Ai support system

Abstract

To provide a user-friendly AI support system which reduces the number of inspection targets to be visually inspected by workers.SOLUTION: An AI support system comprises a photographed image obtaining unit 11, a quality determination unit 2, a display control unit 13, and a storage unit 12. The quality determination unit 2 includes: a first determination unit 22 which excludes a photographed image G in which an inspection target is determined to be normal; and a second determination unit 23 which performs clustering of the inspection target based on the photographed image G in which the inspection target is determined to be defective by the first determination unit 22. The display control unit 13 displays on a worker terminal 6 the photographed image G clustered by the second determination unit 23 and defect classification of the inspection target.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an AI support system for image quality determination work on production lines in factories, etc. [Background technology]

[0002] Conventionally, there is known an AI support system that acquires images of inspection objects captured by a camera installed on a production line via a network and judges the pass / fail of the inspection objects using an AI model (see, for example, Patent Document 1).

[0003] The AI ​​support system described in Patent Document 1 displays the result of the inspection target being normal on the worker's terminal when the inspection target is determined to be normal, and displays the result of the inspection target being defective along with a captured image of the inspection target on the worker's terminal when the inspection target is determined to be defective. Also, when the inspection target is defective, the result of the inspection target being determined to be defective and the captured image are displayed on the worker's terminal, and the worker tags the cause of the defect and re-learns the AI ​​model. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2022 / 124316 Summary of the Invention [Problem to be solved by the invention]

[0005] The AI ​​support system described in Patent Document 1 only displays the judgment result when the inspection object is normal on the worker's terminal, which reduces the consumption of communication resources. However, when the inspection objects are, for example, nonwoven fabrics that flow sequentially from rolls to the production line, the number of inspection objects becomes enormous. As a result, the number of inspection objects that workers visually inspect is large, leaving room for improvement.

[0006] In addition, the AI ​​support system described in Patent Document 1 displays the judgment result and its captured image on the worker's terminal when the inspection object is defective, and the worker tags the cause of the defect and re-learns the AI ​​model. However, when there are a wide variety of causes for defects in the inspection objects, specialized knowledge is required for the worker to tag the causes of the defects, so there is room for improvement in terms of usability.

[0007] Therefore, there is a need for an easy-to-use AI support system that can reduce the number of items that workers need to visually inspect. [Means for solving the problem]

[0008] A characteristic configuration of the AI ​​support system according to the present invention includes an image acquisition unit that acquires an image of an inspection object, a pass / fail judgment unit that judges the pass / fail of the inspection object based on the captured image, a display control unit that controls a display screen of an operator terminal, and a memory unit that stores the captured images for which the inspection object is judged to be defective by the pass / fail judgment unit, wherein the pass / fail judgment unit has a first judgment unit that excludes the captured images for which the inspection object is judged to be normal, and a second judgment unit that clusters the inspection object based on the captured images for which the inspection object is judged to be defective by the first judgment unit, and the display control unit displays the captured images clustered by the second judgment unit and the defect classification of the inspection object on the operator terminal.

[0009] The pass / fail judgment section in this configuration has a first judgment section that excludes captured images for which the inspection object is judged to be normal, and is provided with a memory section that stores captured images for which the inspection object is judged to be defective. Therefore, even if there are a huge number of inspection objects that are normal, the storage capacity can be reduced by not storing captured images that are judged to be normal.

[0010] Furthermore, the pass / fail judgment unit in this configuration has a second judgment unit that performs clustering of the inspection objects based on the captured images in which the inspection objects have been judged to be defective by the first judgment unit. In a work site where there is a huge number of inspection objects that are judged to be normal, the second judgment unit narrows down the captured images that have been judged to be defective by the first judgment unit and performs clustering of the inspection objects, thereby significantly increasing the processing speed.

[0011] Furthermore, the display control unit in this configuration displays the captured images clustered by the second judgment unit and the defect classification of the inspection objects on the worker terminal, so that the worker can visually check the captured images with a significantly reduced number of inspection objects and correct the defect classification. As a result, the worker can determine and correct the defect classification of the inspection objects by looking only at the captured images with a significantly reduced number of sheets, which reduces the number of inspection objects visually checked by the worker, resulting in a user-friendly AI support system.

[0012] Another characteristic configuration is that the memory unit stores the defect classification for the captured image in association with the captured image, and when an input classification input from the operator terminal differs from the defect classification, the pass / fail judgment unit re-learns based on the input classification.

[0013] In this configuration, since the storage unit stores the defective classification for the captured image in association with the captured image, for example, when a certain number of captured images have been accumulated, the captured images can be read out sequentially by batch processing, and the worker can correct the defective classification. Then, the pass / fail judgment unit re-learns based on the input classification input by the worker, thereby improving the judgment accuracy of the AI ​​support system.

[0014] Another characteristic configuration is that when the first judgment unit judges that the inspection object is normal, the display control unit does not display anything on the worker terminal, and when the second judgment unit judges that the inspection object is normal, the display control unit displays the probability of normality and the captured image on the worker terminal.

[0015] In this configuration, since inspection objects judged as normal by the first judgment unit are excluded and nothing is displayed on the worker terminal, when there are a huge number of inspection objects that are normal, the worker does not need to check anything, and work efficiency is extremely improved. Also, since the probability of normality and the captured image are displayed on the worker terminal only when the inspection object is judged as normal by the second judgment unit, the workload of the worker is not increased even if the worker checks them together with the captured image judged as defective.

[0016] Another characteristic configuration is that the memory unit stores the defect classification for the captured image in association with the captured image, and when the first judgment unit judges that the inspection object is defective, the display control unit displays the captured image and the attributes of the inspection object as thumbnails on the operator terminal, and displays a list of judgment probabilities of the defect classification in parallel with the thumbnail display.

[0017] As in this configuration, the storage unit stores the defect classification for the captured image in association with the captured image, so that the worker can input the defect classification by batch processing when he has time. Also, the captured image and the attributes of the inspection object are displayed as thumbnails on the worker terminal, and a list of the judgment probabilities of the defect classification is displayed in parallel with the thumbnail display, so that even if the worker is not a skilled worker with specialized knowledge, the accuracy of the defect classification can be easily confirmed, and the work efficiency of the defect classification input work can be improved.

[0018] Another characteristic configuration is that the memory unit stores the defect classification for the captured image in association with the captured image, the pass / fail judgment unit calculates features of the captured image for the inspection object judged to be defective by the first judgment unit, and the display control unit displays a collection of the features for each defect classification in a two-dimensional space.

[0019] As in this configuration, by displaying a collection of features for each defect classification in a two-dimensional space, defect classifications with close feature distances can be easily understood. As a result, for example, if a defect is re-learned as defect A but is close to defect B in the two-dimensional space, it becomes possible to determine whether the captured image used for re-learning is suitable, such as excluding it from the re-learning target. [Brief description of the drawings]

[0020] [Figure 1] This is an overall diagram of the AI ​​support system. [Diagram 2] FIG. 1 is a block diagram showing the configuration of an AI support system. [Diagram 3] FIG. 1 is a control flow diagram of the AI ​​support system. [Figure 4A] FIG. 1 is an explanatory diagram of re-learning of the AI ​​support system. [Figure 4B] This is a histogram before and after re-learning. [Diagram 5] 13 is an example of a defect classification for an inspection target. [Figure 6] This is an example of the management screen of the AI ​​support system. [Figure 7] 13 is an example of a display screen of a worker terminal. [Figure 8] 13 is an example of a display screen of a worker terminal. [Figure 9] 13 is an example of a display screen of a worker terminal. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] Hereinafter, an embodiment of the AI ​​support system according to the present invention will be described with reference to the drawings. In this embodiment, an AI support system X that inspects nonwoven fabric products will be described as an example of the AI ​​support system. However, the present invention is not limited to the following embodiment, and various modifications are possible within the scope of the gist of the present invention.

[0022] Fig. 1 is an overall explanatory diagram showing an image of the AI ​​support system X, and Fig. 2 is a block diagram showing a schematic configuration of the AI ​​support system X. As shown in Fig. 2, the AI ​​support system X is configured with each of the functional parts of an image acquisition unit 11, a quality determination unit 2, a storage unit 12, and a display control unit 13. Each of these functional parts is constructed with hardware or software or both, with a processor such as a CPU as a core member, in order to perform processing related to defect determination of scratches, dirt, etc. on the surface of an inspection target (product) in a factory production line.

[0023] The captured image acquisition unit 11 acquires a captured image G of a nonwoven fabric product (an example of an inspection target, hereinafter referred to as a "product") flowing through a production line in a factory. The nonwoven fabric is shipped as a product after being cut by a slit blade from a roll of raw fabric sent out from a roll. As shown in FIG. 1, in this embodiment, in order to check the state of these products, a camera 5 installed at a predetermined installation location above the product conveying line sequentially images the nonwoven fabric products that are continuously flowing. Note that the installation form of the camera 5 shown in FIG. 1 is an example, and the camera 5 may be installed in an installation form different from that shown in FIG. 1. That is, the number of cameras 5 may be two or more, or may be one. In addition, the installation location may be a structure dedicated to installing the camera 5, or may be an existing structure already installed in the factory for another purpose. As shown in FIG. 2, the captured image acquisition unit 11 receives a plurality of captured images G captured by the camera 5 via a communication line such as an Internet line, and transmits them to the quality determination unit 2. That is, the captured image acquisition unit 11 is a communication interface.

[0024] The quality judging section 2 judges the quality of the product based on the captured image G acquired by the captured image acquiring section 11. The quality judging section 2 has a feature amount calculation section 21, a first judgment section 22, a second judgment section 23, a deviation map generation section 24, and a re-learning section 25. The quality judging section 2 in this embodiment uses, for example, a neural network to map the feature amount space with plots corresponding to the respective feature amounts so that the higher the similarity between the feature amounts, the shorter the distance, and the lower the similarity, the longer the distance, and calculates the length of this distance as the deviation. In order to use such a neural network to appropriately calculate the degree of discrepancy between two feature quantities, the pass / fail judgment unit 2 is configured with a trained model that has been deep-learned using known metric learning or the like so that when two feature quantities with high similarity are input, the distance between the two plots corresponding to these two feature quantities in the feature space becomes short, and when two feature quantities with low similarity are input, the distance between the two plots corresponding to these two feature quantities in the feature space becomes long.

[0025] The feature amount calculation unit 21 calculates a feature amount (distance vector) of the input captured image G with respect to a collection of normal captured images G, using a trained model that has undergone deep learning of multiple captured images G in which the product state is normal or defective. This feature amount may be a Euclidean distance or a Mahalanobis distance.

[0026] The first determination unit 22 excludes the captured image G that is determined to be normal. Specifically, the first determination unit 22 determines that the captured image G whose abnormality level based on the feature amount calculated by the feature amount calculation unit 21 is equal to or lower than a predetermined threshold value (for example, 0.01) is normal, and does not transmit the captured image G to the second determination unit 23. As shown in FIG. 1, the first determination unit 22 outputs the feature amount calculated by the feature amount calculation unit 21 as an abnormality level based on a distance vector from 0 to 1, sets a threshold value for the captured image G whose probability of being abnormal is 100%, and excludes the captured image G whose probability is equal to or lower than the threshold value. This threshold value may differ depending on the product attribute (product number, etc.), and is set in advance as a boundary where the probability of being abnormal is 100% for a learned model that has been deep-learned for each product attribute. In the example shown in FIG. 1, when the threshold value is set to 0.4, defect data is included, so the threshold value is set to 0.01. This setting may be set by an administrator by visually checking the histogram, or a threshold value that does not include defect data may be automatically set.

[0027] The second judgment unit 23 performs clustering of products based on the captured images G in which the products are judged to be defective by the first judgment unit 22. Specifically, the second judgment unit 23 judges the probability of similarity with respect to a plurality of preset defect classifications based on the feature amount (distance vector) calculated by the feature amount calculation unit 21. As shown in FIG. 1, the second judgment unit 23 judges the captured images G in which the products are judged to be defective to have a 60% probability of defect A, a 25% probability of defect B, and a 15% probability of defect C in order of decreasing magnitude of the feature amount (distance vector) with respect to defects A to C in the defect classifications. In this embodiment, there are multiple types of defect classifications, and stains, lint, dust, tears, gouges, wrinkles, dirt, insects, etc. are set as the defect classifications (see also FIG. 5). Note that the second judgment unit 23 judges the normal captured images G that were not excluded by the first judgment unit 22 to have a higher probability of normality than any of the defect classifications.

[0028] The deviation map generating unit 24 generates a feature space (a two-dimensional space that is a collection of points obtained by plotting distance vectors on XY coordinates) for the feature of the captured image G calculated by the feature calculating unit 21. Specifically, the deviation map generating unit 24 maps the feature groups of normal products excluded by the first judging unit 22 and the feature groups of products clustered by the second judging unit 23 for normal and multiple defective classifications. As shown in FIG. 9, the deviation map generating unit 24 generates a map in which the normal feature groups and the feature groups of defects A to D are plotted in a two-dimensional space so that the distances are separated according to the respective deviations. This makes it possible to determine whether the captured image G used for re-learning is suitable, for example, by excluding the captured image G from the re-learning target when the image is re-learned as defect A but is located near defect B in the two-dimensional space.

[0029] The re-learning unit 25 learns based on the input classification (including a defective classification and a normal classification) input from the operator terminal 6 described later. Even if the defective classification determined by the second judgment unit 23 (including a case where the first judgment unit 22 judges the defect as defective but the second judgment unit 23 judges the defect as normal, the same applies below) is incorrect, re-learning can be performed based on the correct defective classification. Specifically, as shown in FIG. 4A, the second judgment unit 23 classifies defect A as defective because the probability of defect A being 60%, but when the operator visually checks the captured image G, it is defect B. In this case, defect A is corrected to the input defect B and re-learned in the trained model. As a result, the probability of defect B being the judgment result after re-learning becomes 80%. The re-learning unit 25 stores the captured image G and the corrected defect classification in the memory unit 12 for each product attribute (product number), and reads out the captured image G and its defect classification from the memory unit 12 when a certain amount of data has been accumulated or at predetermined intervals, and performs deep learning using known metric learning or the like to reconstruct the learned model.

[0030] 2, the storage unit 12 is configured with a storage medium such as an HDD or SSD that stores a trained model, a plurality of captured images G, and their attributes. In the storage unit 12 in this embodiment, the attributes of each captured image G are stored, such as the inspection date and time, feature amount, degree of abnormality (classification of normal / defective), defect classification, product number, and its specifications (width, length, number of divisions, etc.). This storage unit 12 does not need to store the plurality of captured images G and their attributes that are determined to be normal by the first determination unit 22, but it is preferable to store only the normal feature amount.

[0031] That is, the storage unit 12 stores the defect classification for the captured image G in association with the captured image G. This allows, for example, when a large amount of captured images G has been accumulated, the captured images G to be sequentially read out by batch processing, and enables an operator to perform the defect classification judgment work and correction work when he or she has time. Then, the re-learning unit 25 performs re-learning based on the input classification input by the operator, thereby improving the judgment accuracy of the AI ​​support system X.

[0032] The display control unit 13 controls the display screen of the worker terminal 6, and causes the worker terminal 6 to display the captured images G clustered by the second judgment unit 23 and the defective classification of the product. The worker terminal 6 is configured as a personal computer, a mobile terminal such as a tablet, or the like that can be operated by a worker who performs inspection work in the factory. The display control unit 13 transmits the multiple captured images G and their attributes stored in the storage unit 12, and data related to the feature space generated by the deviation map generation unit 24, to the worker terminal 6, and causes it to be displayed on the screen.

[0033] When the first judgment unit 22 judges that the product is normal, the display control unit 13 causes the worker terminal 6 to display nothing, and when the second judgment unit 23 judges that the product is normal, the display control unit 13 causes the worker terminal 6 to display the probability of normality (including the probability of other defective classifications) and the captured image G. In this way, since the product judged to be normal by the first judgment unit 22 is excluded and nothing is displayed on the worker terminal 6, when there is a huge number of normal products such as nonwoven fabrics, the worker does not need to check anything, and the work efficiency is extremely improved. Also, only when the second judgment unit 23 judges that the product is normal, the probability of normality and the captured image G are displayed on the worker terminal 6, so that the work burden is not increased even if the worker checks them together with the captured image G judged to be defective.

[0034] 7, when the first judgment unit 22 judges that the product is defective, the display control unit 13 causes the captured image G and the attributes of the product to be displayed as thumbnails on the worker terminal 6 as the judgment results of the second judgment unit 23, and also causes a list of judgment probabilities of defective classification (including the probability of judgment as normal) to be displayed in parallel to the thumbnail display. In this way, because the captured image G and the attributes of the product are displayed as thumbnails on the worker terminal 6, and a list of judgment probabilities of defective classification is displayed in parallel to the thumbnail display, even if the worker is not a skilled worker with specialized knowledge, it is possible to easily check the accuracy of the defective classification, and the work efficiency of the defective classification correction work can be improved.

[0035] As described above, the pass / fail judgment unit 2 in this embodiment has a first judgment unit 22 that excludes captured images G for which the product has been judged to be normal, and is provided with a memory unit 12 that stores captured images G for which the product has been judged to be defective. Therefore, even in the case where there is a huge number of normal products, such as nonwoven fabrics, the storage capacity can be reduced by not storing captured images G for which the product has been judged to be normal.

[0036] In addition, the pass / fail judgment unit 2 has a second judgment unit 23 that performs product clustering based on the captured images G in which the first judgment unit 22 judges the product to be defective. In a work site where there is a huge number of products judged to be normal, the second judgment unit 23 narrows down the captured images G that are defective to perform product clustering, thereby significantly increasing the processing speed.

[0037] Furthermore, the display control unit 13 causes the captured images G clustered by the second judgment unit 23 and the defective classification of the products to be displayed on the worker terminal 6, so that the worker can visually check the captured images G in which the number of products has been greatly reduced, and judge and correct the defective classification. As a result, the worker can judge and correct the defective classification of the products by looking only at the captured images G in which the number has been greatly reduced, resulting in an easy-to-use AI support system X.

[0038] Next, the control process of the AI ​​support system X will be explained mainly with reference to Fig. 3. The control process described below is composed of a program executed by a computer, an application that runs on this program, and a non-transitory storage medium that stores the program or the application.

[0039] When products start to flow on the product conveying line, the camera 5 sequentially captures images of the nonwoven fabric products that are continuously flowing. Then, the captured image acquisition unit 11 sequentially captures the captured images G captured by the camera 5 (#31, captured image acquisition process). Next, the captured images G are transmitted to the quality determination unit 2 (trained model) via an Internet line or the like, and the feature amount calculation unit 21 calculates the feature amount corresponding to each captured image G (feature amount calculation process). At this time, the feature amount calculation unit 21 converts the feature amount into a value of 0 to 1 as an abnormality degree, and calculates the abnormality degree (#32, abnormality degree calculation process). A histogram showing the distribution of the abnormality degree is shown in the upper left of FIG. 1, and the threshold value is set to 0.01, for example.

[0040] Next, the first judgment unit 22 judges whether the degree of abnormality calculated by the feature amount calculation unit 21 is equal to or less than a predetermined threshold (e.g., 0.01) (#33, first judgment process). If the degree of abnormality is equal to or less than the predetermined threshold, the first judgment unit 22 judges the captured image G to be normal, does not transmit the captured image G to the second judgment unit 23, transmits only the feature amount, and does not transmit anything to the operator terminal 6 or notify that it is normal (#33 YES, #40, #41). In other words, since the products judged to be normal by the first judgment unit 22 are excluded and nothing is displayed on the operator terminal 6, when there are a huge number of normal products such as nonwoven fabrics, the operator does not need to check anything, and the work efficiency is extremely improved.

[0041] On the other hand, if the degree of abnormality exceeds a predetermined threshold, the first judgment unit 22 judges the captured image G to be defective and transmits the captured image G and its features to the second judgment unit 23, and the memory unit 12 stores the captured image G and its features (#34, storage process).

[0042] Next, the second judgment unit 23 performs clustering of the products based on the captured images G in which the first judgment unit 22 judged the product to be defective (#35, second judgment process). FIG. 5 shows an example of clustering by the second judgment unit 23. In case 1, since there is a large depression in the nonwoven fabric, it is judged to be a gouge, in case 2, since there is small dirt attached, it is judged to be dirt, and in case 3, since there are multiple scattered stains, it is judged to be stains. These judgments are automatically made by a trained model that has been deep-learned by metric learning or the like, so there is no need for the worker to visually track the product or judge by looking at the captured images G.

[0043] When the second judgment unit 23 executes clustering, it calculates the judgment probability of the defective classification based on the deviation from the judged defective classification (e.g., the distance from the average value of the features belonging to the same defective classification) (#36, second judgment process). The defective classification and judgment probability linked to the captured image G are stored in the memory unit 12 (storage process). An example of the judgment probability is shown in FIG. 7. As shown in the right diagram of FIG. 7, it is judged that the judgment probability of defect A is 96%, the judgment probability of defect B is 2%, the judgment probability of defect C is 1%, and the judgment probability of defect D is 1%.

[0044] Further, the deviation map generating unit 24 generates a feature space (a two-dimensional space that is a collection of points in which distance vectors are plotted on XY coordinates) for the feature of the captured image G calculated by the feature calculating unit 21 (deviation map generating process). An example of the deviation map is shown in Fig. 9. As shown in Fig. 9, point cloud maps are created for normal, defect A, defect B, defect C, and defect D, respectively, and it can be determined whether the captured image G is suitable for relearning.

[0045] Next, based on the defective classification (including the judgment that the image is normal) judged by the second judgment unit 23 and the captured image G, the display control unit 13 generates display screen data for the operator terminal 6 and transmits it to the operator terminal 6 (#37, display process). The operator terminal 6 displays a list screen as shown in FIG. 6. In this list, the number of captured images G (the number of defects) stored as defective classifications for each part number of the lot number is displayed, and the status is displayed as to whether "slit registration" has been completed or not and whether "label registration" has been completed or not. "Slit registration" refers to information such as the length of the original roll and the width of the slit processing. "Label registration" refers to the judgment and correction work of defective classifications. In this embodiment, as shown in FIG. 7, when performing label registration, the display control unit 13 displays the captured image G and the product attributes (defect classification such as "Defect A") as thumbnails on the operator terminal 6 (left side of FIG. 7), and displays a list of judgment probabilities of defective classifications in parallel with the thumbnail display (right side of FIG. 7). When the captured image G in this thumbnail display is clicked, the list of judgment probabilities of defective classifications is displayed. As a result, since the list of judgment probabilities of defect classifications is displayed in parallel with the thumbnail display, even an unskilled worker with specialized knowledge can easily check the accuracy of the defect classifications, thereby improving the work efficiency of the defect classification correction work.

[0046] Next, the pass / fail judgment unit 2 judges whether the worker has completed the judgment and correction work of the defective classification for each part number of the lot number based on the status information of the above-mentioned "label registration" (re-learning judgment process). When the worker has completed the judgment work (#38), the re-learning unit 25 executes re-learning of the trained model (#39, re-learning process). As shown in FIG. 4A, the re-learning unit 25 reads out the captured image G and its defective classification from the storage unit 12 when a certain amount of data is accumulated or at predetermined intervals, and reconstructs the trained model by deep learning using well-known deep metric learning or the like. As a result, even if defective data (defective) and normal data are mixed overall as shown in the upper side of FIG. 4B, as a result of re-learning the defective data with the normal data, the variation between the defective data (defective) and normal data is reduced as shown in the lower side of FIG. 4B, and the judgment accuracy of the trained model is improved.

[0047] The display control unit 13 in this embodiment can display the performance of the current trained model as shown in Fig. 8. As described above, since the threshold is set in the first determination unit 22, the probability that the captured image G is determined to be normal despite being defective is 0%. This makes it possible for the first determination unit 22 to determine that the captured image G is normal if the degree of abnormality is equal to or lower than the predetermined threshold, and to neither transmit anything to the operator terminal 6 nor notify that the captured image is normal.

[0048] The "false positive rate (Precision)" is true positive / (true positive+false positive), the "recall rate" is true positive / (true positive+false negative), and the "overall accuracy (Accuracy)" is (true positive+true negative) / total number. In the example shown in FIG. 8, the "false positive rate (Precision)" is 75%, the "recall rate" is 66%, and the "overall accuracy (Accuracy)" is 96%. In the example shown in FIG. 8, the "recall rate" is 66%, but in this embodiment, since the "recall rate" is a binary classification of normal / fail, the normal rate is 100%, and it can be seen that the probability that the first judgment unit 22 judges a defective item as normal is 0%. If a target value of the trained model (for example, an overall accuracy of 95% or more) is set using these indices, overlearning by the re-learning unit 25 can also be prevented.

[0049] 9, the display control unit 13 in this embodiment can map the group of features of normal products excluded by the first judgment unit 22 and the group of features of products clustered by the second judgment unit 23 for normal and multiple defective classifications. This makes it possible to determine whether the captured image G used for relearning is suitable, for example, by excluding it from the relearning target when it is relearned as defect A but is located near defect B in the two-dimensional space.

[0050] [Other embodiments] (a) In the above embodiment, the example of the inspection object is a nonwoven fabric, but the inspection object may be any product that requires inspection.

[0051] (b) Although the above-described re-learning unit 25 executes re-learning by batch processing, it may execute re-learning every time the input classification differs from the failure classification.

[0052] (c) In the above-described embodiment, when the second judgment unit 23 judges that the product is normal, the probability of normality and the captured image G are displayed on the worker terminal 6, but it is not necessary to display anything.

[0053] (d) In the above-described embodiment, when the first judgment unit 22 judges that the product is defective, the display control unit 13 causes the captured image G and the product attributes to be displayed as thumbnails on the operator terminal 6, and also causes a list of judgment probabilities of defective classification to be displayed in parallel with the thumbnail display. Alternatively, the thumbnail display may be omitted, or the list of judgment probabilities of defective classification may be displayed on a separate screen when the thumbnail-displayed captured image G is clicked.

[0054] (e) The browser in the worker terminal 6 in the above embodiment is merely an example, and any suitable browser can be used without departing from the gist of the present invention. [Industrial Applicability]

[0055] The present invention can be used in an inspection system used in a production line in a factory or the like. [Explanation of symbols]

[0056] 2: Good / bad judgement section 6: Worker terminal 11: Image acquisition unit 12: Storage section 13: Display control section 22:First judgment part 23:Second judgment section G: Captured image X: AI Support System

Claims

1. An image acquisition unit that acquires an image of an inspection object; a quality determination unit that determines quality of the inspection object based on the captured image; A display control unit that controls a display screen of the worker terminal; a storage unit that stores the captured image in which the inspection object is determined to be defective by the quality determination unit, the quality determination unit includes a first determination unit that excludes the captured images in which the inspection object is determined to be normal, and a second determination unit that performs clustering of the inspection object based on the captured images in which the inspection object is determined to be defective by the first determination unit, The display control unit is an AI support system that displays the captured images clustered by the second judgment unit and the defect classification of the inspection object on the worker terminal.

2. the storage unit stores the defect classification for the captured image in association with the captured image, 2. The AI ​​support system according to claim 1, wherein when an input classification input from the operator terminal differs from the defective classification, the pass / fail judgment unit re-learns based on the input classification.

3. 3. The AI ​​support system according to claim 1 or 2, wherein the display control unit, when the first judgment unit judges that the inspection object is normal, does not display anything on the operator terminal, and, when the second judgment unit judges that the inspection object is normal, displays the probability of normality and the captured image on the operator terminal.

4. the storage unit stores the defect classification for the captured image in association with the captured image, 3. The AI ​​support system according to claim 1 or 2, wherein when the first judgment unit judges that the inspection object is defective, the display control unit causes the captured image and attributes of the inspection object to be displayed as thumbnails on the operator terminal, and also causes a list of judgment probabilities of the defective classification to be displayed in parallel with the thumbnail display.

5. the storage unit stores the defect classification for the captured image in association with the captured image, The quality determination unit calculates a feature amount of the captured image of the inspection object determined to be defective by the first determination unit, The AI ​​support system according to claim 1 or 2, wherein the display control unit displays a collection of the feature amounts for each of the defect classifications in a two-dimensional space.

Images