Inspection system, inspection method, and program

Abstract

The present invention improves the efficiency of inspection in which there are a plurality of inspection objects.　This inspection system includes: a first map generation unit that extracts a first feature amount from a first image including a plurality of inspection objects, and that generates a first feature amount map; a first score calculation unit that uses the first feature amount map to calculate a first score corresponding to each of the inspection objects; and a defective product identification unit that, on the basis of the first score, identifies an inspection object that is a defective product.

Description

Inspection system, inspection method, and program

[0001] The present invention relates to an inspection system, an inspection method, and a program.

[0002] Patent Document 1 discloses an information processing device that presents the basis for a determination together with the result of determining whether an inspection target included in an input image is abnormal. This information processing device generates a plurality of score maps indicating scores for each feature amount from an image including an inspection target, and integrates the generated plurality of score maps to generate a defect display image representing a defect that serves as a basis for determining that the inspection target is abnormal.

[0003] Japanese Unexamined Patent Application Publication No. 2016-114592

[0004] For example, when an inspection target for image-based inspection is small, a plurality of inspection targets may be inspected simultaneously using an image containing the plurality of inspection targets. In such a case, it is difficult to perform efficient inspection using a system that determines whether a single inspection target is defective, as in Patent Document 1. Further, when a plurality of inspection targets are inspected visually by an operator, in order to identify the inspection target that the operator has determined to be defective, the operator needs to manually select the defective product or use an eye tracker or the like. For this reason, there was room for improvement in work efficiency.

[0005] Accordingly, an object of the present invention is to improve the efficiency of inspection when there are a plurality of inspection targets.

[0006] This specification includes the entire content of Japanese Patent Application No. 2025-053088 filed on March 27, 2025. One aspect of the present invention is an inspection system comprising: a first map generation unit that extracts a first feature amount from a first image including a plurality of inspection targets and generates a first feature amount map; a first score calculation unit that calculates a first score corresponding to each of the inspection targets using the first feature amount map; and a defective product identification unit that identifies the inspection target that is a defective product based on the first score.

[0007] Another aspect of the present invention is an inspection method comprising: a first step in which a computer extracts first features from a first image containing a plurality of objects to be inspected and generates a first feature map; a second step in which the computer uses the first feature map to calculate a first score corresponding to each of the objects to be inspected; and a third step in which the computer identifies the objects to be inspected that are defective based on the first score.

[0008] Another aspect of the present invention is a program that causes the processor to function as: a first map generation unit that extracts first features from a first image containing a plurality of objects to be inspected and generates a first feature map; a first score calculation unit that calculates a first score corresponding to each of the objects to be inspected using the first feature map; and a defective product identification unit that identifies the objects to be inspected that are defective based on the first score.

[0009] According to the present invention, the efficiency of inspections involving multiple objects can be improved.

[0010] Figure 1 is a diagram showing the configuration of the inspection system according to the first embodiment. Figure 2 is a block diagram showing the configuration of the processing unit. Figure 3 is a diagram showing an example of the first image. Figure 4 is a diagram showing an example of the first feature map. Figure 5 is a diagram showing an example of the second image. Figure 6 is a diagram showing an example of the second feature map. Figure 7 is a diagram showing an example of the third image. Figure 8 is a flowchart showing the operation of the processing unit according to the first embodiment. Figure 9 is a flowchart showing the operation of the processing unit according to the second embodiment.

[0011] Embodiments of the present invention will be described below with reference to the drawings. [1. First Embodiment] First, a first embodiment of the present invention will be described. [1-1. Configuration of the Inspection System] Figure 1 is a diagram showing the configuration of the inspection system 1 according to the first embodiment. The inspection system 1 is a system that performs visual inspection on an object to be inspected 90. In this embodiment, the object to be inspected 90 is a transistor.

[0012] The inspection system 1 has a transport device 2. The transport device 2 is a device for transporting the objects to be inspected 90. The transport device 2 has a first transport mechanism 21. The first transport mechanism 21 receives the objects to be inspected 90 transported by a feeder (not shown) or the like, and transports them toward a first collection container 23 for collecting good quality objects to be inspected 90. The first transport mechanism 21 transports multiple objects to be inspected 90 simultaneously. The transport device 2 has a second transport mechanism 22. The second transport mechanism 22 transports the objects to be inspected 90 that have been transported to the first transport mechanism 21 and are determined to be defective toward a second collection container 24. In this embodiment, the first transport mechanism 21 is a circular transport table that transports the objects to be inspected 90 by rotating on it. The objects to be inspected 90 transported by the first transport mechanism 21 are discharged from the first transport mechanism 21 to the first collection container 23 by any means such as a device that blows air. In this embodiment, the second conveying mechanism 22 is a belt conveyor. The first conveying mechanism 21 and the second conveying mechanism 22 can be any type of mechanism capable of conveying the object to be inspected 90.

[0013] The inspection system 1 has an imaging device 3. The imaging device 3 includes a camera, such as a CCD camera or a CMOS camera. The imaging device 3 may consist of one camera or multiple cameras. The imaging device 3 captures an image Im1, which is an image including the object to be inspected 90. In this embodiment, the imaging device 3 captures an image Im1 including the object to be inspected 90 being transported by the first transport mechanism 21.

[0014] The inspection system 1 includes a processing unit 4. The processing unit 4 corresponds to an example of a "computer" in this invention. The processing unit 4 communicates with the transport device 2 and the imaging device 3. The processing unit 4 controls the operation of the transport device 2. The processing unit 4 controls the operation of the imaging device 3. The processing unit 4 also acquires data of the image Im1 captured by the imaging device 3.

[0015] The processing device 4 uses the image Im1 captured by the imaging device 3 to identify the defective product 90. The processing device 4 also identifies the part of the defective product 90 where the defect occurred, i.e., the defective part.

[0016] The inspection system 1 has a rejection mechanism 25. The rejection mechanism 25 is connected to the processing device 4 via communication. The rejection mechanism 25 removes the defective inspection object 90 identified by the processing device 4 from the first transport mechanism 21 and transfers it to the second transport mechanism 22. In this embodiment, the rejection mechanism 25 is a robotic arm. The rejection mechanism 25 can be anything as long as it can remove the inspection object 90 from the first transport mechanism 21. For example, the rejection mechanism 25 may be a device that removes the inspection object 90 using air.

[0017] The inspection system 1 includes a marking device 27. The marking device 27 is connected to the processing device 4. The marking device 27 applies markers to defective areas of the object to be inspected 90 identified by the processing device 4. The marking device 27 is, for example, a device that sprays paint, which serves as a marker, onto the defective areas of the object to be inspected 90.

[0018] The inspection system 1 has a display device 41. The display device 41 is connected to the processing device 4. The display device 41 has a display that shows information such as figures and characters under the control of the processing device 4. The display device 41 displays various information to the visual inspection worker P. The display device 41 corresponds to the "display unit".

[0019] The inspection system 1 has an input device 43. The input device 43 is connected to the processing unit 4. The input device 43 includes a keyboard, mouse, switch, touch sensor, and other input means. The input device 43 detects input operations performed by the operator P on the input means and transmits a detection signal corresponding to the input operation to the processing unit 4. The input device 43 corresponds to the "input unit".

[0020] [1-2. Configuration of the Processing Unit] Figure 2 is a block diagram showing the configuration of the processing unit 4 according to the first embodiment. The processing unit 4 includes a control unit 40 and a communication unit 45. The communication unit 45 is equipped with communication hardware that conforms to a predetermined communication standard, such as a communication circuit, and communicates with each part of the inspection system 1, such as the transport device 2 and the imaging device 3, under the control of the control unit 40.

[0021] The control unit 40 includes a processor 400. The processor 400 is composed of a CPU (Central Processing Unit), an MPU (Micro Processing Unit), and the like.

[0022] The control unit 40 has a memory 420. The memory 420 is a memory that stores programs and data. The memory 420 stores the control program 421, the overall CV model 422, the individual CV model 423, the first reference value 424, and data to be processed by the processor 400. The memory 420 has a non-volatile storage area. The memory 420 also has a volatile storage area and constitutes the work area of ​​the processor 400. The memory 420 is composed of, for example, ROM (Read Only Memory) or RAM (Random Access Memory). The control program 421 corresponds to "program".

[0023] The overall CV (Computer Vision) model 422 is a model that uses a first image Im2, which is an image containing multiple objects 90 to be examined, as input data. The single CV model 423 is a model that uses a second image Im3, which is an image containing one object 90 to be examined, as input data. Details of each CV model 422 and 423 will be described later.

[0024] The first reference value 424 is a reference value for the degree of abnormality of the object 90 being inspected, used to determine whether or not the object 90 is identified as a defective product. The first reference value 424 is compared with the first score described later.

[0025] The control program 421 is executed by the processor 400, causing the processor 400 to function as a communication control unit 401, an equipment control unit 402, a first image generation unit 403, a first map generation unit 404, a first score calculation unit 405, a defective product identification unit 406, a second image generation unit 407, a second map generation unit 408, a second score calculation unit 409, a defective part identification unit 410, and a third image generation unit 411.

[0026] [1-2-1. Communication Control Unit] The communication control unit 401 communicates with the transport device 2, the imaging device 3, the exclusion mechanism 25, the marking device 27, the display device 41, and the input device 43 via the communication unit 45.

[0027] [1-2-2. Equipment Control Unit] The equipment control unit 402 controls the transport device 2, the imaging device 3, the rejection mechanism 25, the marking device 27, and the display device 41. It also detects the detection signal transmitted from the input device 43. In the inspection process of the object to be inspected 90, the equipment control unit 402 transports the object to be inspected 90 using the transport device 2. In the inspection process of the object to be inspected 90, the equipment control unit 402 also causes the imaging device 3 to photograph the object to be inspected 90 at predetermined time intervals.

[0028] [1-2-3. First Image Generation Unit] The first image generation unit 403 modifies the image Im1 captured by the imaging device 3 and generates the first image Im2 which is input to the overall CV model 422.

[0029] Figure 3 shows an example of the first image Im2. As shown in Figure 3, the first image Im2 is an image containing multiple inspection objects 90 captured by the imaging device 3. In this embodiment, the first image Im2 contains nine inspection objects 90 arranged in a grid. Hereafter, when distinguishing between the inspection objects 90 in the first image Im2, they will be referred to as inspection objects 90A to 90I. In the example in Figure 3, inspection object 90G is a defective product, and the other inspection objects 90 are good products. Note that the number of inspection objects 90 contained in the first image Im2 does not need to be two or more, and their arrangement can be set arbitrarily. Also, the number and arrangement of inspection objects 90 may differ for each first image Im2.

[0030] The first image generation unit 403 performs arbitrary modifications, such as adjusting the image size, on the image Im1 captured by the imaging device 3 to generate the first image Im2. The first image generation unit 403 may generate the first image Im2 using only one image Im1, or it may generate the first image Im2 using multiple images Im1. In other words, not all of the multiple inspection objects 90 included in the first image Im2 necessarily need to be included in a single image Im1. For example, the first image generation unit 403 may combine multiple images Im1, each containing one inspection object 90, to generate a single first image Im2.

[0031] Furthermore, if image Im1 includes multiple objects 90 to be inspected, the image Im1 may be used as is as the first image Im2, which is the input to the overall CV model 422. In this case, the first image generation unit 403 can be omitted.

[0032] [1-2-4. First Map Generation Unit] The first map generation unit 404 receives the first image Im2 as input to the overall CV model 422 and generates the first feature map M1. The overall CV model 422 is a model that receives the first image Im2 as input. The overall CV model 422 extracts first features from each part of the input first image Im2. The first features are features that indicate the degree of abnormality of the object to be inspected 90 in each part of the first image Im2. The degree of abnormality of the object to be inspected 90 is greater when the object to be inspected 90 is defective compared to when the object to be inspected 90 is good.

[0033] In this embodiment, the overall CV model 422 is a pre-trained machine learning model that utilizes a CNN (Convolutional Neural Network). In this embodiment, the overall CV model 422 is trained as a classification model that receives a first image Im2 as input and determines whether or not the received first image Im2 contains a defective product to be inspected 90.

[0034] The overall CV model 422 is trained using multiple sets of training images, for example, each containing multiple objects to be inspected 90, and training labels indicating whether or not the training image contains defective objects 90. For example, the training image contains the same number and arrangement of objects to be inspected 90 as the first image Im2. If the number and arrangement of objects to be inspected 90 in the first image Im2 are not fixed but vary, the overall CV model 422 may be trained using multiple training images with different patterns of the number and arrangement of objects to be inspected 90. Also, for example, the training labels may be the result of an inspector P looking at the training image and determining whether or not the training image contains defective objects 90.

[0035] The first feature map M1 is data in which the values ​​of the first features extracted from the first image Im2 by the overall CV model 422 are arranged in a format that allows for mapping to each position in the first image Im2. In this embodiment, the first feature map M1 is a two-dimensional array whose elements are the values ​​of the first features, and each element corresponds to each position or range in the first image Im2. More specifically, in this embodiment, each element of the first feature map M1 is mapped one-to-one with each pixel in the first image Im2.

[0036] In this embodiment, the first map generation unit 404 calculates an activation map of the overall CV model 422 to which the first image Im2 has been input, as the first feature map M1. The first map generation unit 404 applies a decision basis visualization method such as CAM (Class Activation Mapping) or Grad CAM (Gradient-weighted CAM) to the overall CV model 422 to which the first image Im2 has been input, and calculates the first feature map M1.

[0037] Figure 4 shows an example of the first feature map M1, and is a heatmap representation of the first feature map M1. The first feature map M1 shown in Figure 4 corresponds to the first image Im2 illustrated in Figure 3. In Figure 4, each element of the first feature map M1 is plotted to overlap with the corresponding position in the first image Im2 of Figure 3, which is indicated by a dashed line. In the heatmap of Figure 4, the first feature map M1 is plotted so that areas with large values ​​for each element are closer to black.

[0038] In this embodiment, a larger value for the first feature indicates a greater degree of abnormality. Therefore, as shown in the example in Figure 4, the value of the first feature corresponding to the position overlapping with the defective inspection object 90G in the first feature map M1 will be larger than the values ​​of the surrounding first features.

[0039] [1-2-5. First Score Calculation Unit] The first score calculation unit 405 calculates a first score, which is the score corresponding to each of the inspection target objects 90 contained in the first image Im2, based on the first feature map M1. The first score is a value that quantifies the degree of abnormality of the corresponding inspection target object 90.

[0040] The first score calculation unit 405 sets a first map region R1 in the first feature map M1, which is the region associated with the object to be inspected 90. The first score calculation unit 405 sets the first map region R1 for each object to be inspected 90 included in the first image Im2.

[0041] In detail, the first map region R1 is the region of the first feature map M1 that corresponds to the first image region A1 in the first image Im2. The first image region A1 is a region within the first image Im2, with one region set for each object 90 to be inspected. Each first image region A1 includes the entirety of one object 90 to be inspected. Furthermore, each first image region A1 is set so as not to include two or more objects 90 to be inspected.

[0042] As described above, in the present embodiment, the first image Im2 includes nine inspection objects 90A to 90I arranged in a grid. Therefore, in the present embodiment, as shown in FIG. 3, the first score calculation unit 405 sets nine first image regions A1 arranged in a grid for the first image Im2. Similarly, in the present embodiment, the first score calculation unit 405 sets nine first map regions R1 arranged in a grid as first map regions R1, which are regions corresponding to each first image region A1 in the first feature map M1. Note that if the number and arrangement of the inspection objects 90 differ for each first image Im2, any object detection method such as R-CNN (Region Based CNN) may be used to detect the inspection objects 90 in the first image Im2, and a region including one of each detected inspection object 90 may be set as the first image region A1.

[0043] Hereinafter, when distinguishing the first image regions A1, the first image regions A1 each including one of the inspection objects 90A to 90I are referred to as first image regions A1A to A1I with reference numerals. Similarly, when distinguishing the first map regions R1, the first map regions R1 corresponding to the first image regions A1A to A1I are referred to as first map regions R1A to R1I with reference numerals.

[0044] The first score calculation unit 405 calculates a first score corresponding to each inspection object 90 using values of the first feature amount within the set range of each first map region R1. As the first score, the first score calculation unit 405 calculates, for example, any representative value such as a total value, an average value, a median value, or a maximum value of the first feature amount within the range of each first map region R1. In the present embodiment, the first score calculation unit 405 calculates an average value of the first feature amount within the range of each first map region R1 as the first score corresponding to each inspection object 90.

[0045] For example, in the example of FIG. 4, the first score calculation unit 405 calculates the first score of the first map region R1G corresponding to the defective inspection object 90G among the nine inspection objects 90 as the highest value.

[0046] [1-2-6. Defective Product Identification Unit] The defective product identification unit 406 identifies, from among the plurality of inspection objects 90 included in the first image Im2, an inspection object 90 that is a defective product based on the first score. In the present embodiment, the defective product identification unit 406 compares the first score calculated for each inspection object 90 by the first score calculation unit 405 with a first reference value 424 stored in a memory 420. The defective product identification unit 406 identifies, as a defective product, an inspection object 90 corresponding to a first score having a value indicating a greater degree of abnormality than that of the first reference value 424.

[0047] In the present embodiment, since a larger first feature amount indicates a greater degree of abnormality than a smaller first feature amount, the defective product identification unit 406 identifies, as a defective product, an inspection object 90 for which the first score is greater than the first reference value 424. Further, unlike the present embodiment, in a case where a smaller first feature amount indicates a greater degree of abnormality than a larger first feature amount, the defective product identification unit 406 identifies, as a defective product, an inspection object 90 for which the first score is less than the first reference value 424.

[0048] [1-2-7. Second Image Generation Unit] The second image generation unit 407 generates a second image Im3 to be input to a single CV model 423 based on the first image Im2. Specifically, the second image generation unit 407 generates the second image Im3 including the inspection object 90 identified as a defective product by the defective product identification unit 406.

[0049] FIG. 5 is a diagram illustrating an example of the second image Im3. As illustrated in FIG. 5, the second image Im3 is an image including one inspection object 90. The inspection object 90 included in the second image Im3 is the inspection object 90 identified as a defective product by the defective product identification unit. In the present embodiment, the second image Im3 includes only one inspection object 90 identified as a defective product and does not include any other inspection objects 90.

[0050] As shown in Figure 5, the object to be inspected 90 has three lead wires 91, 92, and 93, and a main body 94 connected to each of the lead wires 91, 92, and 93. In the example in Figure 5, the defective part, which is the part where the abnormality occurred, is the lead wire 91. In the second image Im3 of Figure 5, the defect in the lead wire 91 is a curvature in a direction different from that of a good product. Other defects that may occur in the object to be inspected 90 include, for example, the length of the lead wires 91 to 93 being outside the specified range, or chipping, scratches, or discoloration occurring in the main body 94. These defects are reflected in the first image Im2 and the second image Im3, and can therefore be detected by the inspection system 1. The lead wires 91, 92, and 93, and the main body 94 correspond to "parts of the object to be inspected".

[0051] In this embodiment, the second image generation unit 407 generates a second image Im3 by extracting a first image region A1 that includes the inspection target 90 identified as a defective product by the defective product identification unit 406. The second image Im3 extracted by the second image generation unit 407 is not limited to matching the first image region A1.

[0052] [1-2-8. Second Map Generation Unit] The second map generation unit 408 receives the second image Im3 as input to the standalone CV model 423 and generates the second feature map M2. The standalone CV model 423 is a model that receives the second image Im3 as input. The standalone CV model 423 is configured to extract second features from each part of the second image Im3. The second features are features that indicate the degree of abnormality of each part 91 to 94 of the object to be inspected 90 in each part of the second image Im3.

[0053] In this embodiment, the standalone CV model 423 is a pre-trained machine learning model that utilizes a CNN. The standalone CV model 423 is trained as a classification model that receives a second image Im3 as input and determines whether or not one of the inspection targets 90 contained in the received second image Im3 is a defective product.

[0054] The standalone CV model 423 is trained using multiple sets of training images, for example, each containing a single object to be inspected 90, and a training label indicating whether or not the object to be inspected 90 in the training image is defective. For example, the training image contains objects to be inspected 90 arranged in the same way as the objects to be inspected 90 in the second image Im3. If the arrangement of the objects to be inspected 90 in the second image Im3 is not fixed but varies, the standalone CV model 423 may be trained using multiple training images with different arrangement patterns of the objects to be inspected 90. Also, for example, the training label may be the result of an inspector P performing a visual inspection looking at the training image and determining whether or not the object to be inspected 90 contained in the training image is defective.

[0055] The second feature map M2 is data in which the values ​​of multiple second features extracted from the second image Im3 by the standalone CV model 423 are arranged in a format that allows for the mapping of each value to each position in the second image Im3. In this embodiment, the second feature map M2 is a two-dimensional array whose elements are the values ​​of the second features, and each element corresponds to each position or range in the second image Im3. More specifically, in this embodiment, each element of the second feature map M2 is mapped one-to-one with each pixel in the second image Im3.

[0056] In this embodiment, the second map generation unit 408 calculates an activation map of the standalone CV model 423, into which the second image Im3 has been input, as the second feature map M2. The second map generation unit 408 applies, for example, a decision basis visualization method such as CAM or GradCAM to the standalone CV model 423 and the second image Im3 to calculate the second feature map M2.

[0057] Figure 6 shows an example of a second feature map M2, which is represented as a heatmap. The second feature map M2 in Figure 6 corresponds to the second image Im3 illustrated in Figure 5. In Figure 6, each element of the second feature map M2 is plotted to overlap with the corresponding position in the second image Im3 in Figure 3, which is shown by a dashed line. In the heatmap of Figure 6, the second feature map M2 is plotted so that areas with large values ​​for each element are closer to black.

[0058] In this embodiment, a larger value for the second feature indicates a greater degree of abnormality than a smaller value. Therefore, as shown in the example in Figure 6, the value of the second feature corresponding to the location near the defective lead wire 91 in the second feature map M2 will be larger than the values ​​of the surrounding second features.

[0059] [1-2-9. Second Score Calculation Unit] The second score calculation unit 409 calculates a second score based on the second feature map M2, which is the score corresponding to each part 91 to 94 of the object to be inspected 90 included in the second image Im3.

[0060] The second score calculation unit 409 sets a second map region R2 in the second feature map M2, which is a region that corresponds to each part 91 to 94 of the object to be inspected 90. The second score calculation unit 409 sets the second map region R2 for each part of the object to be inspected 90 included in the second image Im3.

[0061] In detail, the second map region R2 is the region in the second feature map M2 that corresponds to the second image region A2 in the second image Im3. The second image region A2 is a region within the second image Im3. At least a portion of each part 91 to 94 of the object to be inspected 90 in the second image Im3 is included in one or more second image regions A2.

[0062] In this embodiment, the second image region A2 is set to include at least a portion of the positions where each part 91 to 94 is located, assuming that the object 90 to be inspected in the second image Im3 is a good product. In addition, one second image region A2 is set for each part 91 to 94 in the second image Im3. For this reason, in this embodiment, the second score calculation unit 409 sets four second map regions R2 corresponding to each part 91 to 94 in the second feature map M2. Hereinafter, when distinguishing between the second image regions A2, the second image regions A2 corresponding to each part 91 to 94 will be referred to as second image regions A2A to A2D. Similarly, the second map regions R2 corresponding to second image regions A2A to A2D will be referred to as second map regions R2A to R2D. Each second map area R2 may be an area manually set to match the object to be inspected 90, an area that is evenly divided, or an area that is automatically recognized and set to match the object to be inspected 90.

[0063] The second score calculation unit 409 calculates a second score corresponding to each part of the object to be inspected 90 using the values ​​of the second features within the range of each set second map region R2. The second score calculation unit 409 calculates an arbitrary representative value as the second score, such as the sum, mean, median, or maximum value of the second features within the range of each second map region R2. In this embodiment, the second score calculation unit 409 calculates the mean of the second features within the range of each second map region R2 as the second score of the object to be inspected 90.

[0064] In the example shown in Figure 6, the second score in the second map region R2A, which corresponds to the defective lead wire 91, is the largest among all the second scores.

[0065] [1-2-10. Defective Area Determination Unit] Based on the second score, the defective area identification unit 410 identifies the area where an abnormality occurred, i.e., the defective area, from among the parts 91 to 94 of the object to be inspected 90 in the second image Im3. More specifically, the defective area identification unit 410 identifies the area corresponding to the second map region R2 associated with the second score that indicates the greatest degree of abnormality among all the second scores as the defective area.

[0066] In this embodiment, if the second feature is larger than it is smaller, the degree of abnormality is greater. Therefore, the defective area identification unit 410 identifies the area corresponding to the second map region R2 associated with the largest second score as a defective area. Also, for example, if the degree of abnormality is larger when the second feature is smaller than it is larger, the defective area identification unit 410 identifies the area corresponding to the second map region R2 associated with the smallest second score as a defective area.

[0067] [1-2-11. Third Image Generation Unit] The third image generation unit 411 generates a third image Im4 based on the first image Im2, which is an image that highlights and displays the inspection target 90 that the defective product identification unit 406 has identified as a defective product.

[0068] Figure 7 shows an example of the third image Im4. In the example in Figure 7, the third image Im4 is shown when the defective product identification unit 406 identifies the inspection target 90G as a defective product, based on the first image Im2 illustrated in Figure 3. The third image generation unit 411 generates the third image Im4 by making the transparency of the inspection target 90 that the defective product identification unit 406 did not identify as defective greater than the transparency of the inspection target 90 that the defective product identification unit 406 identified as defective. In detail, the third image generation unit 411 makes the transparency of the first image region A1 including the inspection target 90 that the defective product identification unit 406 did not identify as defective greater than the transparency of the first image region A1 including the inspection target 90 that the defective product identification unit 406 identified as defective. As a result, the inspection target 90 identified as defective is emphasized more than the other inspection target 90. In addition to changing the transparency of the first image Im2, the third image generation unit 411 may also highlight the inspection target 90 identified as defective in any way. For example, the third image generation unit 411 may generate the third image Im4 by superimposing an image such as a frame surrounding the inspection target 90 identified as defective onto the first image Im2.

[0069] [1-3. Operation] Next, the operation of the inspection system 1 according to the first embodiment will be described with reference to Figure 8. Figure 8 is a flowchart of the processing device 4, showing the operation of the inspection system 1 when inspecting the object to be inspected 90. The operation in Figure 8 is started when the processing device 4 detects a visual inspection event. In this embodiment, the visual inspection event is when the equipment control unit 402 uses the imaging device 3 to take a predetermined number of images Im1 including the object to be inspected 90 being transported to the transport device 2. The operation in Figure 8 is repeatedly executed each time the processing device 4 detects a visual inspection event during the inspection process of the object to be inspected 90.

[0070] Each time the processing unit 4 detects an appearance inspection event, in step SA1, the first image generation unit 403 generates a first image Im2 using the image Im1 captured by the imaging device 3. If the image Im1 is to be used as the first image Im2 input to the overall CV model 422, step SA1 can be omitted.

[0071] In step SA2, the first map generation unit 404 inputs the first image Im2 generated in step SA1 into the overall CV model 422 and generates the first feature map M1 by extracting the first feature. Step SA2 corresponds to the "first step".

[0072] In step SA3, the first score calculation unit 405 uses the first feature map M1 to calculate a first score corresponding to each object 90 to be inspected in the first image Im2. Step SA3 corresponds to the "second step".

[0073] In step SA4, the defective product identification unit 406 compares all the first score values ​​obtained in step SA3 with the first reference value 424 stored in memory 420.

[0074] If the defective product identification unit 406 determines that all first score values ​​are less than or equal to the first reference value 424 (step SA4: No), the defective product identification unit 406 identifies all of the inspection targets 90 included in the first image Im2 as good inspection targets 90. After that, the processing device 4 terminates the operation shown in Figure 8 and, when it detects an appearance inspection event again, starts the operation of step SA1 again.

[0075] If the defective product identification unit 406 determines that one or more of the first scores calculated in step SA3 exceed the first reference value 424 (step SA4: Yes), the operation of the processing unit 4 proceeds to step SA5.

[0076] In step SA5, the defective product identification unit 406 identifies the inspected object 90 as defective based on the first score. Specifically, the defective product identification unit 406 identifies the inspected object 90 that corresponds to the first score exceeding the first standard value 424 as defective. After that, the operation of the processing device 4 proceeds to step SA6. Step SA5 corresponds to the "third step".

[0077] The operations from step SA6 onward are performed on all inspection objects 90 that were identified as defective in step SA5. Furthermore, the operations from step SA6 onward may be performed simultaneously on all inspection objects 90 identified as defective in step SA5, or they may be performed sequentially for each inspection object 90.

[0078] In step SA6, the processing unit 4 performs an action on the identified defective product. In this embodiment, the processing unit 4 performs the following actions on the defective product: sorting the defective product and displaying the third image Im4.

[0079] In detail, in step SA6, the equipment control unit 402 of the processing device 4 controls the exclusion mechanism 25 and removes the inspection object 90 that was determined to be defective in step SA5 from the first transport mechanism 21. The equipment control unit 402 then moves the removed inspection object 90 to the second transport mechanism 22. This separates the inspection object 90 that was identified as defective in step SA5 from the inspection object 90 that was not identified as defective.

[0080] Furthermore, in step SA6, the third image generation unit 411 of the processing device 4 generates a third image Im4 based on the first image Im2, highlighting the inspection target 90 that was identified as defective in step SA5. The device control unit 402 then displays the third image Im4 generated by the third image generation unit 411 using the display device 41.

[0081] After step SA6, in step SA7, the second image generation unit 407 extracts a region from the first image Im2 generated in step SA1 that includes the inspection target 90 corresponding to the inspection target 90 identified as defective in step SA5, and generates the second image Im3. After that, the operation of the processing device 4 proceeds to step SA8.

[0082] In step SA8, the second map generation unit 408 inputs the second image Im3 generated in step SA7 into the standalone CV model 423 and generates the second feature map M2 by extracting the second feature. After that, the operation of the processing unit 4 proceeds to step SA9.

[0083] In step SA9, the second score calculation unit 409 uses the second feature map M2 to calculate a second score corresponding to each part 91 to 94 of the object to be inspected 90 included in the second image Im3.

[0084] In step SA10, the defective part identification unit 410 identifies the part corresponding to the second score that indicates the greatest degree of abnormality among the second scores obtained in step SA9 as a defective part. In other words, the defective part identification unit 410 identifies the part corresponding to the second score with the largest value as a defective part.

[0085] In step SA11, the processing device 4 performs an action on the defective area. In this embodiment, the processing device 4 performs marking on the defective area as an action on the defective area.

[0086] In detail, in step SA11, the equipment control unit 402 controls the marking device 27 to mark the defective area identified in step SA10. This makes it easier to identify the defective area of ​​the object to be inspected 90, and allows for smoother investigation of the cause of the defect, for example.

[0087] After step SA11 is completed, the processing unit 4 terminates the operation shown in Figure 8 and, upon detecting an appearance inspection event, restarts the operation of step SA1.

[0088] Next, the effects of the first embodiment will be described. The inspection system 1 of this embodiment generates a first feature map M1 based on a first image Im2 containing multiple objects to be inspected 90, then calculates a first score corresponding to each object to be inspected 90 based on the first feature map M1, and identifies defective products based on the individual scores. The first score is calculated using the value of the first feature in the first map region R1 corresponding to each object to be inspected 90 within the first feature map M1. Therefore, it is possible to identify defective objects to be inspected from a first image containing multiple objects to be inspected, thereby improving the efficiency of inspection.

[0089] Furthermore, the inspection system 1 of this embodiment identifies the defective part of the inspection target 90 that has been identified as a defective product. Therefore, by marking the identified defective part, the investigation of the cause of the defect can be carried out smoothly.

[0090] Furthermore, the inspection system 1 of this embodiment displays a third image Im4 on the display device 41, which highlights the inspection target 90 identified as defective in the first image Im2. In addition, the third image Im4 has increased transparency of the inspection target 90 that was not identified as defective in the first image Im2 compared to the transparency of the inspection target 90 that was identified as defective. As a result, the visual inspection worker P can easily confirm whether the inspection target 90 identified as defective by the inspection system 1 is actually defective.

[0091] [2. Second Embodiment] Next, a second embodiment will be described. In the following, the differences from the first embodiment will be explained, and the description of similar configurations will be omitted.

[0092] In the first embodiment described above, the defective product identification unit 406 determines whether or not the first image Im2 contains a defective product, the object to be inspected 90. In contrast, in the second embodiment, the determination of whether or not the first image Im2 contains a defective product, the judgment of the operator P is included.

[0093] [2-1. Operation] Figure 9 is a flowchart of the processing device 4, showing the operation of the inspection system 1 when inspecting the object 90. The flowchart shown in Figure 9 is the same as the flowchart in Figure 8, with steps SB1 to SB4 added.

[0094] In the operation shown in Figure 9, when the processing unit 4 detects an appearance inspection event, the processing unit 4 executes steps SA1 to SA3, similar to the first embodiment, to generate the first image Im2, generate the first feature map M1, and calculate the first score.

[0095] In the operation shown in Figure 9, after step SA3, in step SB1, the equipment control unit 402 controls the display device 41 to display the first image Im2 generated in step SA1. As a result, the visual inspection worker P can see the first image Im2 displayed on the display device 41.

[0096] After step SB1, in step SB2, the equipment control unit 402 waits for a first input operation by the worker P and determines whether or not the first input operation has been performed. The first input operation is an input operation performed by the worker P to the input device 43 in order to input that the first image Im2 displayed on the display device 41 contains a defective product to be inspected 90. That is, the input device 43 detects the first input operation, which is an input operation indicating that the first image Im2 contains a defective product to be inspected 90. The worker P determines whether or not the first image Im2 displayed on the display device 41 contains a defective product to be inspected 90. If the worker P determines that a defective product is included, the worker P performs the first input operation to the input device 43.

[0097] If the input device 43 does not detect the first input operation and the equipment control unit 402 does not detect a detection signal corresponding to the first input operation (step SB2: NO), the operation of the processing device 4 proceeds to step SA4. In other words, if the operator P determines that the first image Im2 displayed on the display device 41 does not contain the defective inspection target 90, the operation of the processing device 4 proceeds to step SA4.

[0098] The operation from step SA4 onward is the same as the operation shown in Figure 8 in the first embodiment described above. Therefore, if the input device 43 does not detect the first input operation (step SB2: NO), the defective product identification unit 406 identifies the inspected object 90 corresponding to a first score indicating a greater degree of abnormality than the first reference value 424 as a defective product (step SA5). In detail, if the input device 43 does not accept the first input operation, in step SA5, the defective product identification unit 406 identifies the inspected object 90 corresponding to a first score with a value greater than the first reference value 424 as a defective product.

[0099] In step SB2, if the input device 43 detects the first input operation and the equipment control unit 402 detects a detection signal corresponding to the first input operation (step SB2: YES), the operation of the processing device 4 proceeds to step SB3. In other words, if the operator P determines that the first image Im2 displayed on the display device 41 contains a defective product (inspection target 90), the operation of the processing device 4 proceeds to step SB3.

[0100] In step SB3, the first image generation unit 403 stores the generated first image Im2 as the first image Im2 including defective products in an arbitrary database or the like. By using the data stored at this time for training, it becomes possible to improve the accuracy of the overall CV model 422. After that, the process proceeds to step SB4.

[0101] In step SB4, the defective product identification unit 406 identifies a predetermined number of inspection objects 90 included in the first image Im2 as defective products, starting with those with the greatest degree of abnormality, based on the first score calculated in step SA3. That is, the defective product identification unit 406 identifies a predetermined number of inspection objects 90 included in the first image Im2 as defective products, starting with those with the largest first score. The predetermined number can be one or more, and less than or equal to the number of inspection objects 90 included in the first image Im2, for example, one.

[0102] As described above, in step SB4, the defective product identification unit 406 identifies at least one of the inspection objects 90 included in the first image Im2 as a defective product. In other words, if the input device 43 detects the first input operation in step SB2, the defective product identification unit 406 determines that the first image Im2 contains at least one inspection object 90 that is a defective product.

[0103] After step SB4, the operation of the processing unit 4 proceeds to step SA5. The operation from step SA5 onward is the same as the operation shown in Figure 8 in the above embodiment.

[0104] Next, the effects of the second embodiment will be described. In the second embodiment, when the input device 43 detects the first input operation, the defective product identification unit 406 identifies a predetermined number of inspection targets 90 included in the first image Im2 as defective products, in order of decreasing degree of abnormality, based on the first score. Therefore, while defective products can be detected with high reliability by the operator P's visual inspection, the identification of defective products can be performed automatically, and inspection can be carried out efficiently.

[0105] Furthermore, in the second embodiment, if the input device 43 does not detect the first input operation, the defective product identification unit 406 identifies the inspected object 90 corresponding to the first score indicating that the degree of abnormality is greater than the first reference value 424 as a defective product. Therefore, the inspection system 1 can automatically detect and identify defective products even when the operator P misses a defective product and does not perform the first input operation, thereby improving the reliability of the inspection.

[0106] [3. Other Embodiments] The embodiments described above are merely examples and can be modified and applied as needed. Furthermore, the first embodiment, the second embodiment, and the other embodiments described below can be combined as needed to create new embodiments.

[0107] In other embodiments, in addition to the operation of the embodiments described above, in step SA3, the first score calculation unit 405 may store a pair of the calculated first score and the first image region A1 corresponding to the first map region R1 associated with the first score in an arbitrary database or the like. This allows the data stored in the database or the like to be used, for example, to improve the accuracy of each CV model 422, 423, or to calibrate the first reference value 424.

[0108] In the embodiment described above, the actions taken for defective products in step SA6 were sorting of the defective products and displaying the third image Im4. In other embodiments, the action taken for the inspection target 90 identified as a defective part may be configured such that the equipment control unit 402 displays a manual on the display device 41 showing how to deal with the defective product.

[0109] In the embodiment described above, the action for the defective area in step SA11 was to mark the defective area. In other embodiments, as an action for the identified defective area, the device control unit 402 may visualize the second feature map M2 generated in step SA8 as a heatmap and overlay it on the second image Im3 and display it on the display device 41. This makes it easier for the worker P to identify the defective area. In other embodiments, as an action for the identified defective area, the second image Im3 generated in step SA7 may be stored in an arbitrary database or the like.

[0110] In the above-described embodiment, the inspected objects 90 were configured to be classified into two types by the defective product identification unit 406: inspected objects 90 identified as defective and inspected objects 90 not identified as defective. In other embodiments, in addition to the operation of the above-described embodiment, in step SA4, the defective product identification unit 406 may refer to a first score and identify inspected objects 90 with a moderate degree of abnormality. Specifically, the defective product identification unit 406 identifies inspected objects 90 whose first score value is between a first reference value 424 and a second reference value corresponding to a degree of abnormality smaller than the first reference value 424 as inspected objects with a moderate degree of abnormality. This makes it possible, for example, to send the identified inspected objects with a moderate degree of abnormality to another line for double-checking, thus making it easier to suppress the outflow of defective products.

[0111] In the embodiments described above, each CV model 422, 423 was assumed to be a model utilizing a CNN. In other embodiments, each CV model 422, 423 may be a model utilizing an autoencoder, or a model utilizing ViT (Vision Transformer), etc. For example, if each CV model 422, 423 is a model utilizing ViT, each map generation unit 404, 408 may generate spleness maps such as Attention Maps for each CV model 422, 423, into which the first image Im2 and the second image Im3 are input, as each feature map M1, M2.

[0112] In the embodiment described above, the object to be inspected 90 was defined as a transistor. In other embodiments, any object may be used as the object to be inspected 90. For example, the object to be inspected 90 may be an electronic component other than a transistor, such as a capacitor or inductor, or an electronic component packaged on a carrier tape. The object to be inspected 90 may also be a screw, a spring, a tablet, a boxed beverage, etc. Furthermore, the object to be inspected 90 may be solder connecting an electronic component to a circuit board. The inspection system 1 may also detect any defect occurring in the object to be inspected 90 as long as the defect is reflected in the first image Im2. The first image Im2 may also be a transmission image including multiple objects to be inspected 90 taken by an X-ray imaging device, and the application of the inspection system 1 is not limited to visual inspection.

[0113] The processor 400 may consist of a single processor or multiple processors. The processor 400 may also be hardware programmed to implement the corresponding functional units. That is, the processor 400 may consist of, for example, an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0114] Furthermore, the configuration of the processing unit 4 shown in Figure 2 is merely an example, and the specific implementation is not particularly limited. In other words, it is not necessarily required that hardware corresponding to each part be implemented individually; it is certainly possible to configure the system so that a single processor executes a program to realize the functions of each part. Also, some of the functions realized by software in the above-described embodiment may be implemented by hardware, or vice versa. In addition, the specific detailed configuration of each part of the processing unit 4 can be arbitrarily changed.

[0115] Furthermore, the operation step units shown in Figures 8 and 9 are divided according to the main operation content in order to facilitate understanding of the operation when inspecting the object to be inspected 90, and the present invention is not limited by the way the operation units are divided or the names of the operation units. Depending on the operation content, it may be further divided into more step units. Also, it may be divided so that one step unit includes even more operations. In addition, the order of the steps may be changed as appropriate, as long as it does not impede the spirit of the present invention.

[0116] Furthermore, the control program 421 of the above-described embodiment can also be non-temporarily recorded on a recording medium that is readable by the processor 400, for example. A magnetic, optical, or semiconductor memory device can be used as the recording medium. Alternatively, the control program 421 can be stored in a server device or the like, and the processing unit 4 can download the control program 421 from the server device to realize the above-described operation.

[0117] As a variation, a third feature map may be generated by multiplying a first feature map M1, which is generated by extracting a first feature from the first image Im2, with a second feature map M2, which is generated by creating a second image Im3 containing the inspected object 90 identified as defective from the first image Im2, and extracting a second feature from the second image Im3. Specifically, a more reliable and detailed feature map can be obtained by array multiplication of the scores contained in each feature map.

[0118] [4. Configurations Supported by the Above Embodiments] The above embodiments support the following configurations.

[0119] (Configuration 1) An inspection system comprising: a first map generation unit that extracts first features from a first image containing multiple objects to be inspected and generates a first feature map; a first score calculation unit that calculates a first score corresponding to each of the objects to be inspected using the first feature map; and a defective product identification unit that identifies the objects to be inspected that are defective based on the first score. According to the inspection system of Configuration 1, it is possible to identify defective objects to be inspected from a first image containing multiple objects to be inspected. Therefore, the efficiency of inspections involving multiple objects to be inspected can be improved.

[0120] (Configuration 2) The inspection system according to Configuration 1, wherein the first score calculation unit sets a plurality of first map regions corresponding to each of the inspection targets in the first feature map and calculates the first score using the values ​​of the first features in the first map regions. According to the inspection system of Configuration 2, it is possible to identify defective inspection targets from a first image containing multiple inspection targets. Therefore, the efficiency of inspections involving multiple inspection targets can be improved.

[0121] (Configuration 3) An inspection system according to Configuration 1 or 2, comprising: a second image generation unit that extracts a second image from the first image that includes the object to be inspected which has been identified as a defective product by the defective product identification unit; a second map generation unit that extracts a second feature quantity from the second image and generates a second feature quantity map; a second score calculation unit that uses the second feature quantity map to calculate a second score corresponding to a part of the object to be inspected; and a defective part identification unit that identifies the part that is a defective part based on the second score. According to the inspection system of Configuration 3, it is possible to further identify defective parts in an object to be inspected which has been identified as a defective product.

[0122] (Configuration 4) An inspection system according to any one of Configurations 1 to 3, comprising: a third image generation unit that generates a third image that highlights the object to be inspected which has been identified as a defective product by the defective product identification unit based on the first image; and a display unit that displays the third image. According to the inspection system of Configuration 4, the identified defective product can be highlighted and displayed. Therefore, it becomes easier for the worker to confirm the object to be inspected which the inspection system has identified as a defective product.

[0123] (Configuration 5) The inspection system according to Configuration 4, wherein the third image generation unit generates a third image in which the transparency of the inspection target that was not identified as a defective product by the defective product identification unit is greater than the transparency of the inspection target that was identified as a defective product by the defective product identification unit. According to the inspection system of Configuration 5, identified defective products can be highlighted and displayed by the difference in transparency. Therefore, it becomes easier for the operator to confirm the inspection target that the inspection system has identified as a defective product.

[0124] (Configuration 6) An inspection system according to any one of Configurations 1 to 5, comprising an input unit for detecting a first input operation, wherein when the input unit detects the first input operation, the defective product identification unit identifies a predetermined number of the inspection targets included in the first image as defective products, in descending order of the degree of abnormality based on the first score. According to the inspection system of Configuration 6, it is possible to automatically identify defective products while detecting them with high reliability based on the first input operation from an operator or the like. This makes it easier to streamline the inspection process.

[0125] (Configuration 7) The inspection system according to Configuration 6, wherein, if the input unit does not detect the first input operation, the defective product identification unit identifies the object to be inspected that corresponds to the first score indicating that the degree of abnormality is greater than the first reference value among the plurality of objects to be inspected included in the first image as a defective product. According to the inspection system of Configuration 7, even if the first input operation is not performed, the detection and identification of defective products can be performed automatically. Therefore, it is easy to improve the reliability of the inspection.

[0126] (Configuration 8) An inspection method comprising: a first step in which a computer extracts a first feature from a first image containing multiple objects to be inspected and generates a first feature map; a second step in which the computer uses the first feature map to calculate a first score corresponding to each of the objects to be inspected; and a third step in which the computer identifies the objects to be inspected that are defective based on the first score. According to the inspection method of Configuration 8, the same effects as the inspection system of Configuration 1 can be obtained.

[0127] (Configuration 9) A program that causes the processor to function as: a first map generation unit that extracts first features from a first image containing multiple objects to be inspected and generates a first feature map; a first score calculation unit that uses the first feature map to calculate a first score corresponding to each of the objects to be inspected; and a defective product identification unit that identifies the objects to be inspected that are defective based on the first score. According to the program of Configuration 8, the same effect as the inspection system of Configuration 1 can be obtained.

[0128] 1...Inspection system, 2...Transportation device, 3...Photography device, 4...Processing device (computer), 21...First transport mechanism, 22...Second transport mechanism, 23...First collection container, 24...Second collection container, 25...Removal mechanism, 27...Marking device, 40...Control unit, 41...Display device (display unit), 43...Input device (input unit), 45...Communication unit, 90, 90A-90I...Objects to be inspected, 91-93...Lead wires, 94...Main unit, 400...Processor, 401...Communication control unit, 402...Equipment control unit, 403...First image generation unit, 404...First map generation unit, 405...First score calculation unit, 406...Defective product identification unit, 407... 408...Second image generation unit, 409...Second map generation unit, 410...Second score calculation unit, 411...Defect part identification unit, 420...Third image generation unit, 421...Memory, 422...Control program (program), 422...Overall CV model, 423...Individual CV model, 424...First reference value, A1, A1A to A1I...First image area, A2, A2A to A2D...Second image area, Im1...Image, Im2...First image, Im3...Second image, Im4...Third image, M1...First feature map, M2...Second feature map, P...Worker, R1, R1A to R1I...First map area, R2, R2A to R2D...Second map area

Claims

1. An inspection system comprising: a first map generation unit that extracts first features from a first image containing multiple objects to be inspected and generates a first feature map; a first score calculation unit that calculates a first score corresponding to each of the objects to be inspected using the first feature map; and a defective product identification unit that identifies the objects to be inspected as defective based on the first score.

2. The inspection system according to claim 1, wherein the first score calculation unit sets a plurality of first map regions corresponding to each of the objects to be inspected in the first feature map, and calculates the first score using the values ​​of the first features in the first map regions.

3. An inspection system according to claim 1 or 2, comprising: a second image generation unit that extracts a second image from the first image that includes the object to be inspected which has been identified as a defective product by the defective product identification unit; a second map generation unit that extracts a second feature quantity from the second image and generates a second feature quantity map; a second score calculation unit that calculates a second score corresponding to a part of the object to be inspected using the second feature quantity map; and a defective part identification unit that identifies the part that is a defective part based on the second score.

4. An inspection system according to any one of claims 1 to 3, comprising: a third image generation unit that generates a third image that highlights the object to be inspected which has been identified as a defective product by the defective product identification unit based on the first image; and a display unit that displays the third image.

5. The inspection system according to claim 4, wherein the third image generation unit generates a third image in which the transparency of the inspection target that was not identified as a defective product by the defective product identification unit is greater than the transparency of the inspection target that was identified as a defective product by the defective product identification unit.

6. An inspection system according to any one of claims 1 to 5, comprising an input unit for detecting a first input operation, wherein when the input unit detects the first input operation, the defective product identification unit identifies a predetermined number of the inspection targets included in the first image as defective products, in descending order of the degree of abnormality based on the first score.

7. The inspection system according to claim 6, wherein, if the input unit does not detect the first input operation, the defective product identification unit identifies the object to be inspected among the plurality of objects to be inspected included in the first image that corresponds to the first score indicating that the degree of abnormality is greater than the first reference value as a defective product.

8. An inspection method comprising: a first step in which a computer extracts first features from a first image containing multiple objects to be inspected and generates a first feature map; a second step in which the computer uses the first feature map to calculate a first score corresponding to each of the objects to be inspected; and a third step in which the computer identifies the objects to be inspected as defective based on the first score.

9. A program that causes a processor to function as: a first map generation unit that extracts first features from a first image containing multiple objects to be inspected and generates a first feature map; a first score calculation unit that calculates a first score corresponding to each of the objects to be inspected using the first feature map; and a defective product identification unit that identifies the objects to be inspected as defective based on the first score.

Images