the storage unit 11 may store a workpiece image 115. In this embodiment, the workpiece area extraction unit 101 extracts a workpiece area from a captured image 111 of the phosphor-containing glass, which is the object to be inspected, and stores in the storage unit 11 a workpiece image 115 corresponding to the workpiece area.

[0131] The image acquisition unit 102 acquires the overall average luminance value of the workpiece image 115 as a reference luminance value. As shown in FIG. 12, the image acquisition unit 102 may acquire the average of the luminance values ​​of each pixel in the entire area (AreaAll) of the workpiece image 115 as the overall average luminance value of the workpiece image 115. Alternatively, as shown in FIG. 12, the image acquisition unit 102 may first acquire, as a partial average luminance value, the average of the luminance values ​​of the pixels in each partial area (e.g., Area1, Area2, . . ., Area16) obtained by partitioning the workpiece image 115. Then, the image acquisition unit 102 may acquire, as the overall average luminance value of the workpiece image 115, the average of the partial average luminance values ​​for each partial area (the partial average luminance value of Area1, the partial average luminance value of Area2, . . .).

[0132] Then, as one example, the image acquisition unit 102 may extract, as an image area, an area of ​​the workpiece image received from the workpiece area extraction unit 101 that exceeds a range determined based on the acquired overall average brightness value. The image acquisition unit 102 cuts out an inspection image including the extracted image area from the workpiece image 115. The area that exceeds the range determined based on the overall average brightness value may, as one example, be an area that exceeds the overall average brightness value. In another example, the area that exceeds the range determined based on the overall average brightness value may be an area that exceeds a value obtained by adding or subtracting a predetermined margin to the above-mentioned overall average brightness value.

[0133] In another example, when the workpiece image 115 is divided into several partial regions, the image acquisition unit 102 may refer to each of the partial average luminance values ​​obtained for each partial region as a reference luminance value. That is, the image acquisition unit 102 may compare each partial region of the workpiece image received from the workpiece region extraction unit 101 with the partial average luminance value of the corresponding partial region, and extract, as an image region, a region in the partial region that exceeds a range determined based on the partial average luminance value. The region that exceeds the range determined based on the partial average luminance value may be, for example, a region that exceeds the partial average luminance value, or a region that exceeds a value obtained by adding or subtracting a predetermined margin to or from the partial average luminance value.

[0134] In the inspection device 1 of FIG. 11, the image rotation unit 103 may be omitted.

[0135] The pass / fail judgment unit 106 receives from the image acquisition unit 102 an inspection image including an image region that exceeds a range determined based on the overall average luminance value, and judges the pass / fail of the object to be inspected based on an output value obtained by inputting the inspection image into the pass / fail judgment model 113. The pass / fail judgment model 113 used by the pass / fail judgment unit 106 may be a model constructed using correct answer data indicating the presence or absence of aggregation of phosphor particles that should be judged as failing.

[0136] This allows the inspection device 1 to efficiently detect defects in aggregation of phosphor particles in phosphor-containing glass without being affected by changes in imaging conditions or fluctuations in brightness due to individual differences in the objects to be inspected.

[0137] The pass / fail determination model 113 used by the pass / fail determination unit 106 may be a model constructed using correct answer data indicating the presence or absence of aggregation of phosphor particles that should be determined to be defective. The pass / fail determination model 113 is constructed by the teacher data generation unit 104 and the learning unit 105.

[0138] Specifically, in this embodiment, the image acquisition unit 102 may be configured to acquire a learning image cut out from an image of an optical material, which is the object to be inspected, so as to include an image region extracted based on the overall average brightness value (S3 to S4 in FIG. 4).

[0139] In this embodiment, the pass / fail judgment model 113 used by the pass / fail judgment unit 106 for pass / fail judgment is constructed using the learning images acquired by the image acquisition unit 102. That is, the learning images used in this embodiment are images cut out from the captured image so as to include an image region extracted based on the overall average brightness value.

[0140] The training data generation unit 104 associates the training images acquired as described above with the correct answer data to generate training data 112 (S7). The learning unit 105 constructs a pass / fail determination model 113 by supervised learning using the training data 112.

[0141] <Extracting areas with relatively low brightness values> The image acquiring unit 102 may use an overall average luminance value, which is the average value of the luminance values ​​of the entire captured image of the inspection object itself, as the reference luminance value.The image acquiring unit 102 may then extract, as an image region, a region from the captured image whose average luminance value is below a range determined based on the overall average luminance value.

[0142] According to the above-described configuration, areas with relatively low brightness values ​​are extracted as image areas from the captured image of the object under inspection. Therefore, it is possible to narrow down the image area to be judged as pass / fail without being affected by changes in imaging conditions or brightness fluctuations due to individual differences in the object under inspection. This allows for efficient detection of defects that appear darker than normal areas. Examples of such defects include dirt adhering to optical materials. It is also possible to detect defects caused by special chipping.

[0143] As shown in FIG. 11, the image acquisition unit 102 acquires (calculates) the overall average brightness value of the workpiece image 115 as the reference brightness value.

[0144] Then, as one example, the image acquisition unit 102 may extract, as an image area, an area of ​​the workpiece image 115 extracted by the workpiece area extraction unit 101 and stored in the storage unit 11 that is below a range determined based on the calculated overall average luminance value. The image acquisition unit 102 cuts out an inspection image including the extracted image area from the workpiece image 115. The area below the range determined based on the overall average luminance value may be, for example, an area below the overall average luminance value. In another example, the area below the range determined based on the overall average luminance value may be an area below a value obtained by adding or subtracting a predetermined margin to the above-mentioned overall average luminance value.

[0145] Even when extracting a region with a relatively low brightness value, the image rotation unit 103 may be omitted.

[0146] The pass / fail judgment unit 106 receives from the image acquisition unit 102 an inspection image including an image region below a range determined based on the overall average brightness value, and judges the pass / fail of the object to be inspected based on an output value obtained by inputting the inspection image into the pass / fail judgment model 113. The pass / fail judgment model 113 used by the pass / fail judgment unit 106 may be a model constructed using correct answer data indicating the presence or absence of stains and special chipping that should be judged as failing.

[0147] This allows the inspection device 1 to efficiently detect stains and special chipping defects in phosphor-containing glass without being affected by changes in imaging conditions or fluctuations in brightness due to individual differences in the objects being inspected.

[0148] The pass / fail determination model 113 used by the pass / fail determination unit 106 may be a model constructed using correct answer data indicating the presence or absence of stains and special chipping that should be determined to be defective. The pass / fail determination model 113 is constructed by the teacher data generation unit 104 and the learning unit 105.

[0149] Specifically, in this embodiment, the image acquisition unit 102 may be configured to acquire a learning image cut out from an image of an optical material, which is the object to be inspected, so as to include an image region extracted based on the overall average brightness value (S3 to S4 in FIG. 4).

[0150] In this embodiment, the pass / fail judgment model 113 used by the pass / fail judgment unit 106 for pass / fail judgment is constructed using the learning images acquired by the image acquisition unit 102. That is, the learning images used in this embodiment are images cut out from the captured image so as to include an image region extracted based on the overall average brightness value.

[0151] The teacher data generation unit 104 associates the correct answer data with the learning images acquired as described above to generate teacher data 112 (S7). The learning unit 105 constructs a pass / fail determination model 113 by supervised learning using the teacher data 112. Note that the pass / fail determination model 113 may be constructed for each type of defect. Third Embodiment Other embodiments of the present invention will be described below. For ease of explanation, the same reference numerals will be used to designate components having the same functions as those described in the above embodiment, and the description thereof will not be repeated.

[0152] The present embodiment aims to efficiently detect defects in optical materials that result in irregular colors or shapes, and to more efficiently detect defects that are uniformly spread over a slightly wider area than the localized defects assumed in embodiment 1.

[0153] For this purpose, the image acquisition unit 102 of the inspection device 1 according to this embodiment is configured to acquire, as an inspection image, an image cut out from the captured image so as to include an image area extracted based on the comparison result of the brightness values ​​between the captured image of the optical material, which is the object to be inspected, and a reference image of a good product.

[0154] [Configuration of inspection device] In the inspection device 1 of this embodiment, the image acquisition unit 102 acquires an inspection image that is cut out to include an image area extracted based on the comparison result of brightness values ​​between an image captured of the optical material, which is the object to be inspected, and one or more reference images, which are images captured by capturing good optical materials.

[0155] The pass / fail determination unit 106 determines the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail determination model 113 constructed by machine learning. The pass / fail determination model 113 is, for example, a pass / fail determination model constructed by machine learning using training data that includes at least captured images of defective optical materials.

[0156] More specifically, the pass / fail determination unit 106 inputs the inspection image, which has been enlarged or reduced to the same size as the learning image used as training data for the pass / fail determination model 113, to the pass / fail determination model 113. Alternatively, the image acquisition unit 102 may acquire an inspection image that includes an image region extracted based on the result of comparing the brightness values ​​with the reference image and has the same size as the learning image. In this case, the pass / fail determination unit 106 may input the inspection image acquired by the image acquisition unit 102 to the pass / fail determination model 113 without resizing it.

[0157] [Testing method] The inspection method executed by the inspection device 1 according to this embodiment includes S14 (image acquisition step) and S18 (determining pass / fail) similarly to the inspection method according to embodiment 1 (see FIG. 9). However, the content of the process in S14 differs from that of the inspection method according to embodiment 1 as follows.

[0158] In this embodiment, in S14, the image acquisition unit 102 acquires an inspection image that is cut out to include an image area extracted based on the comparison result of brightness values ​​between an image captured of the optical material that is the object to be inspected and one or more reference images that are images captured of the optical material in a good condition.

[0159] In S18, the quality determination unit 106 determines the quality of the optical material based on the output value obtained by inputting the inspection image into a quality determination model 113 constructed by machine learning.

[0160] The reference image may be an image captured using the same imaging technique and under the same conditions as when the above-mentioned captured image was obtained.

[0161] According to the above-described configuration and method, when a defect occurs in an optical material, causing the color or shape to become irregular, the brightness value of the area where the defect occurs will differ from the reference image obtained by capturing an optical material without the defect. Therefore, the above-described configuration makes it possible to detect such defects.

[0162] In particular, depending on the imaging method and conditions, there can be a bias in brightness depending on the location within a single captured image. For example, in coaxial epi-illumination photography, the center of the image tends to be bright and the edges dark. Therefore, by comparing the brightness values ​​at the same positions with a reference image captured using the same imaging method and conditions, it is possible to narrow down the image area to be judged as pass / fail without being affected by the bias in brightness that occurs within the same image.

[0163] As shown in FIG. 13, in the inspection device 1 according to this embodiment, the memory unit 11 may store a reference image 116. The reference image 116 is a captured image obtained by capturing an image of a non-defective optical material. In this embodiment, as an example, the reference image 116 is a work image of a non-defective workpiece W extracted from a captured image 111 of a non-defective phosphor-containing glass. A plurality of reference images 116 may be stored. In this embodiment, the reference image 116 is cut out from the captured image 111 (obtained by capturing an image of an optical material that is known to be non-defective) so as not to include the groove G.

[0164] In this embodiment, the work area extraction unit 101 extracts a work area from a captured image 111 of the phosphor-containing glass, which is the object to be inspected, and outputs a work image WP corresponding to the work area to the image acquisition unit 102. In this embodiment, the work image WP is cut out from the captured image 111 so as not to include the groove G (see FIG. 2).

[0165] The image acquisition unit 102 compares the workpiece image WP received from the workpiece area extraction unit 101 with the reference image 116 stored in the storage unit 11. As an example, as shown in Fig. 14, the image acquisition unit 102 compares the luminance values ​​of corresponding pixels between the reference image 116 and the workpiece image WP, and obtains a difference image DP that represents the difference in luminance values. In the illustrated example, the greater the difference in luminance values, or the smaller the luminance value of the workpiece image WP is compared to the luminance value of the reference image 116, the brighter (higher luminance value) the corresponding pixels will be in the obtained difference image DP.

[0166] The reference image of a non-defective piece of phosphor-containing glass is generally dominated by bright areas. In phosphor-containing glass, areas with defects (such as scratches or stains) have lower brightness values ​​than non-defective areas. Therefore, by subtracting the brightness values ​​of the workpiece image WP from the brightness values ​​of the reference image 116, a difference image DP is obtained in which the low-brightness areas where defects are likely to exist are bright and prominent.

[0167] In another example, the image acquisition unit 102 may compare the reference image 116 with the workpiece image WP to obtain a difference image DP based on the absolute value of the difference in brightness values. In this case, the difference image DP obtained is one in which both excessively dark and excessively bright areas are brighter and more pronounced than the brightness values ​​of the reference image. Using the difference image DP obtained in this manner, it is possible to detect defects in optical materials that appear excessively dark, as well as defects that appear excessively bright (such as phosphor particle agglomerations).

[0168] The image acquisition unit 102 acquires an inspection image based on the results of the comparison described above. More specifically, the image acquisition unit 102 extracts an area that is brighter than a predetermined value in the difference image DP that indicates the comparison result described above as an image area, and acquires an inspection image cut out from the workpiece image WP so as to include that image area. In other words, the image acquisition unit 102 can cut out the inspection image from the workpiece image WP so as to include a portion of the workpiece image WP that appears excessively dark compared to the reference image 116.

[0169] Furthermore, the image acquisition unit 102 may acquire a learning image based on a difference image obtained by comparing a reference image of a non-defective product with a workpiece image.

[0170] Specifically, in this embodiment, the image acquisition unit 102 may extract an image region based on a comparison result of brightness values ​​between a reference image, which is a captured image of a non-defective optical material, and a captured image of the optical material to be inspected. Then, the image acquisition unit 102 may be configured to acquire a learning image cut out so as to include the extracted image region (S3 to S4 in FIG. 4).

[0171] In this embodiment, the pass / fail judgment model 113 used by the pass / fail judgment unit 106 for pass / fail judgment is constructed using the learning images acquired by the image acquisition unit 102. That is, the learning images used in this embodiment are images cut out from the captured images so as to include an image region extracted based on the comparison result of brightness values ​​between a reference image of a non-defective product and a captured image of the inspection object.

[0172] The training data generation unit 104 associates the training images acquired as described above with the correct answer data to generate training data 112 (S7). The learning unit 105 constructs a pass / fail determination model 113 by supervised learning using the training data 112.

[0173] In yet another example, the storage unit 11 of the inspection device 1 may store a reference image 116 for each address of the workpiece W on the optical material when a captured image 111 of the entire non-defective optical material is obtained. The address of the workpiece W is set for each position of the workpiece W on the optical material, and the address can be used to identify which position on the optical material the workpiece W is at.

[0174] In the inspection system 100 shown in FIG. 2, when imaging is performed using the coaxial epi-illumination method adopted in the imaging device 2, as shown in FIG. 15, the center of the captured image 111 tends to be brightest and the image tends to become darker toward the edges. Thus, if the workpiece image WP is compared with the reference image 116 without taking into account the light-dark bias caused by coaxial epi-illumination imaging, the workpiece image WP located near the edges of the captured image 111 will be judged to be darker than necessary when compared with the reference image 116. Therefore, areas of the workpiece image WP that are judged to be dark may be judged to be defective even though they are actually good products. For accurate quality determination, it is desirable to compare the reference image 116 with the workpiece image WP after absorbing the light-dark bias caused by such imaging conditions.

[0175] Therefore, in this embodiment, the reference image 116 of the workpiece W may be stored in the storage unit 11 for each address of the workpiece W at the time of photographing.

[0176] For example, as shown in FIG. 15, a captured image 111 of phosphor-containing glass includes workpiece images WP of 12 workpieces W arranged in four columns (columns A to D) and three rows (columns 1 to 3). The center of the captured image 111 tends to appear bright, and the images tend to appear darker as they move away from the center. For example, the workpiece image WP (hereinafter referred to as workpiece image WP_A1) of the workpiece W located at address A1 in column A (hereinafter referred to as address A1) tends to appear dark in approximately half of the upper left area near the edge. Therefore, as a reference image for address A1, which always appears in this manner, an image obtained by photographing a non-defective workpiece W located at address A1 using a coaxial epi-illumination method is used. Hereinafter, the reference image for address A1 will be referred to as reference image 116_A1. An example of reference image 116_A1 is shown in FIG. 16.

[0177] The image acquisition unit 102 receives the workpiece image WP from the work area extraction unit 101 along with address information indicating the address. When the image acquisition unit 102 receives the workpiece image WP_A1 at address A1, it reads out the reference image 116_A1 for address A1 stored in the memory unit 11. The image acquisition unit 102 then compares the workpiece image WP_A1 to be inspected with the reference image 116_A1. As an example, as shown in FIG. 16, the image acquisition unit 102 compares the luminance values ​​of pixels between the reference image 116_A1 and the workpiece image WP_A1, and obtains a difference image DP that represents the difference in luminance values. In the illustrated example, the greater the difference in luminance values, specifically, the smaller the luminance value of the workpiece image WP compared to the luminance value of the reference image 116, the brighter the corresponding pixels in the difference image DP that is obtained.

[0178] If this workpiece image WP_A1 were compared with the uniformly prepared reference image 116 shown in FIG. 14, a difference image DP would be obtained in which approximately half of the dark area in the upper left region is brightened and emphasized. The original dark areas, which are inverted and brightened in this way, are dark areas caused by the shooting conditions, not dark areas caused by dirt or special chipping. Therefore, it may be difficult to accurately determine the quality of the dark areas. However, with the above-described configuration, the workpiece image WP_A1 is compared with the reference image 116 prepared for each address, thereby eliminating the imbalance in brightness caused by the address. This allows the image acquisition unit 102 to accurately identify image areas by focusing on dark areas that are likely to be caused by defects such as dirt or special chipping, rather than focusing on dark areas caused by the shooting conditions and unrelated to defects. As a result, the quality determination unit 106 can accurately determine the quality of the inspection image containing the identified image area.

[0179] [Software implementation example] The control block of the inspection device 1 (particularly, each unit included in the control unit 10) may be realized by a logic circuit (hardware) formed on an integrated circuit (IC chip) or the like, or may be realized by software.

[0180] In the latter case, the inspection device 1 includes a computer that executes instructions in a program (inspection program), which is software that realizes each function. The program causes the computer to function as the inspection device 1. The computer includes, for example, one or more processors and a computer-readable recording medium that stores the program. The object of the present invention is achieved when the processor in the computer reads and executes the program from the recording medium. The processor may be, for example, a central processing unit (CPU). The recording medium may be a "non-transitory tangible medium," such as a read-only memory (ROM), tape, disk, card, semiconductor memory, or programmable logic circuit. The computer may also include a random access memory (RAM) for loading the program. The program may be supplied to the computer via any transmission medium capable of transmitting the program (such as a communication network or broadcast waves). Note that one aspect of the present invention may also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission.

[0181] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

[0182] [Additional Notes] The inspection device according to aspect 1 of the present disclosure includes an image acquisition unit that acquires an inspection image cut out from an image of an optical material, which is an object to be inspected, so as to include an image region extracted based on an average brightness value, and a pass / fail judgment unit that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning, and the image acquisition unit extracts, as the image region, a region from among the regions constituting the captured image where the average brightness value of the region falls outside a range determined based on a reference brightness value.

[0183] In the inspection device according to aspect 2 of the present disclosure, in the above-described aspect 1, the reference brightness value is the average brightness value of the area in which the optical material appears in the captured image obtained by capturing an image of the optical material in a good condition, and the image acquisition unit extracts as the image area an area from among the areas constituting the captured image where the average brightness value of the area exceeds a range determined based on the reference brightness value.

[0184] In the inspection device according to aspect 3 of the present disclosure, in the above-described aspect 1, the reference brightness value is the average brightness value of the area in which the optical material appears in the captured image obtained by capturing an image of the optical material in a good condition, and the image acquisition unit extracts as the image area an area from among the areas constituting the captured image where the average brightness value of the area is below a range determined based on the reference brightness value.

[0185] In the inspection device according to aspect 4 of the present disclosure, in the above-described aspect 1, the reference brightness value is an overall average brightness value, which is the average value of the brightness values ​​of the entire captured image of the object to be inspected, and the image acquisition unit extracts, as the image area, an area from among the areas constituting the captured image in which the average brightness value of the area exceeds a range determined based on the overall average brightness value.

[0186] In the inspection device according to aspect 5 of the present disclosure, in aspect 1 above, the reference brightness value is an overall average brightness value, which is the average brightness value of the entire captured image of the object to be inspected, and the image acquisition unit extracts, as the image area, an area from among the areas constituting the captured image where the average brightness value of the area is below a range determined based on the overall average brightness value.

[0187] The inspection device according to aspect 6 of the present disclosure includes an image acquisition unit that acquires an inspection image cut out to include an image region extracted based on a comparison result of brightness values ​​between an image of an optical material that is an object to be inspected and one or more reference images that are images obtained by imaging the optical material in a good condition, and a pass / fail judgment unit that judges the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning.

[0188] An inspection method according to aspect 7 of the present disclosure is an inspection method performed by an inspection device, and includes an image acquisition step of acquiring an inspection image cut out from an image of an optical material, which is an object to be inspected, so as to include an image region extracted based on an average value of brightness values, and a pass / fail judgment step of judging the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning, wherein in the image acquisition step, a region from among the regions constituting the captured image, where the average brightness value of the region falls outside a range determined based on a reference brightness value, is extracted as the image region.

[0189] An inspection method according to aspect 8 of the present disclosure is an inspection method performed by an inspection device, and includes an image acquisition step of acquiring an inspection image cut out to include an image region extracted based on a comparison result of brightness values ​​between an image of an optical material that is an object to be inspected and one or more reference images that are images obtained by imaging the optical material in a good condition, and a pass / fail judgment step of inputting the inspection image into a pass / fail judgment model constructed by machine learning and judging the pass / fail of the optical material based on an output value obtained.

[0190] The inspection program of aspect 9 of the present disclosure is an inspection program for causing a computer to function as the inspection device of aspect 1 or aspect 6 above, and is an inspection program for causing a computer to function as the image acquisition unit and the pass / fail judgment unit. [Explanation of symbols]

[0191] 1. Inspection equipment 102 Image acquisition unit 103 Image Rotation Unit 106 Good / bad judgement section 111 Captured images 113 Good / bad judgement model

Claims

1. an image acquisition unit that acquires an inspection image cut out from an image of an optical material that is an inspection object so as to include an image region extracted based on the variation in brightness values; a quality determination unit that determines the quality of the optical material based on an output value obtained by inputting the inspection image into a quality determination model constructed by machine learning, the optical material is a wavelength conversion member workpiece containing a phosphor, The pass / fail judgment model is constructed to judge the presence or absence of special chipping that occurs at the end of the workpiece and is not formed by broken pieces peeling off, in this inspection device.

2. 2. The inspection device according to claim 1, wherein the image acquisition unit identifies an image area among the plurality of image areas in which one of the optical materials is captured that has a relatively large variation in brightness values, and acquires an inspection image cut out from the captured image so as to include the identified image area.

3. 3. The inspection device according to claim 1, wherein the image acquisition unit acquires the inspection image cut out from the captured image so as to include each image area set along the outer edge of the optical material shown in the captured image, based on the variation in brightness values ​​in the image area.

4. an image rotation unit that rotates the inspection image so that the position of the peripheral portion of the optical material shown in the inspection image is uniform on a predetermined side of the four sides, i.e., top, bottom, left, and right, of the inspection image; 4. The inspection device according to claim 3, wherein the quality determination unit determines the quality of the optical material based on an output value obtained by inputting the inspection image rotated by the image rotation unit into the quality determination model.

5. 5. The inspection device according to claim 1, wherein the quality determination unit determines the quality of the optical material based on an output value obtained by inputting the inspection image into the quality determination model that distinguishes between and detects multiple types of defects.

6. The inspection device according to claim 1 , wherein the image acquisition unit acquires the inspection image of a size smaller than the smallest size of the workpiece among the plurality of types of workpieces with different sizes.

7. 7. The inspection device according to claim 1, wherein the wavelength conversion member containing a phosphor is a phosphor-containing glass.

8. 8. The inspection device according to claim 1, wherein the quality determination model is constructed to determine whether or not there is dirt adhering to the workpiece.

9. 1. An inspection method performed by an inspection device, comprising: an image acquisition step of acquiring an inspection image cut out from an image of an optical material as an inspection object so as to include an image region extracted based on the variation in brightness values; a pass / fail judgment step of judging the pass / fail of the optical material based on an output value obtained by inputting the inspection image into a pass / fail judgment model constructed by machine learning, the optical material is a wavelength conversion member workpiece containing a phosphor, An inspection method in which the pass / fail judgment model is constructed to judge the presence or absence of special chipping that occurs at the end of the workpiece and is not formed by broken pieces peeling off.

10. 2. An inspection program for causing a computer to function as the inspection apparatus according to claim 1, the inspection program causing the computer to function as the image acquisition unit and the quality determination unit.

Images