Inspection System

Abstract

To provide an inspection system capable of improving inspecting accuracy of an abnormal location even after a trained model is created.SOLUTION: An inspection system 100 comprises: a camera 300 to image an inspected object 200; a sampling generating part, wherein the sampling generating part causes an image captured with a camera 300 to be input to a previously generated, trained model, regards an internal variable of the trained model as a probability variable, classifies each pixel of the image, and outputs multiple sets of sampling data; and a data derivation part to derive one set of statistical data by statistically processing multiple sets of sampling data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an inspection system. [Background technology]

[0002] Conventionally, visual inspection of products has been performed using cameras. In visual inspection, images captured by the camera are input into a pre-trained AI model (hereinafter referred to as the trained model), and abnormalities in the product can be detected as output.

[0003] In the field of product appearance inspection, Patent Document 1 discloses a technology for automatically creating a trained model appropriate for a specific product by using images of the specific product captured from multiple angles. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-112483 Summary of the Invention [Problem to be solved by the invention]

[0005] However, even if a trained model appropriate for a specific product is created, if an abnormality different from that of the specific product is found during visual inspection of another product of the same type as the specific product, the trained model may not be able to detect the abnormality.

[0006] In consideration of the above-mentioned problems, the present disclosure aims to provide an inspection system that can improve the accuracy of detecting abnormalities even after a trained model has been created. [Means for solving the problem]

[0007] In order to solve the above problem, an inspection system according to one aspect of the present disclosure includes a camera that captures an image of an object to be inspected, and inputting the image captured by the camera into a trained model that has been created in advance; A single sampling data in which the output value output from the trained model is linked to each pixel of the image, The internal variables of the trained model are used as random variables. multiple a sampling generation unit that outputs the sampling data; and a data derivation unit that performs statistical processing on the plurality of sampling data to derive one piece of statistical data. a scanning determination unit that determines whether the camera is scanning; and a model switching unit that, when it is determined by the scanning determination unit that the camera is scanning, switches the trained model to a first model in which the internal variables are fixed values, and, when it is determined by the scanning determination unit that the camera is not scanning, switches the trained model to a second model in which the internal variables are random variables; Equipped with. In order to solve the above problem, an inspection system according to one aspect of the present disclosure includes a camera that captures an image of an object to be inspected, and inputting the image captured by the camera into a trained model that has been created in advance; A single sampling data in which the output value output from the trained model is linked to each pixel of the image, The internal variables of the trained model are used as random variables. multiple a sampling generation unit that outputs the sampling data; and a data derivation unit that performs statistical processing on the plurality of sampling data to derive one piece of statistical data. a transmission state storage unit that stores sampling data in association with the transmission state of neurons in the trained model when the sampling data is obtained; a performance evaluation unit that derives a detection rate and an overdetection rate of abnormal portions of the inspection object based on the difference between the sampling data and the teacher data, and evaluates a performance value of the sampling data based on the detection rate and the overdetection rate; and an on / off operation unit that performs on / off operation of neuron transmission in the trained model based on the performance value and using the transmission state associated with the performance value. Equipped with.

[0008] The data derivation unit may derive a cumulative value by multiplying each sampling data by a predetermined value and accumulating the multiplied values ​​within a predetermined upper limit value.

[0009] The data derivation unit may derive an average value by dividing an integrated value obtained by integrating a plurality of pieces of sampled data for each pixel by the total number of samples.

[0011] The model switching unit may execute the second model by stochastically turning on and off transmission of neurons in the trained model. [Effects of the Invention]

[0013] According to the present disclosure, it is possible to improve the accuracy of detecting abnormalities even after a trained model has been created. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a schematic diagram of the inspection system according to this embodiment. [Figure 2] FIG. 2 is a block diagram showing the functions of the image processing device. [Figure 3] FIG. 3 is a diagram illustrating the operation of the first model, which is a trained model. [Figure 4] FIG. 4 is a diagram for explaining the operation of the second model, which is a trained model. [Figure 5] FIG. 5 is a diagram for explaining the transmission state of each neuron in the first model. [Figure 6] FIG. 6 is a diagram for explaining the transmission state of each neuron in the second model. DETAILED DESCRIPTION OF THE INVENTION

[0015] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Specific dimensions, materials, numerical values, etc. shown in the embodiments are merely examples for ease of understanding and do not limit the present disclosure unless otherwise specified. In this specification and drawings, elements having substantially the same functions and configurations are designated by the same reference numerals to avoid redundant explanation, and elements not directly related to the present disclosure are not shown.

[0016] 1 is a schematic configuration diagram of an inspection system 100 according to this embodiment. As shown in FIG. 1, the inspection system 100 includes an inspection object 200, a camera 300, an image processing device 400, and a display 500.

[0017] The inspection object 200 is a product whose appearance is inspected by the camera 300. The imaging range 302 of the camera 300 is large enough to capture an image of at least a portion of the inspection object 200. If the inspection object 200 is larger than the imaging range 302 of the camera 300, the appearance of the entire inspection object 200 is inspected by scanning the camera 300. In this embodiment, the inspection object 200 is, for example, a general industrial product. However, the inspection object 200 is not limited to an industrial product as long as it can be imaged by the camera 300.

[0018] The camera 300 captures an image of the inspection object 200 and outputs the captured image to the image processing device 400. In this embodiment, the camera 300 is manually operated by an inspector IS, and the entire inspection object 200 is scanned.

[0019] The camera 300 is moved by the inspector IS in a direction parallel to the surface of the inspection object 200. The inspector IS visually checks for an abnormality 202 in the inspection object 200, and stops scanning with the camera 300 when the abnormality 202 is included in the imaging range 302 of the camera 300. Note that the inspector IS continues scanning with the camera 300 while the abnormality 202 is not included in the imaging range 302 of the camera 300. However, without being limited to this, the camera 300 may be automatically moved in a direction parallel to the surface of the inspection object 200 by a moving device (not shown), and the entire inspection object 200 may be scanned.

[0020] The image processing device 400 acquires the captured image output from the camera 300. The image processing device 400 also performs image processing on the captured image and outputs the captured image and the image processing results to the display 500. However, the image processing device 400 may also acquire the captured image output from the camera 300 via the display 500.

[0021] The image processing device 400 includes one or more processors 400a and one or more memories 400b connected to the processors 400a. The memories 400b include a ROM in which programs and the like are stored, and a RAM as a work area.

[0022] 2 is a block diagram showing the functions of the image processing device 400. A processor 400a of the image processing device 400, in cooperation with a memory 400b, functions as a scanning determination unit 402, a model switching unit 404, a sampling generation unit 406, a data derivation unit 408, a transmission state storage unit 410, a performance evaluation unit 412, and an on / off operation unit 414.

[0023] The scanning determination unit 402 identifies the position of the camera 300 based on the output of a position detection device (not shown) that detects the position of the camera 300, and determines whether the camera 300 is being scanned.

[0024] The model switching unit 404 switches the type of trained model based on the determination result of the scanning determination unit 402. The trained model is, for example, a deep learning model in image analysis. In this embodiment, the trained model is configured, for example, by a combination of a neural network structure and parameters that represent the strength of the connection between each neuron. Each connection between neurons is provided with a parameter that is a coefficient. Each parameter is configured to be adjustable. The internal state of the trained model is represented by a set of numerical values ​​that are a combination of the neural network structure, called internal variables, and the parameters between each neuron. In this embodiment, the model switching unit 404 switches the trained model between a first model and a second model based on the determination result of the scanning determination unit 402.

[0025] Specifically, when the scanning determination unit 402 determines that the camera 300 is scanning, the model switching unit 404 switches the trained model to a first model in which the internal variables are fixed values. Also, when the scanning determination unit 402 determines that the camera 300 is not scanning, the model switching unit 404 switches the trained model to a second model in which the internal variables are random variables. Details of the first model and the second model will be described later.

[0026] A trained model is created by learning a large amount of training data to which correct answers are previously linked. For example, in the training data, in an image of an inspection object 200 having an abnormality 202 on its surface, 1 is linked as an output node to pixels corresponding to the abnormality 202, and 0 is linked as an output node to pixels corresponding to normal areas. The training data is input to the trained model, and values ​​(fixed values) of the internal variables of the trained model are set so that the error between the output of the trained model and the linked correct answer is small.

[0027] 3 is a diagram for explaining the operation of the first model, which is a trained model. As shown in FIG. 3, the sampling generation unit 406 inputs a captured image 600 to the first model, which is a trained model. The sampling generation unit 406 also sets the internal variables of the trained model to fixed values, classifies each pixel of the image using the values ​​of the output nodes output from the trained model, and outputs one piece of sampling data 700.

[0028] In the sampling data 700, pixels corresponding to abnormal areas on the surface of the inspection object 200 (black areas in the sampling data 700 in FIG. 3) are associated with the numerical value "1" as an output node. In the sampling data 700, pixels corresponding to normal areas on the surface of the inspection object 200 (white areas in the sampling data 700 in FIG. 3) are associated with the numerical value "0" as an output node. In other words, the captured image input to the trained model is output in the output format of semantic segmentation.

[0029] 4 is a diagram illustrating the operation of the second model, which is a trained model. As shown in FIG. 4, the sampling generation unit 406 inputs a captured image 600 to the second model, which is a trained model. The sampling generation unit 406 also treats the internal variables of the trained model as random variables, classifies each pixel of the image using the values ​​of the output nodes output from the trained model, and outputs multiple sampling data 800, 802, and 804. In this embodiment, for simplicity of explanation, the number of multiple sampling data 800, 802, and 804 is set to three, but the number of sampling data is not limited to this as long as it is multiple.

[0030] Furthermore, the data derivation unit 408 performs statistical processing on the plurality of sampling data 800, 802, and 804 to derive one piece of statistical data 900. In this embodiment, the data derivation unit 408 multiplies each of the sampling data 800, 802, and 804 by a predetermined value, and accumulates the multiplied values ​​within a predetermined upper limit value to derive a cumulative value.

[0031] The predetermined value is, for example, 1 / x, where x is an arbitrary value set by the inspector IS and is equal to or less than the total number of the plurality of sampling data 800, 802, and 804. The predetermined upper limit is, for example, 1.

[0032] Specifically, suppose that the value of "2" is set for x in the predetermined value 1 / x, and an output node value of "1" is associated with a predetermined pixel corresponding to each of the sampling data 800, 802, and 804. In this case, the output node value "1" of the predetermined pixel in each of the sampling data 800, 802, and 804 is first multiplied by 1 / 2.

[0033] This multiplication value 1 / 2 is then accumulated for the total number of pieces of sampling data 800, 802, and 804. Here, the total number of pieces of sampling data 800, 802, and 804 is 3, and the multiplication value "1 / 2" is associated with a predetermined pixel of all the sampling data 800, 802, and 804.

[0034] In this case, the cumulative value at the specified pixel becomes 3 / 2, which exceeds the predetermined upper limit value of 1. In this case, the portion of the cumulative value 3 / 2 that exceeds the predetermined upper limit value of 1 is rounded down, and the cumulative value becomes the predetermined upper limit value of 1.

[0035] On the other hand, for a predetermined pixel corresponding to each of the sampling data 800, 802, and 804, assume that only one of the sampling data is associated with an output node value of "1" and the remaining two sampling data are associated with an output node value of "0." In this case, the output node value "1" of the predetermined pixel in one sampling data is multiplied by 1 / 2, and the output node value "0" of the predetermined pixel in the remaining two sampling data is multiplied by 1 / 2.

[0036] Then, at a predetermined pixel corresponding to each of the sampling data 800, 802, and 804, the multiplication values ​​1 / 2 and 0 are accumulated for the total number of sampling data 800, 802, and 804. As a result, the accumulated value at the predetermined pixel becomes 1 / 2, which does not exceed the predetermined upper limit value "1", and therefore the accumulated value remains the same value "1 / 2".

[0037] The pixel value or brightness value of the image displayed on the display 500 is changed according to this cumulative value. For example, if the image is represented by pixel values ​​of 256 gradations and the cumulative value of a predetermined pixel is "1", the predetermined pixel is represented by a pixel value of "255" among 0 to 255. Also, if the image is represented by pixel values ​​of 256 gradations and the cumulative value of a predetermined pixel is "1 / 2", the predetermined pixel is represented by a pixel value of "127" among 0 to 255.

[0038] According to this, by changing the predetermined value of x, the pixel value or brightness value of the abnormal part of the image displayed on the display 500 can be easily changed, and the image representation of the abnormal part can be adjusted to suit the preferences of the inspector IS.

[0039] However, the method of deriving the cumulative value is not limited to this. For example, the data derivation unit 408 derives an average value by accumulating multiple pieces of sampled data for each pixel and dividing the accumulated value by the total number of samples. Depending on this average value, the pixel values ​​or brightness values ​​of the image displayed on the display 500 are changed as described above.

[0040] This allows pixel values ​​or brightness values ​​in an abnormal area to be averaged over the total number of samples, making it easier to distinguish and recognize pixel areas that are likely to be abnormal from pixel areas that are unlikely to be abnormal.

[0041] The method of changing the pixel values ​​or brightness values ​​of an image according to the average value is not limited to this. For example, a predetermined pixel in a normal area where the average value is "0" may be represented by a pixel value of 0, and a predetermined pixel in an abnormal area where the average value is greater than "0" may be represented by a pixel value within a range of 127 to 255 according to the average value. However, the range of pixel values ​​in abnormal areas is not limited to this, and may be set to any range as long as the value is greater than 0.

[0042] This allows the pixel value of a specified pixel that is determined to be an abnormal area in at least one of multiple sampling data to be at least 127, making it easier for the inspector IS to recognize abnormal areas that are difficult to find visually.

[0043] Returning to FIG. 2, the model switching unit 404 switches from the first model to the second model using the Monte Carlo (MC) dropout method. The MC dropout method randomly sets the transmission of each neuron in the trained model to a value of 0, putting it in an on-off state. For example, the MC dropout method deactivates some neurons in the trained model with a predetermined probability, thereby turning off transmission from activated neurons to inactivated neurons or transmission from inactivated neurons to activated neurons. More specifically, the MC dropout method turns off transmission between neurons with a predetermined probability by setting the value of a parameter, which is a coefficient provided at the connection between neurons, to 0 with a predetermined probability.

[0044] The predetermined probability is, for example, 50%, but is not limited to this and can be set to any value. In this way, the model switching unit 404 switches from the first model to the second model by turning on and off the transmission of neurons in the trained model based on probability. However, instead of using the MC dropout method, a second model may be created in advance in which each neuron is assigned a probability distribution, and switching between the second model and the first model may be performed.

[0045] 5 is a diagram for explaining the transmission state of each neuron in the first model. As shown in FIG. 5, the trained model includes multiple layers L1, L2, L3, L4, L5, L6, L7, L8, L9, L10, L11, L12, L13, L14, L15, L16, L17, L18, L19, L20, L21, L22, L23, L24, L25, L2 M-1 , L M In Figure 5, each layer L1, L2, and L3 in the trained model are M-1 , L M and neurons N1, N2, N3, N4, N5, N M-1 , N M is shown in a simplified manner, and the actual number of layers and neurons is much higher.

[0046] The trained model (first model) has a neural network (NN). The neural network is a convolutional neural network (CNN) trained through supervised learning. However, neural networks other than convolutional neural networks may also be used. Furthermore, learning models other than neural networks may also be used.

[0047] The first model receives a captured image as input, and each neuron N1, N2, N3, N4, N5, N M-1 , N M By passing through the above, the sampling data 700 in which the output node values ​​0 and 1 are associated with each pixel is output as described above.

[0048] 6 is a diagram for explaining the transmission state of each neuron in the second model. As shown in FIG. 6, the second model has the same multiple layers L1, L2, L3 as the first model. M-1 , L M and neurons N1, N2, N3, N4, N5, N M-1 , N M However, in the second model, the MC dropout method is used to randomly switch on and off the transmission of each neuron in the trained model.

[0049] In Figure 6, the hatched neurons N1, N3, N5, and N M-1is a neuron whose transmission is turned off by MC dropout. In Figure 6, a captured image is input to the second model, and each neuron N2, N4, and N M By passing through the above, sampling data 800 is output in which each pixel is associated with an output node value of 0 or 1, as described above.

[0050] In addition, the same captured image is input to the second model, and the neuron transmission is randomly turned on and off using the MC dropout method, which outputs multiple sampling data 802, 804, with each pixel associated with an output node value of 0 or 1, as described above.

[0051] By randomly turning on and off the neuronal transmission, each layer L1,L M-1 , L M In this case, a probability distribution is generated as shown in Figure 4, and the internal variables in the trained model become random variables. In this way, by randomly turning on and off the transmission of each neuron in the first model, the processing algorithm within the trained model can be switched, and the state within the trained model can be treated as a probabilistic model.

[0052] By using the MC dropout method when executing the second model, the burden on the designer and development costs can be reduced compared to creating a second model in which each internal variable is assigned a probability distribution in advance.

[0053] Returning to FIG. 2, the transmission state storage unit 410 stores the plurality of sampling data 800, 802, 804 output from the second model in association with the transmission states (on / off states) of the neurons in the second model when the sampling data 800, 802, 804 were obtained.

[0054] The performance evaluation unit 412 compares the multiple pieces of sampling data 800, 802, and 804 output from the second model with one or more pieces of training data created in advance by capturing images of an inspection target including a predetermined abnormality. This comparison process is performed from the time the multiple pieces of sampling data 800, 802, and 804 are obtained by the second model until the next time new multiple pieces of sampling data are obtained by the second model. In other words, the comparison process is performed from the time the multiple pieces of sampling data 800, 802, and 804 are obtained while the camera 300 is not scanning, until the camera 300 scans again and the camera 300 is not scanning again and new multiple pieces of sampling data are obtained by the second model. By comparing the sampling data with the training data, the detection rate (the rate at which abnormalities are correctly detected as abnormalities) and the overdetection rate (the rate at which normal areas are mistakenly detected as abnormalities) can be derived from the difference between the sampled data and the training data.

[0055] In this way, the performance evaluation unit 412 derives the detection rate and overdetection rate of abnormal areas based on the difference between the sampled data and the teacher data, and evaluates the performance value of the sampled data based on the detection rate and overdetection rate. The performance value increases as the detection rate increases and the overdetection rate decreases.

[0056] Priority may be set for the detection rate and the overdetection rate. For example, the detection rate may be set to have a higher priority than the overdetection rate. This is because, even if the overdetection rate is high, abnormal locations will not be overlooked, but the lower the detection rate, the more likely it is that abnormal locations will be overlooked. However, conversely, the overdetection rate may be set to have a higher priority than the detection rate.

[0057] The transmission state storage unit 410 stores the corresponding relationship between the sampling data and the transmission state of the neuron, as well as the performance value evaluated by the performance evaluation unit 412 in association with the sampling data. In other words, the transmission state storage unit 410 stores the corresponding relationship between the sampling data, the transmission state of the neuron, and the performance value.

[0058] The on / off operation unit 414 controls the on / off of the neuron transmission in the second model based on the performance value evaluated by the performance evaluation unit 412, using the transmission state associated with the performance value. For example, the on / off operation unit 414 acquires, from the transmission state storage unit 410, the transmission state of the neuron in the trained model when the sampling data with the highest evaluated performance value was output. Then, the performance evaluation unit 412 controls the on / off of the neuron transmission in the second model so as to achieve the acquired transmission state.

[0059] This allows the on / off operation unit 414 to reproduce the transmission state of the neuron in the trained model when the sampling data with the highest evaluated performance value is output. Note that the on / off operation unit 414 may acquire a predetermined number of pieces of information about the transmission states of neurons with highly evaluated performance values ​​from the transmission state storage unit 410, and perform on / off operation of the transmission of the neuron in the second model so that the transmission states acquired in descending order of performance value become the transmission states acquired in descending order of performance value.

[0060] By preferentially reproducing the transmission state of a neuron with a high evaluation of its performance value before performing image processing using the second model during the second or subsequent non-scanning periods, it is possible to facilitate early detection of abnormalities in the inspection object 200. Furthermore, after reproducing the transmission state of a neuron with a high evaluation of its performance value, the internal variables of the second model may be converted into random variables using the MC dropout method, and the sampling results may be aggregated. This increases the detection rate of abnormalities that could not be detected using the transmission state of a neuron with a high evaluation of its performance value while the camera 300 is stopped from scanning. In this way, the image captured by the camera 300 is processed by the image processing device 400.

[0061] Returning to FIG. 1 , the display 500 displays the captured image output from the camera 300. The inspector IS can visually check the captured image displayed on the display 500 to see if there is any abnormality in the inspection target 200.

[0062] Furthermore, the display 500 outputs the sampling data 700 and statistical data 900 processed by the image processing device 400. By viewing the captured image together with the sampling data 700 and statistical data, the inspector IS can check whether or not there are any abnormalities in the inspection object 200.

[0063] According to this embodiment, by classifying each pixel of an image using the internal variables of a pre-created trained model as random variables, the accuracy of detecting abnormalities can be improved even after the trained model has been created.

[0064] In particular, the model switching unit 404 switches the trained model to the first model when the camera 300 is scanning, and switches the trained model to the second model when the camera 300 is not scanning. This makes it possible to detect abnormalities that cannot be detected by the first model when the camera 300 is not scanning. Furthermore, by using the first model, which has a shorter image processing time, when the camera 300 is scanning, it is possible to detect abnormalities more quickly than when the second model is used.

[0065] Although the embodiments have been described above with reference to the accompanying drawings, the present disclosure is not limited to the above-described embodiments. It is clear that a person skilled in the art can conceive of various modifications or alterations within the scope of the claims, and it is understood that such modifications also fall within the technical scope of the present disclosure.

[0066] For example, in the above embodiment, a case where semantic segmentation is applied as the output format of the trained model has been described. However, the output format of the trained model is not limited to this, and may be a discrimination format that simply outputs a scalar value of the abnormality degree 0 to 1 of the inspection target object 200.

[0067] This disclosure can promote better energy supplies through better facilities, thereby contributing to Goal 12 of the United Nations' Sustainable Development Goals (SDGs), "Ensure sustainable consumption and production patterns." [Explanation of symbols]

[0068] IS Inspector 100 Inspection Systems 200 Inspection objects 202 Abnormality 300 cameras 302 Imaging range 400 Image Processing Device 400a processor 400b memory 402 Scanning determination unit 404 Model Switching Unit 406 Sampling Generation Unit 408 Data Derivation Department 410 Transmission state memory unit 412 Performance Evaluation Department 414 On / off control unit 500 displays 600 captured images 700 Sampling Data 800 sampling data 802 Sampling Data 804 Sampling Data 900 Statistics

Claims

1. a camera for capturing an image of an object to be inspected; a sampling generation unit that inputs an image captured by the camera into a trained model created in advance, and outputs a plurality of pieces of sampling data in which output values ​​output from the trained model are linked to each pixel of the image, with internal variables of the trained model as random variables; a data derivation unit that performs statistical processing on the plurality of sampling data to derive one piece of statistical data; a scanning determination unit that determines the scanning of the camera; When the scanning determination unit determines that the camera is scanning, the trained model is switched to a first model in which the internal variables are fixed values; a model switching unit that switches the trained model to a second model in which the internal variable is a random variable when the scanning determination unit determines that the camera is not scanning; An inspection system comprising:

2. a camera for capturing an image of an object to be inspected; a sampling generation unit that inputs an image captured by the camera into a trained model created in advance, and outputs a plurality of pieces of sampling data in which output values ​​output from the trained model are linked to each pixel of the image, with internal variables of the trained model as random variables; a data derivation unit that performs statistical processing on the plurality of sampling data to derive one piece of statistical data; a performance evaluation unit that derives a detection rate and an overdetection rate of an abnormal portion of the inspection object based on a difference between the sampling data and teacher data, and evaluates a performance value of the sampling data based on the detection rate and the overdetection rate; a transmission state storage unit that stores the sampling data, a transmission state of a neuron in the trained model when the sampling data is obtained, and a performance value of the sampling data in association with each other; an on / off operation unit that performs on / off operation of the transmission of the neuron in the trained model based on the performance value and using the transmission state associated with the performance value; An inspection system comprising:

3. The data derivation unit multiplying each sampling data by a predetermined value and accumulating the multiplied values ​​within a predetermined upper limit to derive a cumulative value; 3. The inspection system according to claim 1 or 2.

4. The data derivation unit deriving an average value by integrating the plurality of sampling data for each pixel and dividing the integrated value by the total number of samplings; 3. The inspection system according to claim 1 or 2.

5. The model switching unit Executing the second model by stochastically turning on and off transmission of neurons in the trained model. The inspection system of claim 1 .

Images