Intelligent steel surface inspection device and method

Abstract

According to the present disclosure, an intelligent steel surface inspection device is provided, comprising: a data collection unit disposed in a steel manufacturing facility to collect steel surface data; a labeling unit that receives a label corresponding to the collected steel surface data and maps it to the steel surface data; and a surface data augmentation generation engine that learns a representation based on supervised contrastive learning using labeled steel surface data and augmented steel surface data, wherein the learning model infers unlabeled steel surface data and extracts some steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data.

Description

Technology Field

[0001] The embodiments of the present disclosure relate to an apparatus and method for non-destructively inspecting appearance defects on a steel surface, and more specifically, to an intelligent steel surface inspection apparatus and method that predicts good and defective steel surfaces from unlabeled data by labeling and retraining only data with high learning importance. Background Technology

[0002] The steel manufacturing process proceeds through entry, roll forming, semi-auto welding, sizing, cutting, end facing, and picking. During this stage, the steel surface is a factor directly linked to quality and yield, making it essential to inspect the product's appearance for any defects.

[0003] In conventional technology, inspection performance was dramatically improved by introducing a deep learning classification model to distinguish between good and defective steel surfaces, but training the deep learning model required a labeling task to tag correct answers on a large amount of training data, which required a lot of time and effort.

[0004] In particular, in the case of steel products, it was difficult to obtain sufficient surface data because they are manufactured in hazardous, high-temperature working environments; furthermore, since the surface defect area was only a very small portion relative to the total area, there was a significant data imbalance between good and defective product classes.

[0005] Therefore, the need has arisen for an invention that distinguishes good or defective steel surface data by using a learning model that distinguishes images similar to good or defective products using separately unlabeled data, and in particular by selectively extracting data of high importance for learning from the unlabeled data to generate training data for said learning model. Prior art literature

[0006] Registered Patent Publication, No. 10-2196114 (December 22, 2020) The problem to be solved

[0007] The embodiments of the present disclosure aim to solve various problems, including those mentioned above, by providing an intelligent steel surface inspection device and method that predicts good and defective steel surfaces from unlabeled data by labeling and retraining only data with high learning importance. However, these problems are exemplary and the scope of the present disclosure is not limited by them. means of solving the problem

[0008] According to one aspect of the present disclosure, an intelligent steel surface inspection device is provided, comprising: a data collection unit disposed in a steel manufacturing facility to collect steel surface data; a labeling unit that receives a label corresponding to the collected steel surface data and maps it to the steel surface data; and a surface data augmentation generation engine that learns a representation based on supervised contrastive learning using labeled steel surface data and augmented steel surface data, wherein the learning model infers unlabeled steel surface data and extracts some steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data.

[0009] According to the present embodiment, the labeling unit receives steel surface data having a similarity lower than the preset reference similarity, and receives a label corresponding to the steel surface data and maps it to the received steel surface data.

[0010] According to the present embodiment, the learning model can perform retraining using steel surface data having a similarity lower than the preset reference similarity to which the label is mapped.

[0011] According to the present embodiment, the inferred steel surface data includes positive samples and negative samples, and the preset reference similarity can be determined according to an additive margin softmax loss function to maintain the projection distance between the positive samples and the negative samples.

[0012] According to the present embodiment, steel surface data having a similarity lower than the preset reference similarity may be data for which it is not determined whether it is a good product or a defective product in vision inspection.

[0013] According to the present embodiment, the surface data augmentation generation engine can generate augmented steel surface data by data splitting, size adjustment, or color conversion for the input steel surface data.

[0014] According to the present embodiment, the learning model can be trained such that the vector similarity of each data is maximized in a direction in which the contrast loss function between the steel surface data and the augmented steel surface data is minimized.

[0015] According to the present embodiment, the data collection unit may include a shooting unit that photographs the steel surface and an IoT sensor unit that acquires information about the steel surface.

[0016] According to one aspect of the present disclosure, an intelligent steel surface inspection method is provided, comprising: a step of collecting steel surface data by being placed in a steel manufacturing facility; a step of receiving a label corresponding to the collected steel surface data and mapping it to the steel surface data; a step of learning a representation based on supervised contrastive learning for a learning model using the labeled steel surface data and the augmented steel surface data; and a step of the learned learning model inferring unlabeled steel surface data and extracting steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data.

[0017] According to the present embodiment, the method may further include the steps of: receiving steel surface data having a similarity lower than the preset reference similarity; receiving a label corresponding to the steel surface data and mapping it to the received steel surface data; and performing retraining using the steel surface data having a similarity lower than the preset reference similarity to which the label is mapped.

[0018] Other aspects, features, and advantages other than those described above will become clear from the following specific details, claims, and drawings for implementing the invention.

[0019] In addition, these general and specific aspects may be implemented using a system, method, computer program, or any combination of a system, method, or computer program. Effects of the invention

[0020] According to the exemplary embodiment of the present disclosure as described above, the problem of the representation vector being distorted due to a change in the representation space caused by data imbalance can be prevented, thereby enabling the implementation of an intelligent steel surface inspection device and method for predicting good and defective steel surfaces from unlabeled data. Of course, the scope of the present disclosure is not limited by this effect. Brief explanation of the drawing

[0021] FIG. 1 is a block diagram schematically illustrating an entire system for performing intelligent steel surface inspection according to an exemplary embodiment of the present disclosure. FIG. 2 is a block diagram schematically illustrating an intelligent steel surface inspection device among subsystems according to an exemplary embodiment of the present disclosure. FIG. 3 is a process diagram schematically illustrating the surface defect analysis process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure. FIG. 4 is a process diagram schematically illustrating the learning process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure. FIG. 5 is a process diagram schematically illustrating the data augmentation process of a surface data augmentation generation engine of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure. FIG. 6 is a conceptual diagram schematically illustrating the additive margin softmax of a learning model according to an exemplary embodiment of the present disclosure. FIG. 7 is a schematic diagram illustrating the relearning process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure. FIG. 8 is a flowchart schematically illustrating an intelligent steel surface inspection method according to an exemplary embodiment of the present disclosure. FIG. 9 is a flowchart schematically illustrating the process of labeling some steel surface data of a labeling section according to an exemplary embodiment of the present disclosure. Specific details for implementing the invention

[0022] The present disclosure is capable of various modifications and may have various embodiments; specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the drawings. However, the present disclosure is not limited to the embodiments disclosed below and may be implemented in various forms.

[0023] In the following embodiments, terms such as first, second, etc. are used not in a limiting sense, but for the purpose of distinguishing one component from another component.

[0024] In the following examples, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0025] In the following embodiments, terms such as "include" or "have" mean that the features or components described in the specification are present, and do not preclude the possibility that one or more other features or components may be added.

[0026] In the following embodiments, when a part such as a layer, region, or component is described as being on or above another part, it includes not only cases where it is directly on top of another part, but also cases where another region, component, etc. is interposed in between.

[0027] In the drawings, the size of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are depicted arbitrarily for convenience of explanation, and therefore the present disclosure is not necessarily limited to what is depicted.

[0028] Where an embodiment can be implemented differently, a specific sequence of operations may be performed differently from the order described. For example, two steps described consecutively may be performed substantially simultaneously or proceed in the reverse order of the description.

[0029] In this specification, “A and / or B” indicates the case where it is A, B, or A and B. And, “at least one of A and B indicates the case where it is A, B, or A and B.”

[0030] In the following embodiments, when layers, regions, components, etc. are described as being connected, this includes cases where the layers, regions, components are directly connected, or / or cases where other layers, regions, components are interposed between the layers, regions, components to form an indirect connection. For example, when layers, regions, components, etc. are described as being electrically connected in this specification, it indicates cases where the layers, regions, components, etc. are directly electrically connected, and / or cases where other layers, regions, components, etc. are interposed between them to form an indirect electrical connection.

[0031] The x-axis, y-axis, and z-axis are not limited to the three axes of an orthogonal coordinate system but can be interpreted in a broader sense that includes them. For example, the x-axis, y-axis, and z-axis may be orthogonal to each other, but they may also refer to different directions that are not orthogonal to each other.

[0032] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure is complete and to fully inform those skilled in the art of the scope of the present disclosure, and the present disclosure is defined only by the scope of the claims.

[0033] The terms used in this disclosure are for describing the embodiments and are not intended to limit this disclosure. In this disclosure, the singular form may include the plural form unless specifically stated otherwise in the text. The terms “comprises” and / or “comprising” used in this disclosure do not exclude the presence or addition of one or more other components in addition to the components mentioned. Throughout the disclosure, the same reference numerals refer to the same components, and “and / or” may include each of the mentioned components and all combinations of one or more. Although terms such as “first,” “second,” etc., are used to describe various components, these components are not limited by these terms. These terms are used merely to distinguish one component from another. Accordingly, the first component mentioned below may be the second component within the technical scope of this disclosure.

[0034] The word "exemplary" is used in this disclosure to mean "used as an example or illustration." Any embodiment described as "exemplary" in this disclosure should not be interpreted as being preferred or having an advantage over other embodiments.

[0035] Embodiments of the present disclosure may be described in terms of functions or blocks performing functions. A block, which may be referred to as a “part” or “module” in the present disclosure, may be physically implemented by analog or digital circuits such as logic gates, integrated circuits, microprocessors, microcontrollers, memory, passive electronic components, active electronic components, optical components, hardwired circuits, etc., and may optionally be driven by firmware and software. Additionally, the term “part” as used in the disclosure refers to a hardware element such as software, FPGA, or ASIC, and the “part” may perform certain roles. However, the meaning of “part” is not limited to software or hardware. The “part” may be configured to reside in an addressable storage medium or may be configured to run one or more processors. Accordingly, as an example, a "part" may include elements such as software elements, object-oriented software elements, class elements, and task elements, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functionality provided within the elements and "parts" may be combined into a smaller number of elements and "parts" or further separated into additional elements and "parts."

[0036] Embodiments of the present disclosure may be implemented using at least one software program running on at least one hardware device and may perform network management functions to control elements.

[0037] Spatially relative terms such as "below," "beneath," "lower," "above," and "upper" may be used to facilitate the description of the relationship between one component and other components as illustrated in the drawings. Spatially relative terms should be understood as encompassing different orientations of components during use or operation, in addition to the orientations depicted in the drawings. For example, if a component depicted in a drawing is inverted, a component described as "below" or "beneath" of another component may be placed "above" of that component. Therefore, the exemplary term "below" may encompass both the lower and upper directions. Components may also be oriented in other directions, and accordingly, spatially relative terms may be interpreted according to the orientation.

[0038] Unless otherwise defined, all terms used in this disclosure (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which this disclosure pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0039] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same reference numerals, and redundant descriptions thereof will be omitted.

[0040] FIG. 1 is a block diagram schematically illustrating an overall system for performing intelligent steel surface inspection according to an exemplary embodiment of the present disclosure, and FIG. 2 is a block diagram schematically illustrating an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure.

[0041] The intelligent steel surface inspection device may be implemented as a physical layer (3000) that is deployed in a steel manufacturing facility to collect steel surface data, a subsystem (1000) including a surface inspection system (100) that performs surface monitoring and data analysis based on the collected data and a process optimization system (200), and an integrated management system (2000) that performs image data collection, image data processing, condition data collection, condition data processing, surface condition monitoring, risk situation monitoring, production condition optimization, warning notification, quality prediction monitoring, risk situation detection, production condition standardization, etc. according to the results obtained from the physical layer (3000) and the subsystem (1000).

[0042] The surface inspection system (100) performs surface monitoring based on real-time collected image and sensor data and performs quality analysis through a surface abnormality notification API, and the process optimization system (200) can perform pickling process monitoring, risk state monitoring, surface data augmentation generation, optimization analysis, etc. based on AI deep learning and big data through preprocessing of collected data.

[0043] Additionally, the surface inspection system (100) can perform labeling corresponding to whether the steel surface data collected through the labeling unit (120) is normal or defective.

[0044] In addition, the process optimization system (200) receives steel surface data from the data collection unit (110) that collects image and sensor data from the physical layer (3000), inputs it to the surface data enhancement generation engine (210), and can enhance the steel surface data.

[0045] Steel surface data and augmented steel surface data can be preprocessed in a data preprocessing / matching engine (220) and stored in a DB, and based on the stored data, data can be analyzed in a data analysis engine (230).

[0046] The process optimization system (200) provides analysis results through a result transmission API, and based on the analysis results, if an abnormal situation occurs in the process or if parameters that are not optimized are detected, it can provide an alarm or warning through an alarm / warning API.

[0047] As illustrated in FIG. 2, the intelligent steel surface inspection device according to the present disclosure may include a data collection unit (110) that is placed in a steel manufacturing facility to collect steel surface data, a labeling unit (120) that receives a label corresponding to the collected steel surface data and maps it to the steel surface data, and a learning model (130) that learns a representation based on supervised contrastive learning using the labeled steel surface data and the augmented steel surface data.

[0048] The above learning model (130) may include a surface data augmentation generation engine (210) that infers unlabeled steel surface data and extracts some steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data.

[0049] In other words, the intelligent steel surface inspection device is a combination or combination of functional configurations involved in steel surface inspection in the surface inspection system (100) and the process optimization system (200), and the intelligent steel surface inspection device may be software implemented on separate hardware rather than a single piece of hardware, or a single piece of software implemented on separate hardware.

[0050] The data collection unit (110) may include a shooting unit that photographs the steel surface and an IoT sensor unit that acquires information about the steel surface. For example, it may include a CCTV that photographs the steel manufacturing process, a current sensor that applies current to the steel surface to detect defects by generating a magnetic field on the steel surface, a temperature sensor that measures the temperature of the steel manufacturing process, a concentration sensor that measures the gas concentration, etc.

[0051] The data collection unit (110) may include a part of the steel manufacturing facility's shooting unit or sensor unit, rather than a separate shooting unit or sensor unit.

[0052] The labeling unit (120) is connected to a user interface and can receive a label corresponding to the steel surface data. The label corresponding to the steel surface data can be input from the user interface.

[0053] The learning model (130) can perform supervised contrastive learning. When unlabeled data, for example, unlabeled steel surface data, is input, the learning model (130) can output representations that are similar for the same class and different for different classes by measuring similarity between representation vectors.

[0054] When training a learning model (130) is performed with additional data, that is, a changed training data set, a change in the representation space may occur. In a non-fixed representation space, a problem may arise in which representations of the same class and different classes are repeatedly output, resulting in a similar relationship between the initial model trained earlier and the representation vectors.

[0055] Therefore, in order to maintain the representation vectors between the same classes of the learning model (130) in a fixed manner, the learning model (130) can extract data with a high amount of learning and relearn data of high importance.

[0056] Accordingly, the labeling unit (120) receives steel surface data having a similarity lower than the preset reference similarity among the inference data output by the learning model (130) to which unlabeled steel surface data is input, and receives a label corresponding to the steel surface data and maps it to the received steel surface data.

[0057] Steel surface data having a similarity lower than the aforementioned preset standard similarity may be data for which it is not determined whether it is a good or defective product in vision inspection.

[0058] In other words, it is possible to infer unlabeled steel surface data and extract some of the steel surface data that have uncertainty.

[0059] Data with uncertainty is data that is difficult to determine whether it is a good product or a defective product in vision inspection, and may be data where the representation vector exists outside the range of representation vectors belonging to any one class or on the boundary line.

[0060] The learning model (130) can perform retraining using steel surface data having a similarity lower than the preset reference similarity to which the label is mapped.

[0061] The steel surface data having a similarity lower than the aforementioned preset reference similarity can be labeled, and the labeled steel surface data can be accumulated and stored in the DB, and the learning model (130) can be retrained with the accumulated labeled steel surface data.

[0062] Instead of performing labeling work on the entire steel surface data, by extracting and labeling only the data that is important for training and has a large amount of training volume, time and cost can be reduced and more effective deep learning model training can be carried out.

[0063] FIG. 3 is a process diagram schematically illustrating the surface defect analysis process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure, and FIG. 4 is a process diagram schematically illustrating the learning process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure.

[0064] As illustrated in FIG. 3, the process may go through a data collection step (S310) in which data is collected through a data collection unit (110), a data analysis and application step (S320) in which data is analyzed and applied to steel manufacturing equipment based on the collected data, and a steel manufacturing demonstration step (S330) in which the steel manufacturing process is verified to be actually optimized.

[0065] At this time, in the data collection step (S310), the surface data augmentation generation step (S311), which generates augmented steel surface data based on the collected steel surface data, can be performed repeatedly.

[0066] A learning model (130) can learn the similarity of representation vectors by comparing collected steel surface data and augmented steel surface data, and can classify data before performing data analysis in the data analysis and application step (S320) using the learned learning model (130).

[0067] The data analysis and application step (S320) implements a deep learning-based surface error detection and analysis optimization decision-making system for each manufacturing process, and can apply optimization technology through real-time manufacturing quality detection, 4M risk environment analysis technology, and data and image big data technology.

[0068] In the steel manufacturing demonstration phase (S330), the To be effect can be verified by verifying the quality intelligence system.

[0069] Specifically, the data augmentation generation step (S311) of the data collection step (S310) may include, as illustrated in FIG. 4, data preprocessing (S401), primary data classification (S402), secondary data classification and post-processing (S403), and result transmission (S404) of the input steel surface data.

[0070] When steel surface data is input, in the data preprocessing (S401), the steel surface image captured at an angle can be converted into a defined standard. Alternatively, it can be preprocessed into a form that can be input into an artificial intelligence model.

[0071] At this time, the surface data augmentation generation engine (210) can generate an augmentation data set for a normal data set and a defective data set and perform data preprocessing (S401).

[0072] While it is necessary to classify steel pipes into global and local defects, and to classify them into good or defective products, significant data imbalance by process and classification type can affect model performance.

[0073] Accordingly, in the first stage classification (S402), a classification of pickling defects or normal pickling is performed using a pickling classification model, and in the second stage classification and post-processing (S403), additional classification can be performed using a defect state classification model or a surface defect detection model.

[0074] The impact of data volume imbalance can be minimized by classifying data that may differ in volume into two stages and applying an augmented data set by applying an ensemble to the classified data.

[0075] FIG. 5 is a process diagram schematically illustrating the data augmentation process of a surface data augmentation generation engine of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure, and FIG. 6 is a conceptual diagram schematically illustrating the additive margin softmax of a learning model according to an exemplary embodiment of the present disclosure.

[0076] According to the present embodiment, the surface data augmentation generation engine (210) can generate augmented steel surface data by data splitting, size adjustment, or color conversion for the input steel surface data.

[0077] As illustrated in FIG. 5, image modulation (color conversion) (S611) is performed on a normal data set, and the learning model (130) can receive the image modulated normal data and the original normal data as inputs to learn that the data generated based on the same image are similar for the image modulated normal data and the original normal data (S612).

[0078] The defective data set also undergoes image modulation (size adjustment) (S621), and the learning model (130) can receive the image-modified defective data and the original defective data as inputs to learn that the data generated based on the same image are similar to the image-modified defective data and the original defective data (S622).

[0079] The learning model (130) can generate similar representation vectors for similar images (S613) and learn to maximize the vector similarity of each data in a direction in which the contrast loss function between the steel surface data and the augmented steel surface data is minimized.

[0080] The contrast loss function is calculated using the following mathematical formula 1, and in order for the above contrast loss function to be minimized, it must be calculated such that the cosine similarity is maximized.

[0081]

[0082] Here, P(i) is the set of positive samples in the batch, A(i) is all samples in the batch, represents the temperature hyper-parameter. z represents the coordinates of each representation vector. Through this, the representation between positive and negative samples can be learned.

[0083] According to the present embodiment, the inferred steel surface data includes positive samples and negative samples, and the preset reference similarity can be determined according to an additive margin softmax loss function to maintain the projection distance between the positive samples and the negative samples.

[0084] As shown in Fig. 6, a fixed decision margin can be set between the decision boundary of class 1 and the decision boundary of class 2 according to the additive margin softmax loss function, and data located within the margin may be unclear data that is not class 1 or class 2.

[0085] The additive margin softmax loss function is trained to maintain a distance of at least a certain size for each class, enabling stable training by maintaining the distance between projected representation vectors even when the amount of data is small.

[0086] FIG. 7 is a schematic diagram illustrating the relearning process of an intelligent steel surface inspection device according to an exemplary embodiment of the present disclosure.

[0087] As illustrated in FIG. 7, training of a learning model (130) is performed from a labeled steel surface data DB (310), and during the training process, a surface data augmentation generation engine (210) can generate augmented steel surface data corresponding to the labeled steel surface data.

[0088] The learning model (130) is trained so that the representation vectors of the labeled steel surface data and the augmented steel surface data are similar, and can perform good or defective product inference from the unlabeled steel surface data DB (320).

[0089] In the case of data that is unclear whether it is good data or defective data as a result of the inference of the learning model (130), for example, unclear data included in the fixed decision margin, the learning model (130) can extract it separately and transmit it to the labeling unit (120).

[0090] The labeling unit (120) can map the corresponding labeling data of unclear data extracted through the user interface and store it in the labeled steel surface data DB (310).

[0091] The learning model (130) can perform retraining from the labeled steel surface data DB (310) in which additional data is accumulated and stored to more accurately produce a representation vector for unclear data.

[0092] FIG. 8 is a flowchart schematically illustrating an intelligent steel surface inspection method according to an exemplary embodiment of the present disclosure, and FIG. 9 is a flowchart schematically illustrating a process of labeling some steel surface data of a labeling unit according to an exemplary embodiment of the present disclosure.

[0093] A steel surface inspection method according to the present disclosure may include the steps of: collecting steel surface data by being placed in a steel manufacturing facility; receiving a label corresponding to the collected steel surface data and mapping it to the steel surface data; learning a representation based on supervised contrastive learning for a learning model using the labeled steel surface data and the augmented steel surface data; and the learned learning model inferring unlabeled steel surface data and extracting steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data.

[0094] Additionally, the method may further include the step of receiving a label corresponding to the steel surface data and mapping it to the received steel surface data, and the step of performing retraining using the steel surface data having a similarity lower than the preset reference similarity to which the label is mapped.

[0095] As illustrated in FIG. 8, N steel surface data can be labeled (S1010), and a learning model (130) can be trained using the labeled steel surface data (S1020). Similarity between representation vectors inferred by the learning model (130) can be compared (S1030), and K steel surface data with similarity lower than the reference similarity can be extracted (S1040). The learning model (130) can be retrained using the extracted K steel surface data (S1050), and the training can be completed as is (Yes at S1060), or if K accumulated steel surface data remain or additional training is needed because the training efficiency is low (No at S1060), the process can be returned to S1030 to perform additional training.

[0096] Additionally, as illustrated in FIG. 9, when the step of extracting K steel surface data (S1040) is implemented, steel surface data having a similarity lower than the preset reference similarity is received (S1040), labels corresponding to the K steel surface data are received and labeled on the K steel surface data (S1041), and the labeled steel surface data is accumulated in the initial labeling data (S1042).

[0097] Retraining of the learning model (130) can be performed using the accumulated labeling data.

[0098] Therefore, by selecting only some data with a high amount of training data from the inference results of unlabeled data and performing labeling work, and by performing retraining with only some labeled data, the cost and time associated with the labeling work can be reduced, while the learning efficiency of the learning model (130) can be maximized and the similarity between the representation vectors of the learning model (130) can be consistently maintained.

[0099] Although the present disclosure has primarily described intelligent surface inspection devices, it is not limited thereto. For example, a surface inspection method for performing such an intelligent surface inspection device and a method for manufacturing an intelligent surface inspection device are also considered to fall within the scope of the present disclosure.

[0100] Although the present disclosure has been described with reference to the embodiments illustrated in the drawings, this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present disclosure should be determined by the technical spirit of the appended claims. Explanation of the symbols

[0101] 100: Surface Inspection System 200: Process Optimization System 110: Data Collection Unit 120: Labeling section 130: Learning Model 210: Surface Data Augmentation Generation Engine 220: Data Preprocessing / Consistency Engine 230: Data Analysis Engine 310, 320: DB

Claims

Claim 1 An intelligent steel surface inspection device comprising: a data collection unit deployed in a steel manufacturing facility to collect steel surface data; a labeling unit that receives a label corresponding to the collected steel surface data and maps it to the steel surface data; and a learning model that learns a representation based on supervised contrastive learning using the labeled steel surface data and the augmented steel surface data, wherein the learning model infers unlabeled steel surface data and includes a surface data augmentation generation engine that extracts some of the steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data, and wherein the steel surface data having a similarity lower than the preset reference similarity is data for which it is not determined whether it is a good product or a defective product in vision inspection. Claim 2 In claim 1, the labeling unit receives steel surface data having a similarity lower than the preset reference similarity, and receives a label corresponding to the steel surface data and maps it to the received steel surface data, thereby forming an intelligent steel surface inspection device. Claim 3 In claim 2, the learning model is an intelligent steel surface inspection device that performs relearning using steel surface data having a similarity lower than the preset reference similarity to which the label is mapped. Claim 4 An intelligent steel surface inspection device according to claim 3, wherein the inferred steel surface data includes positive samples and negative samples, and the preset reference similarity is determined according to an additive margin softmax loss function to maintain the projection distance between the positive samples and the negative samples. Claim 5 delete Claim 6 In claim 1, the surface data augmentation generation engine is an intelligent steel surface inspection device that generates augmented steel surface data by data splitting, size adjustment, or color conversion for input steel surface data. Claim 7 In claim 6, the learning model is an intelligent steel surface inspection device that learns such that the vector similarity of each data is maximized in a direction in which the contrast loss function between the steel surface data and the augmented steel surface data is minimized. Claim 8 In claim 1, the data collection unit comprises an intelligent steel surface inspection device including a shooting unit that photographs the steel surface and an IoT sensor unit that acquires the steel surface data. Claim 9 An intelligent steel surface inspection method comprising: a step of collecting steel surface data by being placed in a steel manufacturing facility; a step of receiving a label corresponding to the collected steel surface data and mapping it to the steel surface data; a step of learning a representation based on supervised contrastive learning for a learning model using labeled steel surface data and augmented steel surface data; and a step of the learned learning model inferring unlabeled steel surface data and extracting steel surface data having a similarity lower than a preset reference similarity among the inferred steel surface data; wherein the steel surface data having a similarity lower than the preset reference similarity is data for which it is not determined whether it is a good product or a defective product in vision inspection. Claim 10 An intelligent steel surface inspection method according to claim 9, further comprising: receiving steel surface data having a similarity lower than the preset reference similarity; receiving a label corresponding to the steel surface data and mapping it to the received steel surface data; and performing retraining using the steel surface data having a similarity lower than the preset reference similarity to which the label is mapped.

Images