Anomaly estimation system

Abstract

To provide an abnormality estimation system that accurately estimates abnormalities in a film-like object.SOLUTION: In an abnormality estimation system 1, an estimation device 2 includes: a machine learning unit 10 that uses a plurality of data sets of captured image data of a film-like object and classification data classified by classification labels related to abnormalities of the film-like object, machine learns each data set as teacher data, and generates a plurality of learning models; a photographing unit 21 of a photographing device 3 that photographs the film-like object; and an estimation unit 12 that estimates abnormalities of the film-like object from captured image data by the photographing unit 21 using the plurality of learning models. The plurality of learning models are classified into a plurality of classes for each classification label, and used to estimate abnormalities from the training models in higher classes in turn.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an abnormality estimation system for estimating an abnormality in a film-like material such as a polymer film. [Background Art]

[0002] In recent years, with the development of artificial intelligence, defect inspection systems that identify types of defects in film products by machine learning are known (for example, Patent Document 1). Patent Document 1 describes that the type of defect can be identified by machine learning that uses any one of total luminance, average luminance, median luminance, luminance variance, luminance gradient direction, magnitude of luminance gradient, defect area, defect perimeter, defect circularity, Feret diameter of defect, and aspect ratio of defect as a feature quantity. [Prior Art Documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Unexamined Patent Publication No. 2017-215277 [Summary of the Invention] [Problem to be Solved by the Invention]

[0004] However, there are many types of film defects, and among film defects, there are defects having very similar shapes. Therefore, when identifying the type of defect from a captured image using a single learning model as in the defect inspection system of Patent Document 1, there is a limit to accuracy, and there is a problem that the type of defect cannot be identified with high accuracy.

[0005] Accordingly, an object of the present invention is to provide an abnormality estimation system capable of estimating an abnormality with higher accuracy than conventional systems. [Means for Solving the Problem]

[0006] One aspect of the present invention for solving the above-mentioned problems is an anomaly estimation system for estimating abnormalities in a film-like material, comprising: a machine learning unit that uses multiple datasets of captured image data of the film-like material and classification data classified by classification labels relating to abnormalities in the film-like material, performs machine learning on each dataset as training data to generate multiple learning models; an imaging unit that photographs the film-like material; and an estimation unit that uses the multiple learning models to estimate abnormalities in the film-like material from the captured image data by the imaging unit, wherein the multiple learning models are classified into multiple classes according to the classification labels, and are used to estimate abnormalities in order from the learning models of the higher classes.

[0007] According to this approach, anomalies are estimated individually by each learning model, allowing for more accurate anomaly estimation compared to conventional methods. According to this pattern, if an anomaly is detected in a higher class, there is no need to make an anomaly judgment in a lower class, thus reducing computational complexity and improving accuracy.

[0008] A desirable outcome is that, in the lower-level learning model, captured image data that would be classified as a higher-level classification label is excluded from the training data.

[0009] According to this approach, extraneous learning can be eliminated, reducing computational complexity and improving accuracy.

[0010] Here, among the defects in film-like materials, there are serious defects that, if present, would render the product unusable. The inventors investigated these serious defects and discovered that those that lead to incorrect classifications of serious defects are often classified as specific defects.

[0011] Therefore, a preferred configuration is one in which the learning model classifies by comparing the features calculated from the captured image data with a threshold, the classification labels include important labels and false positive warning labels, the class to which the important label belongs is a higher class than the class to which the false positive warning label belongs, the machine learning unit can generate a rescue learning model by machine learning using a dataset of captured image data of the film-like object and classification data classified by at least the important labels and the false positive warning labels as training data, the rescue learning model has a smaller threshold than the learning model corresponding to the important label, and the estimation unit, when classified as an anomaly corresponding to the false positive warning label in the learning model corresponding to the false positive warning label, classifies it based on the rescue learning model into an anomaly corresponding to the important label and an anomaly not corresponding to the important label.

[0012] According to this approach, if a misidentification warning label that is prone to being misidentified is classified as such in the learning model, the corrective learning model reclassifies it as an important label, thus enabling more accurate classification of important labels.

[0013] Here, among the abnormalities in the film-like material, there are some that can be classified with nearly 100% accuracy. Therefore, the inventor considered reducing the computational load by placing these abnormalities with extremely high classification accuracy into a higher class.

[0014] Therefore, a preferred configuration is one in which the classification labels include important labels and high-precision labels, wherein the high-precision labels are those in which the corresponding learning model has an accuracy rate of 99.9% or higher and are classified into a higher class than the important labels.

[0015] According to this pattern, highly accurate labels with extremely high accuracy rates are classified before important labels, allowing for more accurate estimation of anomalies.

[0016] In a preferred aspect, the estimating unit performs classification by the learning models in order starting from the learning model of an upper class, and estimates, when a positive result is obtained, the abnormality corresponding to the positive learning model as the abnormality of the film-like material.

[0017] According to this aspect, since the abnormality is estimated when a positive result is obtained, the abnormality can be estimated more efficiently. [Effects of the Invention]

[0018] According to the abnormality estimation system of the present invention, an abnormality can be estimated with higher accuracy than in the conventional art. [Brief Description of Drawings]

[0019] [Figure 1] It is an explanatory diagram of the abnormality estimation system according to the first embodiment of the present invention, wherein (a) is a block diagram of the abnormality estimation system, and (b) is a side diagram schematically showing the imaging device. [Figure 2] It is an explanatory diagram of the abnormality estimation program according to the first embodiment, wherein (a) is an explanatory diagram during learning of each classification model in the abnormality estimation program, and (b) is an explanatory diagram of a multi-stage classification model. [Figure 3] It is an explanatory diagram of the multi-stage classification model in the abnormality estimation program according to the first embodiment, wherein important labels are surrounded by thick lines. [Figure 4] It is an explanatory diagram of the multi-stage classification model in the abnormality estimation program according to the second embodiment, wherein important labels are surrounded by thick lines. [Mode for Carrying Out the Invention]

[0020] Hereinafter, embodiments of the present invention will be described in detail.

[0021] The abnormality estimation system 1 according to the first embodiment of the present invention uses an abnormality estimation program to perform machine learning as supervised training for each data set using a plurality of data sets each consisting of captured image data of a past film-like material 100 and classification data classified by a classification label relating to an abnormality of the film-like material 100, thereby generating a plurality of learning models. Furthermore, the abnormality estimation system 1 estimates an abnormality of the film-like material 100 by serially connecting the respective learning models in a decision tree-like manner from a current captured image by means of the abnormality estimation program, performing classification with each learning model in sequence.

[0022] As shown in Fig. 1(a), the abnormality estimation system 1 is configured such that, as main components, an estimation device 2 and an imaging device 3 are connected by wire or wirelessly via a network 5 such as the Internet or an intranet. The estimation device 2 is a computer whose hardware configuration includes: a central processing unit composed of a control device that controls each device and an arithmetic device that performs computation on data; a storage device that stores data; an input device that inputs data from the outside; and an output device that outputs data to the outside. As shown in Fig. 1(a), the estimation device 2 includes, as main structural parts, a machine learning unit 10, a data storage unit 11, an estimation unit 12, and a communication unit 13.

[0023] The machine learning unit 10 has a function of performing learning via so-called supervised learning by means of the abnormality estimation program, and is capable of performing supervised learning in accordance with a machine learning algorithm such as a convolutional neural network. Here, "supervised learning" refers to a method in which a large number of sets of teacher data, that is, data of a certain input (explanatory variable) and a corresponding result (objective variable), are provided to the machine learning unit 10, so that the machine learning unit 10 learns features contained in these data sets and inductively acquires a model (error model) that estimates a result from an input, that is, the relationship between inputs and results. In other words, the machine learning unit 10 has an input unit and an output unit, as shown in Figure 2(a), and by performing machine learning based on training data, it is possible to construct a learning model that calculates the target variable output from the output unit from the explanatory variables input to the input unit.

[0024] The data storage unit 11 stores data such as past and present captured image data, learning models corresponding to each classification label, and the results of estimations performed by the past and present estimation unit 12. The data storage unit 11 stores an anomaly estimation program that causes the estimation device 2 to perform estimation operations.

[0025] The estimation unit 12 is the part that estimates abnormalities in the film-like object 100 from the captured image data using each learning model constructed by the machine learning unit 10.

[0026] The communication unit 13 is a part that can communicate with the network 5 directly or via a router or the like.

[0027] (Imaging device 3) As shown in Figure 1(a), the imaging device 3 mainly consists of a transport unit 20, an imaging unit 21, an illumination unit 22, and a communication unit 23. As shown in Figure 1(b), the conveying unit 20 includes a pair of conveying rollers 30 and 31, and conveys the film-like material 100 suspended on the conveying rollers 30 and 31 in the conveying direction. The imaging unit 21 is the part that photographs the film-like material 100 as it passes between the transport rollers 30 and 31 and generates image data. The lighting unit 22 is the part that irradiates light onto the film-like material 100 as it passes between the transport rollers 30 and 31. The communication unit 23 is a part that can communicate with the network 5 directly or via a router or the like.

[0028] Next, the anomaly estimation method using the anomaly estimation program of this embodiment will be described.

[0029] The anomaly estimation method of this embodiment mainly consists of a learning model creation step and an evaluation step, and uses the learning model created in the learning model creation step to estimate anomalies in the evaluation step.

[0030] (Training model creation process) The learning model creation process involves creating a learning model to be used in the evaluation process by performing machine learning based on captured image data and the classification results of that image data. Specifically, first, classification labels L1 to L16 are assigned to various abnormalities in the film-like material 100, and then the material is classified into multiple classes according to classification labels L1 to L16. Then, for each classification label L1 to L16, machine learning is performed using a dataset of past captured image data and classification data obtained by classifying abnormalities in the film-like material 100 according to classification labels L1 to L16 as training data, and a learning model is created for each, resulting in a multi-stage classification model as shown in Figure 3. In other words, in the learning model creation process, the machine learning unit 10 performs machine learning using the captured image data as explanatory variables and whether or not it corresponds to a classification label as the objective variable, and constructs each learning model. The constructed learning models are then connected in series as shown in Figure 2(b) to construct a multi-level classification model like a classification tree.

[0031] The classification labels for anomalies include important labels L2, L4, and L5, false positive warning labels L8 and L9, and high-precision labels L1, L3, L6, and L7. Important labels L2, L4, and L5 are labels that correspond to serious defects that would render the product unacceptable, such as insects, holes, or foaming. Misidentification warning labels L8 and L9 are important labels that are prone to being misidentified, such as "tea marks" and "condensation." High-precision labels L1, L2, L3, L6, and L7 are labels that can be classified with an accuracy of 99.9% or higher by the corresponding learning model, and include, for example, white spots, insects, glitter, infestation, and gel. The high-precision labels L1, L2, L3, L6, and L7 of this embodiment are labels that can be classified with 100% accuracy by the corresponding learning model, and even when multi-class classification is performed using all classification labels with a single learning model, the labels can be classified with 100% accuracy. Furthermore, classification label L2 is both an important label and a high-precision label.

[0032] Furthermore, in lower-level class learning models, it is preferable to exclude captured image data that is classified as a classification label in a higher-level class from the training data. In other words, it is preferable not to use training data that has been classified as corresponding to a classification label in a higher-level class as training data in a lower-level class. By doing so, it is possible to reduce combinations that are clearly incorrect answers and improve the efficiency of machine learning. Each learning model is a binary classification model that compares features calculated from captured image data with a judgment threshold to determine whether or not the data fits the classification model. In other words, each learning model determines that a feature corresponds to a classification label if its quantity is above a certain threshold, and that it does not correspond to a classification label if its quantity is below the threshold. The learning models include a basic learning model and a corrective learning model corresponding to the misidentification warning labels L8 and L9. These learning models preferably use the sigmoid function as the activation function. The classification threshold when using the sigmoid function can be designed as appropriate, but for the basic learning model, it is preferably 0.4 to 0.6, and for the rescue learning model, it is set lower than the basic learning model, to 1 / 100 or less of the classification threshold of the basic learning model. The classification threshold for the basic learning model in this embodiment is 0.5, and the classification threshold for the rescue learning model is 0.0001.

[0033] The machine learning algorithms used in the learning model creation process are not particularly limited as long as they are supervised learning algorithms capable of classification evaluation. Examples include linear regression models, logistic regression, lasso regression, random forests, neural networks, and linear kernel support vector machines (SVM (linear)), among which convolutional neural networks are preferred. The machine learning algorithm of this embodiment is a convolutional neural network, which includes convolutional layers, pooling layers, and fully connected layers. For convolutional neural networks, publicly known algorithms such as GoogLeNet can be used.

[0034] (Evaluation process) The evaluation process involves using the multi-stage classification model created in the learning model creation process to estimate and evaluate the type of anomaly from the image data captured by the imaging unit 21. In other words, as shown in Figure 2(b), the evaluation process involves inputting the captured images from the imaging unit 21 into the input unit of the multi-stage classification model, and the anomalies estimated through each learning model are output from the output unit of the multi-stage classification model.

[0035] Specifically, first, when the image captured by the imaging unit 21 is input to the input unit of the multi-stage classification model, step S1 determines whether the anomaly in the captured image corresponds to the high-precision classification label L1 (white dot).

[0036] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L1 (i.e., it is not a white spot) (No in step S1), then in step S2, it is determined whether the anomaly in the captured image corresponds to the important classification label L2 (insect).

[0037] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L2 (i.e., it is not an insect) (No in step S2), then in step S3, it is determined whether the anomaly in the captured image corresponds to the high-precision classification label L3 (Kira).

[0038] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L3 (i.e., it is not a glitter) (No in step S3), then in step S4, it is determined whether the anomaly in the captured image corresponds to the important classification label L4 (hole).

[0039] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L4 (i.e., it is not a hole) (No in step S4), then in step S5, it is determined whether the anomaly in the captured image corresponds to classification label L5 (foaming), which is the weight label.

[0040] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L5 (i.e., it is not foaming) (No in step S5), then in step S6, it is determined whether the anomaly in the captured image corresponds to the high-precision classification label L6 (intrusion).

[0041] If the anomaly feature quantity is below the judgment threshold and the abnormality does not correspond to classification label L6 (i.e., it is not an indentation) (No in step S6), then in step S7, it is determined whether the abnormality in the captured image corresponds to the high-precision classification label L7 (gel).

[0042] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L7 (i.e., it is not a gel) (No in step S7), then in step S8, it is determined whether the anomaly in the captured image corresponds to classification label L8 (brown dot), which is a false positive warning label.

[0043] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L8 (i.e., it is not a brown dot) (No in step S8), then in step S9, it is determined whether the anomaly in the captured image corresponds to classification label L9 (condensed), which is a false positive warning label.

[0044] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L9 (i.e., it is not condensed) (No in step S9), then in step S10, it is determined whether the anomaly in the captured image corresponds to classification label L10 (fiber).

[0045] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L10 (i.e., it is not a fiber) (No in step S10), then in step S11, it is determined whether the anomaly in the captured image corresponds to classification label L11 (film debris). Here, among the gel-like or oligomer lumps, those with low light transmittance, those with a refractive index different from the average refractive index of the main components of the film, and fine fragments of the film are grouped together and called film waste.

[0046] If the anomaly feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L11 (i.e., it is not film debris) (No in step S11), then in step S12, it is determined whether the anomaly in the captured image corresponds to classification label L12 (foreign object).

[0047] If the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L12 (i.e., it is not a foreign object) (No in step S12), it is determined that estimation is impossible, and the estimation result (inferred) is output from the output unit.

[0048] Furthermore, in step S1, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L1 (white dot) (Yes in step S1), the process proceeds to step S13 to determine whether it corresponds to classification label L13 (nucleus present). In step S13, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L13 (with nucleus) (Yes in step S13), the anomaly in the captured image is estimated to be a white dot with a nucleus, and the estimated result (white dot with nucleus) is output from the output unit. On the other hand, in step S13, if the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L13 (with nucleus) (No in step S13), the anomaly in the captured image is estimated to be a nucleusless white spot, and the estimated result (nucleusless white spot) is output from the output unit.

[0049] In step S2, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L2 (insect) (Yes in step S2), the anomaly in the captured image is estimated to be an insect, and the estimated result (insect) is output from the output unit.

[0050] In step S3, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L3 (Kira) (Yes in step S3), the anomaly in the captured image is estimated to be Kira, and the estimated result (Kira) is output from the output unit.

[0051] In step S4, if the feature quantity is above the judgment threshold and the anomaly corresponds to the classification label L4 (hole) (Yes in step S4), the anomaly in the captured image is estimated to be a hole, and the estimated result (hole) is output from the output unit.

[0052] In step S5, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L5 (foaming) (Yes in step S5), the anomaly in the captured image is estimated to be foaming, and the estimated result (foaming) is output from the output unit.

[0053] In step S6, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L6 (intrusion) (Yes in step S6), the anomaly in the captured image is estimated to be intrusion, and the estimated result (intrusion) is output from the output unit.

[0054] In step S7, if the feature is above the threshold and the anomaly corresponds to classification label L7 (gel) (Yes in step S7), the process proceeds to step S14 to determine whether it corresponds to classification label L14 (nucleus). In step S14, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L14 (with nucleus) (Yes in step S14), the anomaly in the captured image is estimated to be a gel with a nucleus, and the estimated result (gel with nucleus) is output from the output unit. On the other hand, in step S14, if the feature quantity is below the judgment threshold and the anomaly does not correspond to classification label L14 (with nucleus) (No in step S14), the anomaly in the captured image is estimated to be a nucleus-less gel, and the estimated result (nucleus-less gel) is output from the output unit.

[0055] In step S8, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L8 (brown dot), which is a false alarm label (Yes in step S8), the process moves to step S15, where a rescue learning model with a smaller judgment threshold determines whether it corresponds to classification label L15 (foam). In step S15, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L15 (foaming) (Yes in step S15), the anomaly in the captured image is estimated to be foaming corresponding to the important label, and the estimated result (foaming) is output from the output unit. On the other hand, in step S15, if the feature quantity is below the judgment threshold and the anomaly does not correspond to the classification label L15 (foaming) (No in step S15), the anomaly in the captured image is estimated to be a brown spot, and the estimated result (brown spot) is output from the output unit.

[0056] In step S9, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L9 (condensed), which is a false alarm label (Yes in step S9), the process moves to step S16, where a rescue learning model with a smaller judgment threshold determines whether it corresponds to classification label L16 (hole). In step S16, if the feature quantity is above the judgment threshold and the anomaly corresponds to the classification label L16 (hole) (Yes in step S16), the anomaly in the captured image is estimated to be a hole corresponding to the important label, and the estimated result (hole) is output from the output unit. On the other hand, in step S16, if the feature quantity is below the judgment threshold and the anomaly does not correspond to the classification label L16 (hole) (No in step S16), the anomaly in the captured image is estimated to be condensation, and the estimated result (condensation) is output from the output unit.

[0057] In step S10, if the feature quantity is above the judgment threshold and the anomaly corresponds to the classification label L10 (fiber) (Yes in step S10), the anomaly in the captured image is estimated to be a fiber, and the estimated result (fiber) is output from the output unit.

[0058] In step S11, if the feature quantity is above the judgment threshold and the anomaly corresponds to classification label L11 (film debris) (Yes in step S10), the anomaly in the captured image is estimated to be film debris, and the estimated result (film debris) is output from the output unit.

[0059] In step S12, if the feature quantity is above the judgment threshold and the anomaly corresponds to the classification label L12 (foreign object) (Yes in step S12), the anomaly in the captured image is estimated to be a foreign object, and the estimated result (foreign object) is output from the output unit.

[0060] According to the anomaly estimation system 1 of this embodiment, each learning model performs individual binary classification like a classification tree, so it can make a judgment even with small features compared to conventional methods, and can estimate anomalies with high accuracy.

[0061] According to the anomaly estimation system 1 of this embodiment, in the learning model for lower classes, captured image data that is classified into higher class classification labels is excluded from the training data, thereby eliminating unnecessary learning, reducing computational load, and improving accuracy. According to the anomaly estimation system 1 of this embodiment, when an anomaly is estimated in a higher class, it is not necessary to make an anomaly judgment in a lower class, thus reducing the amount of computation and improving accuracy.

[0062] According to the anomaly estimation system 1 of this embodiment, when a misidentification warning label that is prone to being misidentified is classified as a misidentification warning label in the learning model, the corrective learning model reclassifies it as an important label. This suppresses the omission of important labels that cannot be misidentified, and enables more accurate classification of important labels.

[0063] According to the anomaly estimation system 1 of this embodiment, high-precision labels L1, L2, and L3 (classification label L2 is both an important label and a high-precision label) with extremely high accuracy rates are classified before important labels L2, L4, and L5, so anomalies can be estimated with greater accuracy.

[0064] According to the anomaly estimation system 1 of this embodiment, since the anomaly is estimated at the time of a positive result (the time when it is determined that the condition applies), the anomaly can be estimated more efficiently.

[0065] Next, a method for estimating anomalies according to a second embodiment of the present invention will be described.

[0066] The anomaly estimation method of the second embodiment differs from the multi-stage classification model of the first embodiment in that the multi-stage classification model is different. In the second embodiment, as shown in Figure 4, there is no determination based on classification label L15 in step S15 and classification label L16 in step S16. If it corresponds to classification label L8 (brown dot) in step S8, it is automatically assumed to be a brown dot, and if it corresponds to classification label L9 (condensed) in step S9, it is automatically assumed to be condensed.

[0067] In the embodiments described above, the learning model corresponding to the important label used the same judgment threshold as the learning model for other classification labels, but the present invention is not limited thereto. The judgment threshold may be set lower in the learning model corresponding to the important label.

[0068] In the embodiments described above, binary classification was performed for each classification label, but the present invention is not limited thereto. Partial multi-class classification may also be performed. For example, multi-class classification may be performed for high-precision labels, and binary classification may be performed for the remaining classification labels.

[0069] In the embodiment described above, the classes to which the important labels belong and the classes to which the high-precision labels belong were arranged randomly, but the present invention is not limited thereto. The classes to which the important labels belong may be arranged together, and the classes to which the high-precision labels belong may be arranged together.

[0070] In the embodiments described above, the components can be freely substituted or added between each embodiment, as long as they fall within the technical scope of the present invention. [Examples]

[0071] The present invention will be described in detail below with reference to examples and comparative examples, but the present invention is not limited to these examples.

[0072] (Example 1) First, using a dataset of image data of film defects and classification labels, machine learning was performed for each classification label to generate individual learning models. A multi-level classification tree model, as shown in Figure 3, was created and designated as Example 1. Machine learning was performed using GoogLeNetwork with the trainNetwork function. Furthermore, the features of each trained model were normalized using the sigmoid function, and the judgment threshold for the rescue trained model was set to 0.0001, while the judgment thresholds for the other trained models were set to 0.5.

[0073] (Example 2) Example 2 was created in the same manner as Example 1, except that a multi-level classification model of the classification tree type as shown in Figure 4 was created. In other words, in Example 2, classification was performed without using a rescue learning model. Furthermore, the features of each learning model were normalized using a sigmoid function, and the decision threshold for each learning model was set to 0.5.

[0074] (Comparative Example 1) Using a dataset of image data of film defects and classification labels, a machine learning model was generated, which was designated as Comparative Example 1. Machine learning was performed using GoogLeNetwork with the trainNetwork function. Furthermore, the features of the trained model were normalized using the sigmoid function, and the judgment threshold of the trained model was set to 0.5.

[0075] (Model evaluation) For 744 captured image data images containing known defects, defects were estimated using the models from Examples 1 and 2 and Comparative Example 1, and the accuracy rate was calculated. In calculating the accuracy rate, each classification label was individually evaluated. A correct answer was given when the target classification label was correctly identified, and an incorrect answer was given when the target classification label was correctly identified. The accuracy rate was then calculated using the following formula (1).

[0076]

number

[0077] Table 1 shows the accuracy rate (%) for each classification label in Examples 1 and 2 and Comparative Example 1.

[0078] [Table 1]

[0079] Examples 1 and 2 showed a higher accuracy rate for each classification label compared to Comparative Example 1, which used multi-class classification, demonstrating a significant improvement in precision. Furthermore, although the accuracy rate for condensation in Example 1 was lower than in Example 2, the accuracy rates for the important labels of insects, holes, and foaming all reached 100%. [Explanation of Symbols]

[0080] 1. Anomaly Estimation System 10 Machine Learning Department 12 Estimation part 21 Photography Department 100 Film-like material

Claims

1. An anomaly estimation system for estimating abnormalities in film-like materials, A machine learning unit uses multiple datasets of image data of the film-like material and classification data classified by classification labels related to abnormalities in the film-like material, performs machine learning on each dataset as training data, and generates multiple learning models. A shooting unit for photographing the aforementioned film-like material, The system includes an estimation unit that uses the aforementioned multiple learning models to estimate abnormalities in the film-like material from the image data captured by the imaging unit. The aforementioned multiple learning models are classified into multiple classes according to the classification label, and are used for anomaly estimation in order from the learning models of the higher classes. An anomaly estimation system in which lower-level class learning models exclude captured image data that would be classified under higher-level class classification labels from the training data.

2. The aforementioned learning model classifies features calculated from the captured image data by comparing them with a threshold, The aforementioned classification labels include important labels and warning labels for misidentification. The class to which the aforementioned important label belongs is a higher class than the class to which the aforementioned misidentification warning label belongs. The machine learning unit can generate a rescue learning model by machine learning using a dataset of captured image data of the film-like object and classification data classified by at least the important label and the false recognition warning label as training data. The aforementioned rescue learning model has a smaller threshold than the learning model corresponding to the important label. The anomaly estimation system according to claim 1, wherein the estimation unit, when classified as an anomaly corresponding to the false alarm label in the learning model corresponding to the false alarm label, classifies it into an anomaly corresponding to the important label and an anomaly other than the important label based on the remediation learning model.

3. An anomaly estimation system for estimating an anomaly in a film-like material, A machine learning unit uses multiple datasets of image data of the film-like material and classification data classified by classification labels related to abnormalities in the film-like material, performs machine learning on each dataset as training data, and generates multiple learning models. A shooting unit for photographing the aforementioned film-like material, The system includes an estimation unit that uses the aforementioned multiple learning models to estimate abnormalities in the film-like material from the image data captured by the imaging unit. The aforementioned multiple learning models are classified into multiple classes according to the classification label, and are used for anomaly estimation in order from the learning models of the higher classes. The aforementioned learning model classifies features calculated from the captured image data by comparing them with a threshold, The aforementioned classification labels include important labels and warning labels for misidentification. The class to which the aforementioned important label belongs is a higher class than the class to which the aforementioned misidentification warning label belongs. The machine learning unit can generate a rescue learning model by machine learning using a dataset of captured image data of the film-like object and classification data classified by at least the important label and the false recognition warning label as training data. The aforementioned rescue learning model has a smaller threshold than the learning model corresponding to the important label. An anomaly estimation system in which, when an anomaly corresponding to the false alarm label is classified in the learning model corresponding to the false alarm label, the estimation unit classifies it into an anomaly corresponding to the important label and an anomaly not corresponding to the important label, based on the corrective learning model.

4. An anomaly estimation system for estimating an anomaly in a film-like material, A machine learning unit uses multiple datasets of image data of the film-like material and classification data classified by classification labels related to abnormalities in the film-like material, performs machine learning on each dataset as training data, and generates multiple learning models. A shooting unit for photographing the aforementioned film-like material, The system includes an estimation unit that uses the aforementioned multiple learning models to estimate abnormalities in the film-like material from the image data captured by the imaging unit. The aforementioned multiple learning models are classified into multiple classes according to the classification label, and are used for anomaly estimation in order from the learning models of the higher classes. The aforementioned classification labels include important labels and high-precision labels. The aforementioned high-precision label is an anomaly estimation system in which the accuracy of the corresponding learning model is 99.9% or higher, and the label is classified into a higher class than the aforementioned important label.

5. The abnormality estimation system according to any one of claims 1 to 4, wherein the estimation unit classifies the learning models in order from the highest class, and when a learning model is found to be positive, it estimates the abnormality corresponding to the positive learning model as an abnormality of the film-like material.

Images