External appearance inspection system and external appearance inspection method

Abstract

An external appearance inspection system according to the present invention comprises: a trained model that is created using a training image that is an external appearance image of a workpiece having no defect in terms of external appearance, outputs a training image feature amount corresponding to the training image, and outputs an inspection image feature amount corresponding to an inspection image that is an external appearance image of the workpiece to be inspected; a degree of deviation calculation unit that calculates a degree of deviation indicating the level of deviation between the training image feature amount and the inspection image feature amount; a processing device history management unit that manages the history related to a processing device that processes the workpiece to be inspected; and a cause determination unit that, on the basis of the history related to the processing device, determines whether a cause for the degree of deviation having exceeded a prescribed reference is caused by a change related to the processing device.

Description

Visual inspection system and visual inspection method

[0001] The present invention relates to an appearance inspection system and an appearance inspection method.

[0002] In visual inspections using deep learning, drift of an inspection object can sometimes result in false positives. There are techniques for detecting drift and relearning. Patent Document 1 discloses a management device for managing updates to a trained model in a classification device that receives state data indicating the state of an object and outputs a class into which the state of the object is classified and a confidence level for the classified class according to a trained model that classifies the state of the object into multiple classes. The management device includes: a derivation unit that derives, for each class, a distribution of confidence levels for classification results of a group of state data in a first time period and a second time period; and a determination unit that determines whether to update the trained model based on a first comparison result based on a first index between a group of state data in the first time period whose confidence level for the classification result is equal to or greater than a threshold and a group of state data in the second time period whose confidence level for the classification result is equal to or greater than the threshold; and a second comparison result based on a second index different from the first index between a group of state data in the first time period whose confidence level for the classification result is less than the threshold and a group of state data in the second time period whose confidence level for the classification result is less than the threshold.

[0003] Japanese Patent Application Publication No. 2023-073140

[0004] The invention described in Patent Document 1 cannot determine the cause of a change in the output of a trained model.

[0005] The appearance inspection system according to a first aspect of the present invention comprises a trained model created using training images, which are appearance images of workpieces without visual defects, and which outputs training image features corresponding to the training images and outputs inspection image features corresponding to inspection images, which are appearance images of the workpieces to be inspected; a deviation calculation unit that calculates a deviation indicating the degree of deviation between the training image features and the inspection image features; a processing device history management unit that manages the history of the processing device that processes the workpieces to be inspected; and a cause determination unit that determines, based on the history of the processing device, whether the cause of the deviation exceeding a predetermined standard is due to a change related to the processing device. An appearance inspection method according to a second aspect of the present invention is an appearance inspection method executed by a computing device having a trained model that is created using training images, which are appearance images of workpieces that have no visual defects, and that outputs features corresponding to input images, and includes a feature calculation step that outputs training image features corresponding to the training images and output inspection image features corresponding to inspection images, which are appearance images of the workpieces to be inspected, a deviation calculation step that calculates a deviation indicating the degree of deviation between the training image features and the inspection image features, a processing device history management step that manages the history of a processing device that processes the workpieces to be inspected, and a cause determination step that determines, based on the history of the processing device, whether the cause of the deviation exceeding a predetermined standard is due to a change related to the processing device.

[0006] According to the present invention, it is possible to determine the cause of a change in the output of a trained model.

[0007] Functional block diagram of the visual inspection system in the first embodiment. Overview diagram of a piston, which is the object to be inspected. Diagram showing the processing progress in the first embodiment. Hardware block diagram of the visual inspection system. Functional block diagram of the visual inspection system in the second embodiment. Flowchart showing the operation of the cause determination unit in the second embodiment. Functional block diagram of the visual inspection system in the third embodiment. Flowchart showing the processing of the cause determination unit in the third embodiment. Functional block diagram of the visual inspection system in the fourth embodiment. Diagram showing an area where the processing conditions of the crown surface are the same. Functional block diagram of the visual inspection system in the fifth embodiment. Flowchart showing the processing of the deviation degree calculation unit in the fifth embodiment.

[0008] -First Embodiment- A first embodiment of a visual inspection system will be described below with reference to FIGS.

[0009] FIG. 1 is a functional configuration diagram of an appearance inspection system 10. In this embodiment, an engine piston is shown as an example of an inspection object, but the inspection object is not limited thereto. The appearance inspection system 10 includes a trained model 11, a deviation calculation unit 12, a cause determination unit 13, and a processing device history management unit 14. The trained model 11 reads training images 21 and outputs training image features 23. The trained model 11 reads inspection images 22 and outputs inspection image features 24. The deviation calculation unit 12 calculates a deviation 25 indicating the degree of deviation between the training image features 23 and the inspection image features 24. The cause determination unit 13 determines whether a large deviation 25 is caused by a change in processing.

[0010] FIG. 2 is a schematic diagram of a piston 201, which is the object to be inspected. The piston 201 is used, for example, in an automobile engine. The piston 201 is generally cylindrical and is a cast material that has been subjected to multiple processes. The circular end portion at the top of the piston 201 in the figure is called the crown surface 202. The crown surface 202 includes a casting surface portion 203, which is the bare surface of the casting, a flat portion 204 formed by flattening the casting, a bored portion 205 formed by boring the casting, and an engraved portion 206 in which characters or the like are engraved on the flat portion 204.

[0011] 3A and 3B are diagrams showing the progress of the process. Fig. 3A is a diagram showing an example of a learning image 21 and an inspection image 22. The learning image 21 and the inspection image 22 are images obtained by photographing the piston 201 shown in Fig. 2 from directly above. Hereinafter, the learning image 21 and the inspection image 22 are also collectively referred to as appearance images 20. Returning to Fig. 1, the explanation will continue.

[0012] The trained model 11 receives an appearance image 20 as input and outputs a feature. As shown in FIG. 3( a), the training image 21 and the inspection image 22 are the same type of image. The image used to train the trained model 11 is called the training image 21, and the image to be inspected or evaluated using the trained model 11 is called the inspection image 22. However, the training image 21 is an image of a workpiece that has no defects in appearance, and the presence or absence of defects in the inspection image 22 is unknown. The feature is a training image feature 23 and an inspection image feature 24. The output of the trained model 11 corresponding to the training image 21 is the training image feature 23. The output of the trained model 11 corresponding to the inspection image 22 is the inspection image feature 24. The trained model 11 is a machine learning model trained using the training image 21, for example, a deep neural network such as a U-net.

[0013] The training images 21 are preferably all images used in training the trained model 11, and the test images 22 are preferably multiple samples, for example, 100 or more samples. In other words, it is preferable to calculate multiple training image features 23 and multiple test image features 24. Using multiple features makes it possible to grasp the overall trend of the set of training images 21 and the set of test images 22.

[0014] The training image feature 23 and the test image feature 24 are each a collection of values ​​indicating the features of each region obtained by dividing the training image 21 and the test image 22. The region division is the same for the training image 21 and the test image 22. For example, the resolution of the training image 21 and the test image 22 is 1980 x 1080, the feature is calculated for each 10 x 10 region, and the training image feature 23 and the test image feature 24 include 198 x 108 values. Figure 3(b) shows an example in which the training image 21 and the test image 22 are divided into 9 x 9 regions.

[0015] The deviation calculation unit 12 calculates a deviation 25 indicating the degree of deviation between the training image feature 23 and the inspection image feature 24. However, the deviation 25 also includes identification data indicating the workpiece shown in the inspection image 22. This identification data indicates the type of the workpiece, so to speak, and is not a product unique number or serial number for distinguishing each workpiece.

[0016] The deviation calculation unit 12 compares the training image feature amounts 23 with the test image feature amounts 24 to calculate the deviation amount 25. For example, the deviation calculation unit 12 calculates the average value for each region of the plurality of training image feature amounts 23 and the plurality of test image feature amounts 24, and calculates the ratio of the average values ​​as the deviation amount 25. That is, the deviation amount 25 in this case is the ratio of the average values ​​of the feature amounts for each region, and the cause determination unit 13 determines that an abnormality exists if there is even one region where this ratio is greater than a predetermined threshold. In the example shown in FIG. 3( c), one region W shown by hatching is greater than the threshold and is determined to be abnormal.

[0017] The cause determination unit 13 outputs an evaluation of the inspection image 22 based on the deviation 25 and the output of the machining device history management unit 14. The machining device history management unit 14 manages the machining device history 902 of the machining device 900 that machines the piston 201. The machining device history 902 in this embodiment includes whether or not the machining process for each workpiece has been changed. A change in the machining process includes a change in the machining shape, a change in the tool used for machining, etc.

[0018] The cause determination unit 13 determines that an abnormality exists when the deviation 25 is equal to or greater than a predetermined threshold, and determines that no abnormality exists when the deviation 25 is less than the predetermined threshold. If the deviation 25 is equal to or greater than the predetermined threshold, the cause determination unit 13 further performs the following determination. If it is determined from the output of the processing device history management unit 14 that there is a change in processing, the cause determination unit 13 determines that the abnormality was detected due to the change in processing. If it is determined from the output of the processing device history management unit 14 that there is no change in processing, the cause determination unit 13 determines that there is no change in processing and that an abnormality was detected. The processing device history management unit 14 may output a combination of the identification data of the workpiece and whether or not there has been a change to the cause determination unit 13, or may receive the identification data of the workpiece from the cause determination unit 13 and output whether or not there has been a change in the processing process for that workpiece to the cause determination unit 13.

[0019] FIG. 4 is a hardware configuration diagram of the visual inspection system 10. The visual inspection system 10 includes a central processing unit (CPU) 41, a read-only storage device (ROM) 42, a read / write storage device (RAM) 43, and an input / output device 44 serving as a user interface. The CPU 41 loads programs stored in the ROM 42 into the RAM 43 and executes them to perform the various calculations described above. The visual inspection system 10 may be realized by a field programmable gate array (FPGA), a rewritable logic circuit, or an application specific integrated circuit (ASIC), instead of the combination of the CPU 41, ROM 42, and RAM 43. Furthermore, the visual inspection system 10 may be realized by a different combination of configurations, such as a combination of the CPU 41, ROM 42, RAM 43, and FPGA.

[0020] The appearance inspection system 10 may be realized by one hardware device as shown in Fig. 4, or may be realized by two or more hardware devices. When the appearance inspection system 10 is realized by two or more hardware devices, a communication device 45 is provided to enable the hardware devices to communicate with each other. The communication device 45 is, for example, a network interface card. The input / output device 44 is, for example, a display that presents information to the user, a mouse or keyboard that accepts input from the user, etc.

[0021] The first embodiment described above provides the following advantageous effects. (1) The visual inspection system 10 includes a trained model 11 that is created using training images 21, which are appearance images 20 of workpieces that are free from visual defects, and that outputs training image features 23 corresponding to the training images 21 and inspection image features 24 corresponding to inspection images 22, which are appearance images 20 of the workpieces to be inspected; a deviation calculation unit 12 that calculates a deviation 25 indicating the degree of deviation between the training image features 23 and the inspection image features 24; a processing device history management unit 14 that manages the history of a processing device 900 that processes the workpieces to be inspected; and a cause determination unit 13 that determines, based on the history of the processing device 900, whether the deviation 25 has exceeded a predetermined standard because of a change to the processing device 900. Therefore, the visual inspection system 10 can determine the cause of the deviation between the inspection images 22 and the learning images 21.

[0022] - Second embodiment - A second embodiment of the visual inspection system will be described with reference to Figures 5 and 6. In the following description, the same components as in the first embodiment are designated by the same reference numerals, and differences will be mainly described. Points that are not particularly described are the same as in the first embodiment. This embodiment differs from the first embodiment mainly in that the date and time of changes related to processing are taken into consideration.

[0023] 5 is a functional configuration diagram of an appearance inspection system 10A according to the second embodiment. The differences from the first embodiment and the like are that the cause determination unit 13 is replaced with a cause determination unit 13A, the processing device history 902 is replaced with a processing device history 902A, and the processing device history management unit 14 is replaced with a processing device history management unit 14A. The processing of the trained model 11 and the deviation calculation unit 12 is the same as in the first embodiment.

[0024] The processing device history 902A includes data indicating the date and time of a change in the processing process for each workpiece. The deviation degree 25 includes not only identification data indicating the workpiece, but also data on the date and time when the workpiece related to the inspection image 22 was processed. Data on the date and time of processing may be included in each inspection image 22, or the date and time of processing may be managed independently of the inspection image 22, and the data on that date and time may be added to the deviation degree 25.

[0025] In addition to the operations of the first embodiment, when there is an area where the deviation 25 is equal to or greater than a predetermined threshold and when it is determined that there has been a change in the workpiece, the cause determination unit 13A determines whether the change occurred before the workpiece was processed. If the cause determination unit 13A determines that the change occurred before the workpiece was processed, it determines that the abnormality was caused by a change in the processing process, and if it determines that the change occurred after the workpiece was processed, it determines that the abnormality was caused by some other cause, or at least was not caused by a change in the processing process.

[0026] 6 is a flowchart showing the operation of the cause determination unit 13A in the second embodiment. First, in step S301, the cause determination unit 13A determines whether the ratio of the feature amounts in all regions is less than the threshold. If the cause determination unit 13A determines that the ratio of the feature amounts in all regions is less than the threshold, the process proceeds to step S302. If the cause determination unit 13A determines that there is at least one region where the ratio of the feature amounts is equal to or greater than the threshold, the process proceeds to step S303. In step S302, the cause determination unit 13A notifies the user using the input / output device 44 that no abnormality is found, and then ends the process shown in FIG. 6.

[0027] In step S303, the cause determination unit 13A determines whether a change in the processing process is recorded in the processing device history 902A, using the identification data of the workpiece included in the deviation 25. If the cause determination unit 13A determines that a change in the processing process is recorded, the process proceeds to step S304, and if the cause determination unit 13A determines that a change in the processing process is not recorded, the process proceeds to step S306. In step S304, the cause determination unit 13A determines whether the date and time of the change recorded in the processing device history 902A is earlier than the date and time included in the deviation 25. If the cause determination unit 13A determines that the change in the processing process precedes the processing, the process proceeds to step S305, and if the cause determination unit 13A determines that the change in the processing process follows the processing, the process proceeds to step S306.

[0028] In step S305, the cause determination unit 13A notifies the user using the input / output device 44 that an abnormality has been detected and that the cause is a change in the machining process, and then ends the processing shown in Fig. 6. In step S306, the cause determination unit 13A notifies the user using the input / output device 44 that an abnormality has been detected and that the cause is something other than a change in the machining process, and then ends the processing shown in Fig. 6.

[0029] The second embodiment described above provides the following advantageous effects: (2) If the date and time when the workpiece captured in the inspection image 22, in which the deviation 25 exceeds the predetermined standard, was processed is after the date and time when a change to the processing device 900 occurred, the cause determination unit 13A determines that the cause of the deviation 25 exceeding the predetermined standard is due to the change to the processing device 900. Therefore, the cause determination unit 13A can determine the cause of a large deviation 25 for each period.

[0030] (Variation 1 of the Second Embodiment) In the second embodiment described above, the cause determination unit 13A determines the cause of the deviation 25 exceeding the predetermined standard only once. However, the cause determination unit 13A may determine the cause multiple times. For example, the cause determination unit 13A may determine the cause each time a predetermined time has elapsed since the date and time of a change in the machining process included in the machining device history 902A of the target workpiece, or each time a predetermined time has elapsed since the date and time of machining the workpiece.

[0031] (Variation 2 of Second Embodiment) In the second embodiment described above, if the date and time when the workpiece captured in the inspection image 22, in which the deviation 25 exceeds the predetermined standard, was processed is after the date and time when a change to the processing device 900 occurred, the cause determination unit 13A determines that the cause of the deviation 25 exceeding the predetermined standard is due to a change to the processing device 900. However, instead of comparing the dates and times, the cause determination unit 13A may determine whether the date and time is after the date and time when the change occurred and within a predetermined period of time. This predetermined period is, for example, one day or one month.

[0032] -Third embodiment- A third embodiment of a visual inspection system will be described with reference to Figures 7 and 8. In the following description, the same components as those in the second embodiment are denoted by the same reference numerals, and differences will be mainly described. Points that are not particularly described are the same as those in the second embodiment. This embodiment differs from the second embodiment mainly in that re-learning of the trained model 11 is performed when the cause of the abnormality is not a change in the machining process.

[0033] 7 is a functional configuration diagram of an appearance inspection system 10B according to the third embodiment. The difference from the second embodiment is that the cause determination unit 13A is replaced with a cause determination unit 13B, and a learning execution unit 15 is added. The processing of the trained model 11, deviation calculation unit 12, and processing device history management unit 14A is the same as in the second embodiment. When the cause of the abnormality is not a change in the processing process, the cause determination unit 13B causes the learning execution unit 15 to re-train the trained model 11. The learning execution unit 15 may use not only the training images 21 used in previous training, but also inspection images 22 captured after the training images 21 were captured, for re-training the trained model 11.

[0034] 8 is a flowchart showing the processing of the cause determination unit 13B in the third embodiment. The difference from the flowchart shown in the second embodiment is that step S307 is added after step S306, and step S308 is added after step S305. In step S307, the cause determination unit 13B causes the learning execution unit 15 to re-learn the trained model 11, thereby terminating the processing shown in FIG. 8. In step S308, the cause determination unit 13B notifies the user, using the input / output device 44, that threshold adjustment is required, thereby terminating the processing shown in FIG. 8.

[0035] According to the third embodiment described above, the following advantageous effects can be obtained. (3) The appearance inspection system 10B includes a learning execution unit 15 that updates the parameters of the trained model 11. If the date and time when the workpiece captured in the inspection image 22, in which the deviation 25 exceeds a predetermined standard, was processed is before the date and time when a change related to the processing device 900 occurred (S304: NO in FIG. 8 ), the cause determination unit 13B causes the learning execution unit 15 to update the parameters of the trained model 11. Therefore, re-learning is performed only when necessary, thereby appropriately suppressing the cost of re-learning.

[0036] -Fourth embodiment- A fourth embodiment of a visual inspection system will be described with reference to Figures 9 and 10. In the following description, the same components as those in the first embodiment are denoted by the same reference numerals, and differences will be mainly described. Points that are not particularly described are the same as those in the first embodiment. This embodiment differs from the first embodiment mainly in that the deviation 25 is calculated for each region with different processing conditions.

[0037] 9 is a functional configuration diagram of a visual inspection system 10C according to the fourth embodiment. The difference from the first embodiment is that the deviation calculation unit 12 is replaced with a deviation calculation unit 12C. Fig. 10 will be explained first, and then the operation of the deviation calculation unit 12C will be explained.

[0038] FIG. 10 is a diagram showing regions (hereinafter also referred to as "blocks") on the crown surface 202 that have the same machining conditions. In FIG. 10, blocks with the same machining conditions are indicated by the same hatching. There are four blocks in total in FIG. 10, with the casting surface block 601 indicated by diagonal hatching, the flat block 602 indicated by cross hatching, the bored block 603 indicated by vertical hatching, and the engraved block 604 indicated by horizontal hatching. The regions may be divided along the boundaries of each machined portion on the image, in addition to being divided into a grid pattern. In this embodiment, differences in "machining conditions" can also be rephrased as differences in the machining method, type of machining, machines used for machining, etc.

[0039] The deviation calculation unit 12C tally and compares the training image feature amounts 23 and the test image feature amounts 24 for each block shown in Fig. 10. That is, in the first three embodiments, the ratio was calculated for each grid area such as 198x108 or 9x9, but in this embodiment, the ratios are summed for each block, so in the example of Fig. 10, the ratio is calculated for each of four blocks. The process of the cause determination unit 13C is the same as in the third embodiment, except that the total number of blocks is reduced.

[0040] According to the fourth embodiment described above, the following advantageous effects can be obtained: (4) The deviation calculation unit 12C calculates the deviation 25 for each region where the machining conditions for the workpiece are different. Therefore, the cause of the deviation can be determined for each machining condition.

[0041] Fifth Embodiment A fifth embodiment of a visual inspection system will be described with reference to Figures 11 and 12. In the following description, the same components as those in the fourth embodiment are designated by the same reference numerals, and differences will be mainly described. Points that are not specifically described are the same as those in the fourth embodiment. This embodiment differs from the fourth embodiment mainly in that the deviation 25 is calculated by comparing histograms of feature amounts.

[0042] 11 is a functional configuration diagram of a visual inspection system 10D according to the fifth embodiment. The difference from the fourth embodiment is that the deviation calculation unit 12C is replaced with a deviation calculation unit 12D. The processing of the trained model 11 and the cause determination unit 13B is the same as in the fourth embodiment.

[0043] FIG. 12 is a flowchart showing the processing of the deviation calculation unit 12D. In the description of this flowchart, "n" represents an integer. In step S321, the deviation calculation unit 12D obtains all feature values ​​of each block in the training image 21 to create a first histogram H1. In the following step S322, the deviation calculation unit 12D obtains all feature values ​​of each block in the inspection image 22 to create a second histogram H2. In the following step S323, the deviation calculation unit 12D accumulates and tabulates the first histogram H1 using a predetermined bin width. Specifically, the variable H1_1 is the sum of the number of data items contained in bins 0 to 1 of the first histogram H1, the variable H1_2 is the sum of the number of data items contained in bins 0 to 2 of the first histogram H1, and the variable H1_3 is the sum of the number of data items contained in bins 0 to 3 of the first histogram H1. This calculation is repeated up to the maximum bin value. H1_n is a cumulative value, so for example, H1_8 is a value equal to or greater than H_7.

[0044] In the following step S324, the deviation calculation unit 12D accumulates and tabulates the second histogram H2 over a predetermined bin width. This bin width is the same as in step S323, and the calculation method is also the same as in step S323, so details will be omitted. In the following step S325, the deviation calculation unit 12D calculates the absolute value of the difference between the accumulated values ​​in steps S323 and S324. Specifically, D1 is calculated as the absolute value of the difference between H1_1 and H2_1, and D2 is calculated as the absolute value of the difference between H1_2 and H2_2. Calculations are made in the same manner up to Dn.

[0045] In the following step S326, the deviation calculation unit 12D sets the maximum value of D1, D2, ..., Dn as the cumulative maximum difference D. In the following step S327, the deviation calculation unit 12D calculates the test statistic K using the following equation 1.

[0046] K=D*{Na*Nb / (Na+Nb)}^0.5 (Formula 1)

[0047] In Equation 1, Na is the number of samples of the training image 21, and Nb is the number of samples of the test image 22. In Equation 1, the test statistic K is the product of D and the square root of the value obtained by dividing the product of Na and Nb by the sum of Na and Nb. In the following step S328, the deviation calculation unit 12D refers to the test table for the Kolmogorov-Smirnov test and identifies the significance level α corresponding to the limit value equal to the test statistic K. In the following step S329, the deviation calculation unit 12D calculates the deviation 25 using Equation 2 shown below.

[0048] Dev = 1 - α (Equation 2)

[0049] In Equation 2, Dev represents the deviation 25. The above is the process of the deviation calculation unit 12D in the fifth embodiment.

[0050] According to the fifth embodiment described above, the following advantageous effects can be obtained: (5) The deviation calculation unit 12D calculates the deviation 25 by comparing histograms obtained based on the training image feature amounts 23 and the test image feature amounts 24. Therefore, the deviation 25 can be calculated with high accuracy.

[0051] In the above-described embodiments and modifications, the functional block configurations are merely examples. Some functional configurations shown as separate functional blocks may be integrated, or a configuration shown in a single functional block diagram may be divided into two or more functions. Furthermore, some of the functions of each functional block may be provided by other functional blocks.

[0052] In the above-described embodiments and modifications, the program is stored in ROM 42. However, the program may be stored in a rewritable nonvolatile storage device (not shown). Furthermore, the visual inspection system may be provided with an input / output interface (not shown), and the program may be loaded from another device as needed via the input / output interface and a medium available to the visual inspection system. Here, the medium refers to, for example, a storage medium detachable from the input / output interface, or a communication medium, i.e., a wired, wireless, or optical network, or a carrier wave or digital signal propagating through the network. Furthermore, some or all of the functions realized by the program may be realized by a hardware circuit or FPGA.

[0053] The above-described embodiments and modifications may be combined with each other. In particular, the learning execution unit 15 may be omitted from the configurations of the fourth and fifth embodiments, and the cause determination unit 13 of the first and second embodiments may be provided. Although various embodiments and modifications have been described above, the present invention is not limited to these. Other aspects conceivable within the scope of the technical concept of the present invention are also included within the scope of the present invention.

[0054] 10, 10A, 10B, 10D: Appearance inspection system 11: Trained model 12, 12B, 12C, 12D: Deviation calculation unit 13, 13A, 13B, 13C: Cause determination unit 14, 14A: Processing device history management unit 15: Learning execution unit 20: Appearance image 21: Training image 22: Inspection image 23: Training image feature amount 24: Inspection image feature amount 25: Deviation amount

Claims

1. An appearance inspection system comprising: a trained model created using training images, which are appearance images of workpieces without visual defects, and which outputs training image features corresponding to the training images and also outputs inspection image features corresponding to inspection images, which are appearance images of the workpieces to be inspected; a deviation calculation unit which calculates a deviation indicating the degree of deviation between the training image features and the inspection image features; a processing device history management unit which manages the history of the processing device that processes the workpieces to be inspected; and a cause determination unit which determines, based on the history of the processing device, whether the cause of the deviation exceeding a predetermined standard is due to a change in the processing device.

2. An appearance inspection system as described in claim 1, wherein the cause determination unit determines that the cause of the deviation exceeding the predetermined standard is due to a change related to the processing device if the date and time when the processed product photographed in the inspection image in which the deviation exceeds the predetermined standard was processed is after the date and time when a change related to the processing device occurred.

3. An appearance inspection system as described in claim 2, wherein, after a predetermined time has elapsed from the date and time when the change to the processing device occurred, the appearance inspection system again determines whether the cause of the deviation exceeding the predetermined standard is due to the change to the processing device.

4. An appearance inspection system as described in claim 1, wherein the cause determination unit determines that the cause of the deviation exceeding the predetermined standard is due to a change in the processing device if the date and time when the processed product photographed in the inspection image in which the deviation exceeds the predetermined standard was processed is within a predetermined period after the date and time when the change in the processing device occurred.

5. An appearance inspection system as described in claim 1, further comprising a learning execution unit that updates the parameters of the trained model, wherein the cause determination unit causes the learning execution unit to update the parameters of the trained model if the date and time when the workpiece captured in the inspection image in which the deviation exceeds a predetermined standard was processed is before the date and time when a change related to the processing device occurred.

6. An appearance inspection system according to claim 1, wherein the deviation calculation unit calculates the deviation for each area where the processing conditions for the workpiece are different.

7. An appearance inspection system according to claim 1, wherein the deviation calculation unit calculates the deviation by comparing a histogram obtained based on the training image features and the inspection image features.

8. A visual inspection method executed by a computing device having a trained model that is created using training images, which are appearance images of workpieces that have no visual defects, and that outputs features corresponding to input images, the visual inspection method comprising: a feature calculation step that outputs training image features corresponding to the training images and output inspection image features corresponding to inspection images, which are appearance images of the workpiece to be inspected; a deviation calculation step that calculates a deviation indicating the degree of deviation between the training image features and the inspection image features; a processing device history management step that manages the history of a processing device that processes the workpiece to be inspected; and a cause determination step that determines, based on the history of the processing device, whether the cause of the deviation exceeding a predetermined standard is due to a change in the processing device.

9. An appearance inspection method as set forth in claim 8, wherein in the cause determination step, if the date and time when the workpiece photographed in the inspection image in which the deviation exceeds a predetermined standard was processed is after the date and time when a change occurred in the processing device, it is determined that the cause of the deviation exceeding the predetermined standard is due to a change in the processing device.

10. A visual inspection method as set forth in claim 8, wherein in the cause determination step, after a predetermined time has elapsed from the date and time when the change to the processing device occurred, it is determined again whether the cause of the deviation degree exceeding the predetermined standard is due to the change to the processing device.

11. An appearance inspection method as set forth in claim 8, wherein in the cause determination step, if the date and time when the workpiece photographed in the inspection image in which the deviation exceeds a predetermined standard was processed is within a predetermined period after the date and time when a change occurred in the processing device, it is determined that the cause of the deviation exceeding the predetermined standard is due to a change in the processing device.

12. An appearance inspection method as described in claim 8, wherein in the cause determination step, the parameters of the trained model are updated if the date and time when the workpiece photographed in the inspection image in which the deviation exceeds a predetermined standard was processed is before the date and time when a change occurred to the processing device.

13. A visual inspection method according to claim 8, wherein in the deviation calculation step, the deviation is calculated for each area where the processing conditions for the workpiece are different.

14. An appearance inspection method according to claim 8, wherein in the deviation calculation step, the deviation is calculated by comparing a histogram obtained based on the training image features and the inspection image features.

Images