Method of operating a computing device for defect inspection and system for inspecting defects

Abstract

This invention discloses a method for operating a computing device for defect inspection and a system for inspecting defects. When using a deep learning-based classification model to determine whether a product is good or defective based on an image of the product, this embodiment provides a defect inspection device and method for providing a basis for determining good / defective products by using a deep learning-based classification model with interpretable artificial intelligence (XAI), generating a category set for use as a basis, and continuously updating the parameters of the deep learning-based classification model using the category set.

Description

[0001] Cross-reference to related applications

[0002] This application claims the benefit of Korean Patent Application No. 10-2021-0033928, filed on March 16, 2021, with the Korean Intellectual Property Office, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0003] This disclosure relates to a defect inspection apparatus and method based on Explainable Artificial Intelligence (XAI). More specifically, when determining good / defective products based on product images using a deep learning-based classification model, this disclosure relates to a defect inspection apparatus and method that provides a basis for determining good / defective products by using a deep learning-based classification model based on Explainable Artificial Intelligence (XAI), generating a category set for the determination basis, and continuously updating the parameters of the deep learning-based classification model using the category set. Background Technology

[0004] The statements in this section provide background information in connection with this disclosure only and do not necessarily constitute prior art.

[0005] Compared to rule-based defect inspection methods, deep learning-based defect inspection methods have the advantage of improving quality management by performing more complex product inspections. However, when deep learning technology is applied to practical processing, management supervision may struggle to surpass existing rule-based machine vision techniques. This is because deep learning-based defect inspection methods may not provide clear criteria for determining whether a product is good or defective. Therefore, when defects occur, it may be difficult not only to analyze the cause but also to check whether the algorithm is functioning correctly. While deep learning algorithms have proven to have good performance, this issue remains a technological barrier when applied to practical processing.

[0006] Furthermore, while deep learning-based machine vision defect inspection methods offer various criteria for determining good / defective, they only provide two outcomes: good or defective. Due to environmental changes or processing problems, the parameters of the deep learning algorithm need to be updated. If these updates are not timely because they only provide definitive results, the deep learning algorithm may cause subsequent misjudgments of the product. The low unadjusted ratio resulting from such continuous misjudgments can ultimately cause serious damage in production processing. Therefore, considering the use and maintenance of defect inspection methods, solutions are needed to present interpretable criteria for the definitive results of the deep learning algorithm, as well as solutions for updating the deep learning algorithm based on these interpretable criteria. Summary of the Invention

[0007] When using a deep learning-based classification model to determine good / defective products based on product images, this disclosure aims to provide a defect inspection apparatus and method for providing a basis for determining good / defective products by using an interpretable artificial intelligence (XAI)-based deep learning classification model, by generating a category set for the determination basis, and by continuously updating the parameters of the deep learning-based classification model using the category set, enabling the deep learning-based classification model to adapt to environmental changes.

[0008] According to one aspect of this disclosure, a method for operating a computing device for defect inspection is provided, comprising: acquiring an image of a product to be inspected; generating a determination result indicating whether the product is good or defective based on the acquired image using a first classification model based on deep learning, and providing a determination basis for the determination result, wherein the first classification model has one or more parameters identical to those of a second classification model, the second classification model being pre-trained or stored in a data storage after being updated; and performing an adaptation process on the first classification model when the determination basis is determined to be an outlier, wherein performing the adaptation process comprises: updating a set of categories comprising multiple categories as the determination basis; and training the first classification model using the second classification model and the updated set of categories.

[0009] According to another aspect of this disclosure, an apparatus for defect inspection is provided, comprising: an input unit configured to acquire an image of a product to be inspected; a product inspection unit configured to use a deep learning-based first classification model to generate a determination result indicating whether the product is good or defective based on the acquired image, and to provide a determination basis for the determination result, wherein the first classification model has one or more parameters identical to those of a second classification model, the second classification model being pre-trained or stored in a data storage after being updated; and an adaptation unit configured to perform an adaptation process on the first classification model when the determination basis is determined to be an outlier, wherein the adaptation unit comprises: a category set generation unit configured to update a category set including multiple categories as the determination basis; and a training unit configured to train the first classification model using the second classification model and the updated category set.

[0010] According to another aspect of this disclosure, a non-transitory computer-readable recording medium having instructions stored thereon is provided, wherein the instructions cause a computer to perform: acquiring an image of a product to be inspected; using a first classification model based on deep learning, generating a determination result indicating whether the product is good or defective based on the acquired image, and providing a basis for the determination result, wherein the first classification model has one or more parameters identical to those of a second classification model, the second classification model being pre-trained or stored in a data storage device after being updated; and when the basis for determination is determined to be an outlier, performing an adaptation process on the first classification model, wherein performing the adaptation process includes: updating a set of categories comprising multiple categories as the basis for determination; and training the first classification model using the second classification model and the updated set of categories.

[0011] As described above, according to this embodiment, a defect inspection apparatus and method are provided. The apparatus and method are used to provide a basis for determining good / defective products using a deep learning-based classification model with interpretable artificial intelligence, generate a category set for the basis, and continuously update the parameters of the deep learning-based classification model using the category set. Therefore, considering the utilization and maintenance of the deep learning algorithm, it is possible to adapt in real time to the causes of defects caused by environmental changes and the effectiveness of flexible countermeasures against new types of defects.

[0012] Furthermore, according to this embodiment, a defect inspection apparatus and method are provided, which are used to provide a basis for determining good / defective products using a deep learning-based classification model with interpretable artificial intelligence, and thus have the effect that new proprietary technology associated with production processes can be accumulated by capturing specific potential defect elements that give rise to the basis for determining good / defective within an image. Attached Figure Description

[0013] Figure 1 This is a conceptual block diagram of a defect inspection apparatus according to an embodiment of the present disclosure.

[0014] Figure 2 This is a block diagram conceptually illustrating a classification model according to an embodiment of the present disclosure.

[0015] Figure 3 This is a flowchart of a defect inspection method according to an embodiment of the present disclosure.

[0016] Figure 4 This is a conceptual illustration of a latent space matching process for a classification model according to an embodiment of this disclosure.

[0017] Figure 5 This is a conceptual illustration of an example diagram showing the fine-tuning process of a classification model according to an embodiment of this disclosure. Detailed Implementation

[0018] In the following description, some embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, although elements are shown in different drawings, the same reference numerals preferably indicate the same elements. Furthermore, in the following description of some embodiments, for clarity and brevity, detailed descriptions of known functions and configurations incorporated herein will be omitted.

[0019] Furthermore, various terms (e.g., first, second, A, B, (a), (b), etc.) are used only to distinguish one component from other components and do not imply or suggest the material, order, or sequence of the components. Throughout the specification, when a section “includes” or “comprises” a component, unless otherwise specifically stated, that section is intended to further include other components, but does not exclude them. Additionally, terms such as “unit” or “module” refer to one or more units for performing at least one function or operation, which can be implemented by hardware, software, or a combination thereof.

[0020] The following will be related to the appendix Figure 1 The detailed description provided is intended to describe exemplary embodiments of the invention and is not intended to represent the only embodiments in which the invention can be performed.

[0021] This embodiment discloses a description of an AI-based defect inspection apparatus and method. More specifically, when determining whether a product is good or defective based on an image of the product using a deep learning-based classification model, this embodiment provides a defect inspection apparatus and method for providing a basis for determining good / defective products by using an interpretable artificial intelligence (XAI)-based deep learning classification model, generating a category set for use as a basis, and continuously updating the parameters of the deep learning-based classification model using the category set.

[0022] In the following description of this disclosure, the classification model represents the same deep learning-based neural network as the first classification model. Meanwhile, the second classification model is a neural network with the same structure as the first classification model and is used to back up the first classification model. The second classification model can also be used in the processing of updating the parameters of the first classification model.

[0023] Figure 1 This is a conceptual block diagram of a defect inspection apparatus according to an embodiment of the present disclosure.

[0024] According to this embodiment, the defect inspection device 100 uses a deep learning-based classification model to determine whether a product is good or defective based on an image of the product. Furthermore, the defect inspection device 100 uses interpretable AI (XAI) to provide criteria for determining good / defective products provided by the classification model, generates a category set for determining the criteria, and continuously updates the parameters of the classification model using the category set. The defect inspection device 100 includes all or some of the following: input unit 110, product inspection unit 120, and category set generation unit 130. Here, the components included in the defect inspection device 100 according to this embodiment are not necessarily limited to these. For example, the defect inspection device 100 may further include a training unit (not shown) for pre-training and training the classification model, or it may be implemented in a form that interoperates with an external training unit.

[0025] Figure 1 The configuration shown is an exemplary configuration according to this embodiment, and various implementations are possible, including different components or different connections between components depending on the input type, the type of determination basis based on XAI, the structure of the classification model, the category set generation method, etc.

[0026] Input unit 110 acquires images of the product from a machine vision camera. Here, the images are acquired by capturing the target product to be inspected. For example, when inspecting a printed circuit board (PCB) included in the target product, an image is acquired by capturing the PCB. When inspecting the appearance of the target product, an image can be acquired by capturing the appearance of the product. Furthermore, when the target product includes a display, an image of the display can be captured by the machine vision camera.

[0027] Product inspection unit 120 uses a deep learning-based classification model (first classification model) to generate a result determining whether a product is good or defective based on an image. Furthermore, the classification model generates XAI-based determination criteria for the determination results. These criteria can be represented as intuitively recognizable visual results. Product inspection unit 120 can provide the determination criteria in the form of a heatmap on the image used as input.

[0028] like Figure 2 As shown, the classification model (also known as the second classification model) includes an encoder 210, a fully connected layer 220, and an XAI 230. Here, the encoder 210 performs the function of extracting feature maps from the image of the product, and can be implemented using, for example, a convolutional neural network (CNN), which is a neural network known to be suitable for image processing. The feature maps extracted and finally output by the encoder 210 are represented as latent vectors.

[0029] The fully connected layer 220 generates a result that determines whether the product is good or defective based on the latent vectors that are the output of the encoder 210.

[0030] Meanwhile, XAI230 uses latent vector generation to determine the corresponding criteria for the results. XAI230 is an algorithm with fixed parameters. As such an XAI algorithm, response map-based Class Activation Graph (CAM) and Attention Branch Network (ABN) algorithms can be used, as well as gradient-weighted CAM (GradCAM) and GradCAM++ algorithms based on saliency masks.

[0031] The training unit according to this embodiment pre-trains the second classification model using the seen data for learning and the corresponding target determination results. Here, the seen data refers to the images previously protected for pre-training.

[0032] The training unit defines a loss function based on the distance between the determination result inferred by the second classification model from the seen data and the target determination result, and performs pre-training by updating the parameters of the second classification model in the direction that reduces the loss function. Here, the distance (e.g., L1 and L2 metrics) can be any distance that can represent the distance difference between two compared objects.

[0033] Meanwhile, during the pre-training process, the training unit can update the parameters of all components (encoder 210 and fully connected layer 220) of the second classification model.

[0034] After pre-training the second classification model, the category set generation unit 130 initializes the category set by classifying the heatmap, which serves as the basis for determination generated from the observed data by the second classification model, and generating multiple categories for determining good / defective results. K-means, K-nearest neighbors (kNN), Gaussian mixture models (GMM), etc., can be used as classification algorithms for classifying the basis for determination. The classification algorithm can perform classification on the basis for determination by grouping similar basis for determinations based on the similarity between them. Here, Euclidean distance, Mahalanobis distance, density function, etc., can be used as similarity.

[0035] A category set can be represented as a sum of sets, which are the sets of categories corresponding to good results and the sets of categories corresponding to defective results. The category set includes a representative heatmap (or representative vector) for each included category. Representative heatmaps can be generated differently depending on the classification algorithm used to categorize the criteria and the similarity score.

[0036] Product inspection unit 120 copies the parameters of the second classification model, which is pre-trained or updated and stored in the data storage, to the classification model, and then uses the classification model to perform defect inspection.

[0037] When an outlier is found in the determination criteria, the product inspection unit 120 sends the corresponding determination criteria and determination result to the category set generation unit 130. Here, when the similarity difference between any determination criteria for each category included in each category set and the representative heatmap is greater than a preset reference value, the arbitrary determination criteria is represented as an outlier.

[0038] At the same time, outliers may occur regardless of whether the result is determined to be good or defective.

[0039] When an outlier is found in the determination criteria, the category set generation unit 130 can update the category set by classifying the outlier as a new determination criterion. The updated category set can be used to check for the occurrence of outliers in the determination criteria during subsequent defect inspection processing of the next product.

[0040] The category set generation unit 130 classifies an image as unseen data, which causes the classification model to generate outliers and corresponding determination results. Furthermore, the category set generation unit 130 can include the determination results corresponding to the outliers in the target determination results used for training, thereby updating the target determination results. The category set generation unit 130 provides unseen data, updated target determination results, etc., to the training unit.

[0041] The training unit can use (i) the stored seen data and the second classification model, and (ii) the unseen data and the updated target determination results to train the classification model. The training process of the classification model will be described below.

[0042] Figure 3 This is a flowchart of a defect inspection method according to an embodiment of the present disclosure.

[0043] The defect inspection device 100 acquires an image of the product to be inspected (S300).

[0044] The defect inspection device 100 generates a result determining whether a product is good or defective based on an image using a classification model, and provides a determination basis based on XAI for the determination result (S302). Here, the classification model is generated by copying the parameters of a pre-trained second classification model. Alternatively, the classification model has the same parameters as the second classification model stored in the data storage after the update. As described above, the training unit of the defect inspection device 100 can perform pre-training on the second classification model based on the observed data used for learning and the corresponding target determination result.

[0045] The defect inspection device 100 uses a category set to determine whether the determination criterion for a product is an outlier (S304). Here, when the similarity difference between any determination criterion for each category included in the category set and the representative heatmap is greater than a preset reference value, the arbitrary determination criterion is represented as an outlier.

[0046] When it is determined that the criterion is not an outlier, the inspection process for that product is terminated, and the defect inspection for the next product can continue.

[0047] When the determination is based on an outlier, the defect inspection device 100 performs adaptive processing on the classification model.

[0048] In another embodiment of this disclosure, when a preset number or more products with outliers are accumulated, an adaptation process for the classification model can be performed.

[0049] The adaptation processes for the classification model (S306 and S308) are as follows.

[0050] The defect inspection device 100 updates the category set (S306). The defect inspection device 100 can update the category set by using outlier classification as a new determination criterion. The updated category set can be used to check for the occurrence of outliers in the determination criterion during subsequent defect inspection processing of the next product.

[0051] The defect inspection device 100 classifies images as unseen data, which causes the classification model to generate outliers and corresponding determination results. Furthermore, the defect inspection device 100 can update the target determination results by including good / defective determination results corresponding to the outliers in the target determination results used for training. The defect inspection device 100 provides the training unit with unseen data, updated target determination results, etc.

[0052] As described above, according to this embodiment, the defect inspection device 100 updates the existing category set in response to the occurrence of outliers, rather than generating a new category set. Therefore, the defect inspection device can operate continuously without interrupting the defect inspection process due to the generation of new data.

[0053] The defect inspection device 100 trains the classification model using a second classification model and an updated set of categories (S308).

[0054] The training process for the classification model includes latent space matching (S320) and fine-tuning (S322).

[0055] Fine-tuning of the classification model involves adapting it to unseen data. Depending on the application of the fine-tuning process using unseen data, the latent vectors generated by the encoder 210 of the classification model based on the seen data may be affected. Latent space matching can be performed to reduce this effect. Furthermore, when the influence on the latent vectors can be suppressed, it is not necessary to impose variability on the fully connected layer 220. Therefore, during the training process of the classification model, the parameters of the encoder 210 can be updated while the fully connected layer 220 within the classification model remains fixed.

[0056] Figure 4 This is a conceptual illustration of a latent space matching process for a classification model according to an embodiment of this disclosure.

[0057] In the latent space matching process, the training unit of the defect inspection device 100 trains a classification model for the seen data S0, ensuring that the latent vectors generated by both the classification model and the second classification model are consistent. Based on the difference between the determination criteria z1 generated by the classification model using the seen data and the determination criteria z1 generated by the second classification model using the same seen data, the training unit defines a loss function L. match (z1, z1). The training unit can move towards the loss function L. match The parameters of encoder 210 within the classification model are updated in the direction of reduction. As mentioned above, the parameters of the fully connected layer 220 within the classification model remain fixed.

[0058] Figure 5 This is a conceptual illustration of an example diagram showing the fine-tuning process of a classification model according to an embodiment of this disclosure.

[0059] In the fine-tuning process, the determination result (y1) generated by the classification model using unseen data (S1) is compared with the corresponding target determination result (y). t The difference between ) is used to define the loss function L for the training unit. class (y1,y t The training unit can move towards the loss function L. class The parameters of encoder 210 within the classification model are updated in the direction of reduction. As mentioned above, the parameters of the fully connected layer 220 within the classification model remain fixed.

[0060] The defect inspection device 100 checks whether the performance of the classification model meets the preset reference performance (S324). Here, the fact that the performance of the classification model meets the preset reference performance indicates, for example, that the loss function L match When the value is reduced to below the preset first reference value, the loss function L class Reduced to below the preset second reference value.

[0061] When the performance of the classification model does not meet the preset reference performance, the defect inspection device 100 repeatedly performs training processing on the classification model (S320 and S322).

[0062] When the performance of the classification model meets the preset reference performance, the defect inspection device 100 copies the parameters of the classification model to the second classification model and includes the unseen data in the seen data (S310). The defect inspection device 100 stores the second classification model and updates the seen data, thereby terminating the training of the classification model.

[0063] The device (not shown) equipped with the defect inspection apparatus 100 according to this embodiment may be a programmable computer and includes at least one communication interface that can be connected to a server (not shown).

[0064] The computing power of the device equipped with the defect inspection device 100 can be used to perform the training of the classification model as described above within the device equipped with the defect inspection device 100. Alternatively, the training of the classification model can be performed on a server.

[0065] As described above, according to this embodiment, a defect inspection apparatus and method are provided to provide a basis for determining good / defective products using a deep learning-based classification model with interpretable artificial intelligence. Taking into account the utilization and maintenance of deep learning algorithms, the apparatus and method can adapt in real time to the causes of defects caused by environmental changes and the effectiveness of flexible countermeasures against new types of defects.

[0066] Each flowchart according to this embodiment is described as a sequential execution of processes, but is not necessarily limited to this. In other words, the flowcharts are not limited to a time-series order, as the processes described in the flowcharts can be modified and executed, or one or more processes can be executed in parallel.

[0067] Various implementations of the systems and techniques described herein can be implemented using digital electronic circuits, integrated circuits, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These different implementations may include implementation by one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor (which may be a dedicated processor or a general-purpose processor) coupled to receive and send data and commands to and from a storage system, at least one input device, and at least one output device. The computer program (also referred to as a program, software, software application, or code) includes instructions for the programmable processor and is stored in a computer-readable recording medium.

[0068] Computer-readable recording media include various recording devices that store data readable by a computer system. These media can be non-volatile or non-transitory media, such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, or storage devices. Furthermore, computer-readable recording media can be distributed across computer systems connected via a network, enabling distributed storage and execution of computer-readable code.

[0069] Although exemplary embodiments of this disclosure have been described for illustrative purposes, those skilled in the art will understand that various modifications, additions, and substitutions can be made without departing from the spirit and scope of the claimed invention. Therefore, exemplary embodiments of this disclosure have been described for the sake of brevity and clarity. The scope of the technical concept of these embodiments is not limited to the illustrations. Therefore, those skilled in the art will understand that the scope of the claimed invention is not limited to the embodiments explicitly described above, but is limited by the claims and their equivalents.

Claims

1. A method of operating a computing device for defect inspection, comprising the following steps: Obtain an image of the product to be inspected; Based on the acquired image, a first classification model based on deep learning is used to generate a determination result indicating whether the product is good or defective, and the corresponding determination basis. Wherein, the first classification model has one or more parameters that are the same as those of the second classification model, and the second classification model is pre-trained or stored in a data storage after being updated; The determination criterion is used to determine whether it is an outlier, wherein the category set includes multiple categories for the determination criterion; and When the determination criterion is determined to be the outlier, an adaptation process is performed on the first classification model. The adaptation process includes: Update the category set based on the outliers; and The first classification model is trained using the second classification model and the updated category set. The first classification model includes an encoder, a fully connected layer, and interpretable AI. The encoder extracts feature maps from the product image and outputs the feature maps as latent vectors. The fully connected layer generates a determination result based on the latent vectors output from the encoder, indicating whether the product is good or defective. The interpretable AI uses the latent vectors to generate a determination criterion corresponding to the determination result. The second classification model is pre-trained using the observed data and corresponding target determination results for training, and The category set is initialized by classifying the multiple categories based on the determination criteria generated by the second classification model using the observed data. Updating the category set includes: The image that caused the first classification model to generate the outlier and the corresponding determination result is classified as unseen data; and This ensures that the determination result corresponding to the outlier is included in the corresponding target determination result. Training the first classification model includes: Perform latent spatial matching using the observed data; and Fine-tuning was performed using the unseen data. Specifically, latent space matching and fine-tuning are repeatedly performed on the first classification model until the performance of the first classification model meets a preset reference performance. The latent space matching process includes updating the parameters of the encoder of the first classification model to reduce the first loss function while keeping the parameters of the fully connected layers of the first classification model fixed. The first loss function is defined based on the difference between the determination criteria generated by the first classification model using the seen data and the determination criteria generated by the second classification model using the seen data. The fine-tuning includes updating the parameters of the encoder of the first classification model to reduce the second loss function while keeping the parameters of the fully connected layer of the first classification model fixed. The second loss function is defined based on the difference between the determination result generated by the first classification model using the unseen data and the corresponding target determination result.

2. The method according to claim 1, wherein, When the difference between the determination criterion and the representative value of each category included in the category set is greater than a preset reference value, the determination criterion is determined to be an outlier.

3. The method according to claim 1, wherein, Training the first classification model includes: Copy one or more parameters of the first classification model to the second classification model to update the second classification model; and When the performance of the first classification model meets the preset reference performance, the unseen data is included in the seen data.

4. A system for inspecting defects, comprising: The input unit is configured to acquire an image of the product to be inspected; A product inspection unit is configured to generate a determination result indicating whether the product is good or defective, and a corresponding determination criterion, based on the acquired image and a first classification model based on deep learning. The first classification model has one or more parameters identical to those of a second classification model, and the second classification model is pre-trained or stored in a data storage device after being updated. An adaptation unit is configured to determine whether the determination criterion is an outlier using a category set, wherein the category set includes multiple categories for the determination criterion, and when the determination criterion is determined to be an outlier, to perform adaptation processing on the first classification model. The adaptation unit includes: The category set generation unit is configured to update the category set based on the outliers; and The training unit is configured to train the first classification model using the second classification model and the updated category set. The first classification model includes an encoder, a fully connected layer, and interpretable AI. The encoder extracts feature maps from the product image and outputs the feature maps as latent vectors. The fully connected layer generates a determination result based on the latent vectors output from the encoder, indicating whether the product is good or defective. The interpretable AI uses the latent vectors to generate a determination criterion corresponding to the determination result. The training unit is configured to pre-train the second classification model using the seen data for training and the corresponding target determination results, and The category set generation unit is configured to classify the image that causes the first classification model to generate the outlier and the corresponding determination result as unseen data; and to include the determination result corresponding to the outlier in the corresponding target determination result. The training unit is configured to perform latent space matching using the seen data and fine-tuning using the unseen data, wherein latent space matching and fine-tuning are repeatedly performed on the first classification model until the performance of the first classification model meets a preset reference performance. The training unit is configured to update the parameters of the encoder of the first classification model to reduce the first loss function while keeping the parameters of the fully connected layers of the first classification model fixed. The first loss function is defined based on the difference between the determination criteria generated by the first classification model using the seen data and the determination criteria generated by the second classification model using the seen data. The training unit is configured to update the parameters of the encoder of the first classification model to reduce the second loss function while keeping the parameters of the fully connected layer of the first classification model fixed. The second loss function is defined based on the difference between the determination result generated by the first classification model using the unseen data and the corresponding target determination result.

5. The system according to claim 4, wherein: The second classification model includes an encoder that generates latent vectors from the image, interpretable artificial intelligence that generates the determination criteria from the latent vectors, and a fully connected layer that generates the determination result from the latent vectors.

6. The system according to claim 4, wherein, The category set generation unit is configured to initialize the category set by classifying the multiple categories on which the determination basis is generated by the second classification model using the seen data used to perform the pre-training.

7. A non-transitory computer-readable recording medium storing instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 3.

Images