Inspection device

Abstract

To provide an inspection device capable of accurately detecting an abnormality of an inspection object without increasing a calculation source.SOLUTION: An inspection device comprises: an acquisition section which acquires a plurality of images obtained by photographing different parts of an inspection object; a feature extraction section which outputs a feature quantity of the images for each part with the use of a learned model where a weight is set by machine learning; an inference section which includes a plurality of inference models each having a parameter set for each part, and, with input of the feature quantity to the inference model corresponding to the part corresponding to the feature quantity, outputs an index value indicating a probability that there is an abnormality for each part; a part determination section which includes a plurality of determination models each having a threshold set for each part, and, with input of the index value to the determination model corresponding to the part corresponding to the index value, outputs a part determination result indicating presence / absence of an abnormality for each part; and an integrated determination section which determines presence / absence of the abnormality of the inspection object based on a plurality of part determination results.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] An embodiment of the present invention relates to an inspection apparatus. [Background technology]

[0002] Non-destructive testing is used as a technology for detecting abnormalities (defects, defects, etc.) in inspected objects such as industrial products, and is capable of detecting abnormalities without destroying the inspected object by analyzing images captured of the inside or surface of the inspected object. In some cases, a trained model generated by machine learning (deep learning) is used for the image analysis process in such non-destructive testing. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-24201 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the above-mentioned non-destructive inspection, a plurality of images (e.g., a plurality of cross-sectional images of the internal structure of the object to be inspected divided into a plurality of layers) are obtained by imaging a plurality of parts of one object to be inspected, and image analysis processing is performed on each of these images. In such a case, it is necessary to perform different image analysis processing for each part (e.g., layer, etc.), but if a different trained model is used for each part, a large amount of computational resources (memory capacity, processing power, etc.) is required.

[0005] Therefore, one of the problems to be solved by the embodiments of the present invention is to provide an inspection device capable of detecting an abnormality in an inspection object with high accuracy without increasing the computational resources. [Means for solving the problem]

[0006] The inspection apparatus of the embodiment includes an acquisition unit that acquires multiple images capturing different portions of an object to be inspected; a feature extraction unit that outputs image features for each portion using a learned model in which weights are set by machine learning; an inference unit that includes multiple inference models each having parameters set for each portion, and outputs an index value indicating the likelihood that an abnormality exists for each portion by inputting the feature values ​​output from the feature extraction unit into the inference model corresponding to the portion corresponding to the feature values; a partial judgment unit that includes multiple judgment models each having a threshold value set for each portion, and outputs a partial judgment result indicating the presence or absence of an abnormality for each portion by inputting the index values ​​output from the inference unit into the judgment models corresponding to the portions corresponding to the index values; and an integrated judgment unit that judges the presence or absence of an abnormality in the object to be inspected based on the multiple partial judgment results output from the partial judgment unit.

[0007] According to the above configuration, an abnormality in an object to be inspected is detected using one trained model, multiple inference models having parameters set for each part, and multiple judgment models having thresholds set for each part. Then, by setting (optimizing) each parameter of the multiple inference models and each threshold of the multiple judgment models to suit the corresponding parts, it becomes possible to appropriately analyze multiple images according to different conditions for each part without using multiple trained models that require large computational resources. This makes it possible to perform high-precision nondestructive inspection without increasing computational resources. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration of an inspection device according to an embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of an inspection object and an image according to the embodiment. [Diagram 3] FIG. 3 is a diagram illustrating an example of a relationship between layers and images according to the embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a functional configuration of the inspection apparatus according to the embodiment. [Diagram 5]FIG. 5 is a flowchart illustrating an example of an inspection process performed by the inspection device of the embodiment. [Figure 6] FIG. 6 is a flowchart illustrating an example of the learning process according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Exemplary embodiments of the present invention are disclosed below. The configurations of the embodiments described below, and the actions, results, and effects brought about by the configurations are merely examples. The present invention can be realized by configurations other than those disclosed in the following embodiments, and it is possible to obtain at least one of various effects and derivative effects based on the basic configurations.

[0010] 1 is a diagram showing an example of the configuration of an inspection device 1 according to an embodiment. The inspection device 1 according to this embodiment is an apparatus capable of non-destructive inspection, which detects abnormalities in a predetermined inspection object O without destroying the inspection object O. The inspection object O should not be particularly limited, and may be, for example, a semiconductor device, an in-vehicle part, or any other type of industrial product.

[0011] The inspection device 1 of this embodiment includes an imaging device 11 and an information processing device 12.

[0012] The imaging device 11 is a device that acquires an image (image data) of a predetermined portion of the inspection object O. The imaging device 11 may be, for example, a CT (Computed Tomography) device capable of acquiring a cross-sectional image showing the internal structure of the inspection object O by using X-rays or the like. Note that the imaging device 11 is not limited to a CT device, and may be, for example, a visible light camera, an infrared camera, an ultrasonic camera, or the like capable of acquiring an appearance image showing the surface state of the inspection object O.

[0013] The information processing device 12 is a device that executes a process for detecting an abnormality in the inspection object O based on an image acquired by the imaging device 11. The information processing device 12 may be, for example, a general-purpose computer including a CPU (Central Processing Unit) 21, a memory 22, a communication I / F (Interface) 23, a user I / F 24, and the like. The CPU 21 executes various processes for detecting an abnormality in the inspection object O according to a program stored in the memory 22. In addition to the program, the memory 22 stores various data required for executing the processes. The communication I / F 23 is a device that establishes communication with an external device such as the imaging device 11 via a predetermined network. The user I / F 24 is a device that enables input and output of information between the user and the information processing device 12, and may be, for example, a display, a speaker, a keyboard, a pointing device, a touch panel mechanism, a microphone, and the like.

[0014] FIG. 2 is a diagram showing an example of an inspection object O and images I1 to In of an embodiment. In FIG. 2, a top view of the inspection object O is illustrated. The inspection object O illustrated here is a BGA (Ball Grid Array) having a plurality of bumps B that function as electrodes of electronic circuit components such as a processor. The BGA illustrated here is configured such that a plurality of bumps B are arranged at predetermined intervals on the surface of a substrate S. Each bump B is made of a conductive material, is formed so as to protrude from the surface of the substrate S, and is a member that is approximately elliptical or spherical in side view.

[0015] The imaging device 11 of this embodiment obtains a plurality of images I1-In showing the internal structure of the bump B using X-rays or the like. These images I1-In are cross-sectional images corresponding to a plurality of layers (the first layer to the n-th layer in this example) dividing the bump B in the height direction (the direction perpendicular to the surface of the substrate S). The information processing device 12 of this embodiment performs a predetermined image analysis process on each of the plurality of images I1-In, and detects abnormalities in all the bumps B or one or more selected bumps B based on the results. Note that the plurality of images I1-In are not limited to X-ray images, and may be, for example, images of a defect photographed from a plurality of angles. By photographing from a plurality of angles, there is an effect that defects that cannot be seen from a certain angle can be detected.

[0016] Fig. 3 is a diagram showing an example of the relationship between layers and images in the embodiment. Fig. 3 illustrates a state in which a void 55 exists inside a bump B having a substantially elliptical spherical shape. For example, a bump B containing such voids 55 at a relatively high rate can be determined to be abnormal. Note that the void 55 is an example of an abnormality factor that makes the inspection object O abnormal, and the abnormality factor is not limited to the void 55.

[0017] The bump B exemplified here is substantially ellipsoidal, and therefore its width (diameter of the cross section parallel to the substrate S) differs for each layer. Specifically, the width of the bump B exemplified here is largest at the (n / 2)th layer, which is the central layer, and becomes smaller toward the end layers, i.e., the first layer or the nth layer.

[0018] 3, an image Ia in the a-th layer corresponding to a position where a certain void 55 exists, and an image Ib in the b-th layer corresponding to a position where another void 55 exists are illustrated. Here, the a-th layer is closer to the (n / 2)-th layer than the b-th layer. In such a case, the width D1 of the bump B in the a-th layer is larger than the width D2 of the bump B in the b-th layer. In other words, the ratio D1 / D of the width D1 of the bump B to the width D of the imaging region RI in the image Ia in the a-th layer is larger than the ratio D2 / D of the width D2 of the bump B to the width D of the imaging region RI in the image Ib in the b-th layer.

[0019] As described above, the relationship between the imaging region RI, the image of the inspection object O (bump B), and the image of the abnormality cause (void 55, etc.) differs for each layer depending on the shape characteristics, etc. of the inspection object O. Therefore, in order to detect the abnormality of the bump B with high accuracy, it is necessary to adjust the parameters, thresholds, etc. in the image analysis process for the images I1 to In for each layer.

[0020] Fig. 4 is a diagram showing an example of a functional configuration of the inspection device 1 of the embodiment. The inspection device 1 of the embodiment includes an acquisition unit 101, a feature extraction unit 102, an inference unit 103, a partial judgment unit 104, an integrated judgment unit 105, an output unit 106, and a learning processing unit 107. These functional units can be realized by cooperation between hardware and software as exemplified in Fig. 1. Furthermore, at least one of these functional units may be realized by dedicated hardware (circuitry, etc.).

[0021] The acquiring unit 101 acquires a plurality of images I1-In obtained by capturing a plurality of portions (e.g., layers) of an inspection object O (e.g., a bump B). Here, a case will be described in which the plurality of images I1-In are a plurality of cross-sectional images showing the internal structure of the inspection object O for each layer, but the plurality of images is not limited to this. For example, the plurality of images may be exterior images captured from a plurality of angles of the surface of the inspection object.

[0022] The feature extraction unit 102 outputs feature quantities F1 to Fn of images I1 to In for each portion (layer) using a trained model in which weights W are set by machine learning (deep learning). The specific structure of the trained model is not limited, but may be, for example, a structure using a neural network. Each of the feature quantities F1 to Fn can take one or more values. For example, the feature quantities F1 to Fn may be a scalar value or a vector.

[0023] The values ​​of the feature quantities F1 to Fn should be appropriately determined according to the type and characteristics of the inspection object O. When the inspection object O is a bump B, the feature quantities F1 to Fn may be values ​​that increase as the proportion of the area of ​​the image of the void 55 in the images I1 to In increases.

[0024] The inference unit 103 includes a plurality of inference models (1st to nth inference models) provided for each part (layer), and outputs abnormality degrees S1 to Sn as index values ​​indicating the likelihood that an abnormality exists for each part (layer) by inputting the plurality of feature amounts F1 to Fn output from the feature extraction unit 102 into the corresponding inference models. Each of the plurality of inference models (1st to nth inference models) has unique parameters α1 to αn set for each part (layer).

[0025] The abnormality levels S1 to Sn, which are an example of index values, may be values ​​that increase as the possibility of an abnormality increases, or may be values ​​that increase as the distance between the feature quantities F1 to Fn and the distribution set by the specific parameters α1 to αn increases. Each of the multiple inference models has specific parameters α1 to αn that differ for each layer. These parameters α1 to αn are set to suit the state of the inspection object O that differs for each layer (for example, the width of the bump B differs for each layer). The specific configuration of the inference model is appropriately determined according to the type, characteristics, etc. of the inspection object O. The specific parameters α1 to αn may have the same value. Furthermore, each of the specific parameters α1 to αn can take one or more values, such as a scalar value or a vector.

[0026] By using a method of measuring normalized distance to calculate the anomaly levels S1 to Sn, the anomaly levels S1 to Sn can take values ​​according to a certain probability density distribution. Therefore, since a judgment can be made by a threshold value obtained from the characteristics of the distribution, the threshold value can be set automatically, and the setting man-hours can be reduced. For example, by using the Mahalanobis distance to calculate the anomaly levels S1 to Sn as the normalized distance, the anomaly levels S1 to Sn follow a chi-square distribution. Therefore, by setting a threshold value with a random variable value whose cumulative probability of the chi-square distribution and the probability of anomaly occurrence match, a threshold value that can satisfy the desired accuracy of anomaly judgment can be derived only by calculation, and the threshold value can be set automatically. Moreover, cosine similarity can be used instead of the normalized distance.

[0027] The partial judgment unit 104 includes a plurality of judgment models (1st to nth judgment models) provided for each part (layer), and outputs partial judgment results R1 to Rn indicating the presence or absence of anomaly for each part (layer) by inputting the plurality of abnormality degrees S1 to Sn output from the inference unit 103 into the corresponding judgment models. Each of the plurality of judgment models (1st to nth judgment models) has a unique threshold value β1 to βn set for each part (layer).

[0028] Each judgment model may be, for example, a logic model that judges that there is an abnormality when the input abnormality degrees S1 to Sn are greater than the corresponding thresholds β1 to βn. Each of the judgment models has a unique threshold β1 to βn that differs for each hierarchical level. These thresholds β1 to βn are set to suit the state of the inspection object O, which differs for each hierarchical level, similar to the parameters α1 to αn of the inference model. It is also possible to set a common value for the thresholds β1 to βn depending on the state and degree of suitability of the inspection object O, which differs for each hierarchical level. In this case, the thresholds β1 to βn can be set collectively for the hierarchical levels to which a common value can be set, which has the effect of reducing the number of steps for setting the thresholds β1 to βn.

[0029] The integrated judgment unit 105 judges the presence or absence of an abnormality in the inspection object O based on the multiple partial judgment results R1 to Rn output from the partial judgment unit 104, and outputs an integrated judgment result RT indicating the result. Note that what is to be the inspection object O can be appropriately set depending on the usage situation, etc. For example, each of the multiple bumps B may be the inspection object O, or a BGA including multiple bumps B may be the inspection object O.

[0030] The output unit 106 outputs the integrated judgment result RT output from the integrated judgment unit 105 in a predetermined format.

[0031] The learning processing unit 107 executes a learning process for setting weights of the trained model, a learning process for setting parameters α1 to αn of each of the multiple inference models, and a learning process for setting thresholds β1 to βn of each of the multiple judgment models. The learning processing unit 107 calculates parameters of an inference model corresponding to a first part, which is one of the multiple parts (layers), based on an output obtained by inputting a predetermined data set into the trained model in which the weight W is set, and calculates a threshold of a judgment model corresponding to the first part based on an output obtained by inputting a predetermined data set into the inference model having the calculated parameters. For example, the learning processing unit 107 calculates a parameter α1 of a first inference model corresponding to the first layer based on an output obtained by inputting a predetermined data set into the trained model in which the weight W is set. Then, the learning processing unit 107 calculates a threshold β1 of a first judgment model corresponding to the first layer based on an output obtained by inputting a predetermined data set into the first inference model having the parameter α1.

[0032] According to the above configuration, image analysis processing can be performed on a plurality of images I1 to In using one trained model (weight W) so as to adapt to different situations for each layer, thereby reducing computational resources such as memory.

[0033] 5 is a flowchart showing an example of an inspection process by the inspection device 1 of the embodiment. When the acquisition unit 101 acquires images I1 to In of each layer of an inspection object O (for example, a bump B), a variable d indicating a layer is set to 1 (S101).

[0034] The feature extraction unit 102 inputs an image Id in the dth layer (image I1 in the first layer at the initial stage) into a trained model in which a weight W has been set by prior machine learning, and obtains a feature Fd (feature F1 at the initial stage) indicating the characteristics of the image Id (S102).

[0035] Then, the inference unit 103 inputs the feature Fd acquired by the feature extraction unit 102 into the d-th inference model (the first inference model having the parameter α1 corresponding to the first hierarchical level in the initial stage) and acquires the anomaly degree Sd of the d-th hierarchical level (the anomaly degree S1 of the first hierarchical level in the initial stage) (S103).

[0036] Thereafter, the partial judgment unit 104 inputs the abnormality degree Sd acquired by the inference unit 103 into the d-th judgment model (the first judgment model having the threshold value β1 corresponding to the first hierarchical level in the initial stage) and acquires a partial judgment result Rd indicating the presence or absence of an abnormality in the d-th hierarchical level (the partial judgment result R1 of the first hierarchical level in the initial stage) (S104).

[0037] Then, it is determined whether d=n, i.e., whether the variable d has reached the maximum number of hierarchical layers n (S105), and if d=n is not true (S105: No), 1 is added to the current variable d (S106) and the processing from step S102 onwards is executed again. That is, after the partial judgment result R1 of the first layer is obtained, the processing of steps S102 to S104 is executed to obtain the partial judgment result R2 of the second layer, and then the processing of steps S102 to S104 is repeated until the partial judgment result Rn of the nth layer is obtained.

[0038] If d=n (S105: Yes), that is, after partial judgment results R1 to Rn for all layers 1 to n are acquired, the integrated judgment unit 105 judges the presence or absence of an abnormality in the inspection object O based on these partial judgment results R1 to Rn (S107). The specific method of the judgment should be appropriately determined according to the type and characteristics of the inspection object O, but for example, if even one of the multiple partial judgment results R1 to Rn is abnormal, it may be judged that the inspection object O is abnormal. Then, the output unit 106 outputs an integrated judgment result RT indicating the judgment result by the integrated judgment unit 105 in a predetermined format (S108).

[0039] The learning process for setting the weight W of the trained model, the parameters α1 to αn of the multiple inference models, and the thresholds β1 to βn of the multiple determination models will be described below.

[0040] 6 is a flowchart showing an example of the learning process of the embodiment. First, the learning processing unit 107 determines whether or not to use a weight calculated using a data set outside the domain as the weight W of the learned model (S201). The data set outside the domain is a data set including an image of an object other than the actual inspection object O. By using such a data set outside the domain, it is not necessary to perform machine learning for calculating the weight for each inspection object O (bump B), and therefore the number of steps required for the learning process can be reduced.

[0041] When the weights calculated using a dataset outside the domain are not to be used (S201: No), the learning processing unit 107 calculates weights using a predetermined dataset (for example, a dataset consisting only of an image of the actual inspection object O) different from the dataset outside the domain (S202), and then sets a variable d indicating the hierarchy to 1 (S203). When the weights calculated using a dataset outside the domain are to be used (S201: Yes), the variable d is set to 1 (S203) without performing step S202.

[0042] Then, the learning processing unit 107 calculates a parameter αd of the dth inference model (a parameter α1 of the first inference model in the initial stage) based on the output obtained by inputting a data set for parameter setting into the learned model in which the weight W calculated as described above is set (S204).

[0043] Then, the learning processing unit 107 calculates a threshold value βd of the dth judgment model (threshold value β1 of the first judgment model in the initial stage) based on the output obtained by inputting a data set for setting a threshold value into the dth inference model having the parameter αd calculated as described above (S205).

[0044] Then, it is determined whether d=n, i.e., whether the variable d has reached the maximum number of hierarchical layers n (S206), and if d=n is not true (S206: No), 1 is added to the current variable d (S207), and the processing from step S204 onwards is executed again. That is, after the setting of the parameter α1 and threshold β1 corresponding to the first layer is completed, the processing of steps S204 and S205 is executed to set the parameter α2 and threshold β2 corresponding to the second layer, and then the processing of steps S204 and S205 is repeated until the setting of the parameter αn and threshold βn for the nth layer is completed.

[0045] If d=n (S206: Yes), that is, when the parameters α1 to αn and thresholds β1 to βn for all layers 1 to n have been set, this routine ends. By this learning process, the parameters α1 to αn and thresholds β1 to βn corresponding to each of the layers can be simply and appropriately set.

[0046] As described above, according to this embodiment, feature quantities F1 to Fn are extracted from each of the multiple images I1 to In using one trained model. Then, the feature quantities F1 to Fn are input to the corresponding inference models to obtain the abnormality degrees S1 to Sn for each layer, and the abnormality degrees S1 to Sn are input to the corresponding judgment models to obtain the partial judgment results R1 to Rn for each layer, and the presence or absence of an abnormality in the inspection object O is judged based on the partial judgment results R1 to Rn. Then, the parameters α1 to αn of the multiple inference models and the threshold values ​​β1 to βn of the multiple judgment models are set (optimized) to suit the corresponding layers. That is, image analysis processing can be performed on the multiple images I1 to In using one trained model to suit different situations for each layer. This makes it possible to perform high-precision nondestructive inspection without increasing computational resources.

[0047] In the above embodiment, the portions of the inspection object O are layers dividing the inside of the inspection object O, and the images are X-ray images showing the internal structure of the inspection object O in each layer. With this configuration, non-destructive inspection can be performed with high accuracy using small computational resources by using a plurality of X-ray images captured by dividing the internal structure of the inspection object O into a plurality of layers.

[0048] In the above embodiment, the inspection device 1 includes a learning processing unit 107 that calculates parameters of an inference model corresponding to a first part, which is one of the parts, based on an output obtained by inputting a predetermined data set into the trained model, and calculates thresholds of a judgment model corresponding to the first part based on an output obtained by inputting a predetermined data set into the inference model having the calculated parameters. With this configuration, the parameters of the multiple inference models and the thresholds of the multiple judgment models can be simply and appropriately set (optimized) to fit the corresponding parts.

[0049] Furthermore, in the above embodiment, the learning processing unit 107 calculates the weights of the trained model using a data set including images of objects other than the inspection target O. This configuration eliminates the need to perform machine learning for calculating the weights for each inspection target, thereby reducing the number of steps required for the learning process.

[0050] In addition, in the above configuration, the inference unit 103 outputs an index value for an input of a feature amount, using a normalized distance for each feature. In this way, by using a method of measuring a normalized distance as a method of calculating an index value, the index value can take a value according to a certain probability density distribution. This makes it possible to make a judgment based on a threshold value obtained from the characteristics of the distribution, so that the setting of the threshold value can be automated, resulting in a reduction in the number of steps required for setting.

[0051] The program for causing the inspection device 1 to execute the above-mentioned functions may be stored in an installable or executable file format on a computer-readable storage medium such as a CD-ROM, CD-R, memory card, DVD (Digital Versatile Disk), or flexible disk (FD) and provided as a computer program product. The program may also be stored on a computer connected to a network such as the Internet and provided by downloading the program via the network. The program may also be provided or distributed via a network such as the Internet.

[0052] Although the embodiment of the present invention has been described above, the above embodiment is presented as an example and is not intended to limit the scope of the invention. This new embodiment can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. This embodiment and its modifications are included in the scope and gist of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0053] 1... inspection device, 11... imaging device, 12... information processing device, 55... void, 101... acquisition unit, 102... feature extraction unit, 103... inference unit, 104... partial judgment unit, 105... integrated judgment unit, 106... output unit, 107... learning processing unit, B... bump, F1 to Fn... feature amount, I1 to In, Ia, Ib... image, O... inspection object, R1 to Rn... partial judgment result, RI... imaging area, RT... integrated judgment result, S... substrate, S1 to Sn... abnormality level, W... weight, α1 to αn... parameter, β1 to βn... threshold value

Claims

1. an acquisition unit that acquires a plurality of images of different parts of an inspection object; a feature extraction unit that outputs a feature amount of the image for each of the portions using a trained model in which weights are set by machine learning; an inference unit including a plurality of inference models each having parameters set for each of the portions, the inference unit inputting the feature amount output from the feature extraction unit to the inference model corresponding to the portion corresponding to the feature amount, and outputting an index value indicating the likelihood that an abnormality exists for each of the portions; a partial judgment unit including a plurality of judgment models each having a threshold value set for each of the portions, the partial judgment unit inputting the index value output from the inference unit to the judgment model corresponding to the portion corresponding to the index value, thereby outputting a partial judgment result indicating the presence or absence of an abnormality for each of the portions; an integrated judgment unit that judges the presence or absence of an abnormality in the inspection object based on the plurality of partial judgment results output from the partial judgment unit; An inspection device comprising:

2. the portions are layers that divide the inside of the inspection object, the images are X-ray images showing the internal structure of the object to be inspected in each of the layers; 2. The inspection device according to claim 1.

3. a learning processing unit that calculates the parameters of the inference model corresponding to a first portion that is one of the multiple portions, based on an output obtained by inputting a predetermined data set into the trained model, and calculates the threshold value of the judgment model corresponding to the first portion based on an output obtained by inputting a predetermined data set into the inference model having the calculated parameters; The inspection apparatus of claim 1 further comprising:

4. The learning processing unit calculates weights of the trained model using a dataset including images of objects other than the inspection target.

4. The inspection apparatus according to claim 3.

5. the inference unit uses a normalized distance for each feature to output the index value in response to an input of the feature amount.

2. The inspection device according to claim 1.

Images