Apparatus and method for automatically inspecting external defects in welds

Abstract

The present invention relates to an automatic external defect inspection device and method. The automatic external defect inspection device according to the present invention includes an image acquisition unit for acquiring image data of a welded product, a defect detection unit for detecting external defects of the welded product using an artificial intelligence-based external defect detection model based on the image data of the welded product, and a weld quality judgment unit for determining the quality of the welded product based on the type and size estimation values ​​of the external defects.

Description

Technology Field

[0001] The present invention relates to a device and method for quality inspection of welded products. Background Technology

[0002] During welding processes, penetration depth and strength can vary depending on the characteristics of the welding method, the material properties of the base and consumables, and environmental conditions; this can lead to defects in the welded joint. While large-scale manufacturing plants largely automate the management of such defects, quality inspection in small and medium-sized factories—which constitute a significant portion of the manufacturing industry—especially in foundational industries like welding and joining, is still conducted through visual inspection by workers or cross-section inspection (destructive testing). However, these methods are susceptible to varying results depending on the inspector's experience, and cross-section inspection, in particular, faces limitations in ensuring reliability as 100% inspection is practically impossible.

[0003] To overcome the limitations of the aforementioned cross-sectional inspection, various non-destructive tests such as visual inspection (VT), ultrasonic inspection (UT), magnetic particle inspection (MT), and radiology inspection (RT) are being attempted.

[0004] FIGS. 1a to 1f are drawings illustrating conventional external defect inspection techniques for welds. FIG. 1a is an example drawing of cross-section inspection, FIG. 1b is an example drawing of inspection of weld throat and weld leg, FIG. 1c is an example drawing of Visual Test (VT), FIG. 1d is an example drawing of Ultrasonic Test (UT), FIG. 1e is an example drawing of Magnetic Test (MT), and FIG. 1f is an example drawing of Radiology Test (RT).

[0005] Meanwhile, regarding spatter among external defects, most are generally removed through subsequent processes (such as grinding); however, if the spatter is large or the weld strength is high, it may remain unremoved, so it is necessary to inspect the post-processed results for any residual spatter. Additionally, the fact that the allowable size of spatter varies depending on the intended use of the welded product must also be taken into account during defect inspection. The problem to be solved

[0006] The present invention relates to a vision inspection of welded joints. A function capable of automatically determining defects by considering not only the shape characteristics but also the size characteristics of external defect elements. The purpose is to provide a welding external defect inspection device and a method including this.

[0007] The present invention aims to provide an automatic welding external defect inspection device and method capable of detecting an external defect area (such as spatter) based on shape characteristics detected by a vision camera and estimating the size of the external defect using a distance measuring means, thereby determining the final status of whether the spatter is normal or defective by comprehensively judging the shape information and size information of the external defect element.

[0008] The objectives of the present invention are not limited to those mentioned above, and other unmentioned objectives will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0009] An automatic external defect inspection device for welding according to one embodiment of the present invention includes: an image acquisition unit for acquiring image data of a welded product; a defect detection unit for detecting external defects of the welded product using an artificial intelligence-based external defect detection model based on the image data of the welded product; and a welding quality judgment unit for determining the quality of the welded product based on the type and size estimation values ​​of the external defects.

[0010] In one embodiment of the present invention, the types of external defects may be classified according to the shape characteristics of the external defects.

[0011] In one embodiment of the present invention, the welding quality judgment unit determines whether the type of external defect corresponds to a defect requiring consideration of size characteristics, and if the type of external defect corresponds to a defect requiring consideration of size characteristics, it can determine the quality of the welded product based on the size estimate value.

[0012] In one embodiment of the present invention, the automatic external defect inspection device may further include a size estimation unit. In this case, the defect detection unit generates the type of external defect and bounding box information using the external defect detection model based on image data of the weld result. The size estimation unit generates an estimated size value of the external defect using a previously generated size conversion formula based on the bounding box information of the external defect.

[0013] In one embodiment of the present invention, the defect detection unit may generate bounding box information of the weld sample using the external defect detection model based on image data of the weld sample acquired by the image acquisition unit. In this case, the size estimation unit generates the size transformation equation using a regression analysis algorithm based on the bounding box information of the weld sample, the distance measurement between the image acquisition unit and the weld sample, the focal length of the camera lens included in the image acquisition unit, and the actual size of the defect included in the weld sample.

[0014] In one embodiment of the present invention, the regression analysis algorithm may be Support Vector Regression.

[0015] In one embodiment of the present invention, the size estimation unit may generate a size estimation value of the weld sample using the size transformation formula based on the bounding box information of the weld sample, the distance measurement value, and the focal length, and may generate a size correction formula using a regression analysis algorithm based on the size estimation value of the weld sample and the actual size of the defect included in the weld sample. Furthermore, the size estimation unit may correct the size estimation value of the external defect using the size correction formula.

[0016] In one embodiment of the present invention, the size estimation unit may generate the size correction equation using a linear regression analysis technique.

[0017] In one embodiment of the present invention, the automatic external welding defect inspection device may further include an AI model learning unit. The AI ​​model learning unit generates the external defect detection model by training an artificial intelligence model based on a learning dataset.

[0018] In one embodiment of the present invention, the AI ​​model learning unit can generate the external defect detection model by training a pre-trained model using a transfer learning technique based on the training dataset.

[0019] In one embodiment of the present invention, the training dataset includes a plurality of welding sample images and annotation files.

[0020] In one embodiment of the present invention, the pre-trained model may include any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof.

[0021] A method for automatically inspecting external defects in a weld according to an embodiment of the present invention comprises: a step of acquiring image data of a welded product; a step of detecting external defects in the welded product using an artificial intelligence-based external defect detection model based on the image data of the welded product; and a step of determining the quality of the welded product based on the detection result of the external defects.

[0022] In one embodiment of the present invention, the automatic external defect inspection method for welding may further include the step of generating the external defect detection model by training an artificial intelligence model based on a training dataset.

[0023] In one embodiment of the present invention, the step of determining the quality of the welded product may include: determining whether an external defect is detected in the welded product; determining the quality of the welded product as normal if no external defect is detected in the welded product; determining whether the type of external defect corresponds to a defect requiring consideration of size characteristics if an external defect is detected in the welded product; determining the quality of the welded product as defective if the type of external defect does not correspond to a defect requiring consideration of size characteristics; calculating an estimated size value of the external defect if the type of external defect corresponds to a defect requiring consideration of size characteristics; and determining the quality of the welded product as defective if the estimated size value is greater than or equal to a set threshold.

[0024] In one embodiment of the present invention, the types of external defects may be classified according to the shape characteristics of the external defects.

[0025] In one embodiment of the present invention, the step of calculating the size estimate may include: a step of generating bounding box information of the external defect using the external defect detection model based on image data of the weld result; and a step of calculating the size estimate using a size conversion formula generated based on the bounding box information of the external defect, a distance measurement value between the vision camera and the weld result, and the lens focal length of the vision camera, and correcting the size estimate using a size correction formula generated.

[0026] In one embodiment of the present invention, the automatic external defect inspection method for welding may further include the step of generating bounding box information of the welding sample using the external defect detection model based on image data of the welding sample, and generating the size transformation formula using a regression analysis algorithm based on the bounding box information of the welding sample, a distance measurement between the vision camera and the welding sample, the lens focal length of the vision camera, and the actual size of the defect included in the welding sample.

[0027] A method for automatically inspecting external defects in welding according to an embodiment of the present invention comprises: a step of generating an external defect detection model by training a pre-trained model using a transfer learning technique based on a training dataset; a step of acquiring image data of a welding result; and a step of detecting external defects in the welding result using the external defect detection model based on the image data of the welding result.

[0028] In one embodiment of the present invention, the training dataset includes a plurality of welding sample images and annotation files.

[0029] In one embodiment of the present invention, the pre-trained model may include any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof. Effects of the invention

[0030] According to one embodiment of the present invention, when a defect is determined in a weld vision inspection, Automatically determines whether a defect exists by considering not only the shape characteristics but also the size characteristics of the defective element. By doing so, there is an advantage in that the accuracy of the inspection can be increased.

[0031] And, according to one embodiment of the present invention, by estimating the size of the final defect using a size transformation formula and a correction formula, The error can be reduced compared to the case where only the existing size conversion formula is used. .

[0032] In addition, according to one embodiment of the present invention, objective and rapid inspection results can be obtained by a non-destructive method replacing conventional visual inspection, thereby improving the reliability of the inspection results, and the welded product Full inspection is possible Therefore, it has the advantage of facilitating quality control.

[0033] The effects obtainable from the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the description below. Brief explanation of the drawing

[0034] Figures 1a to 1f are drawings illustrating conventional external defect inspection techniques for welded joints. FIG. 2 is a block diagram showing the configuration of an automatic external welding defect inspection device according to one embodiment of the present invention. FIGS. 3a and 3b are drawings for illustrating an automatic external welding defect inspection method according to an embodiment of the present invention. Figure 4 is an example drawing of image data of the welding result. Figure 5 is an example diagram of the result of detecting external defects in the welding result image data. Figure 6 is an example diagram of the result of extracting the area where external defects exist from the welding result image data. Figure 7 is a diagram illustrating the process of calculating the pixel area of ​​an external defect-existing region. FIG. 8 is an example drawing of the welding quality judgment result of the automatic external welding defect inspection device and method according to one embodiment of the present invention. FIGS. 9a to 9c are example drawings for explaining the process of deriving size conversion formulas and size correction formulas. Figure 10 is an example diagram illustrating the process of training an artificial intelligence model to generate an external defect detection model. Figure 11 is an example diagram illustrating the process of inferring external defects using an external defect detection model based on a test dataset. Figures 12a and 12b are diagrams showing examples of inference results of an external defect detection model. FIG. 13 is a block diagram showing a computer system for implementing an automatic external welding defect inspection method according to an embodiment of the present invention. Specific details for implementing the invention

[0035] The present invention relates to an apparatus and method applicable to quality inspection of welded products. More specifically, the present invention relates to a structure of a weld external defect inspection apparatus and a method thereof, wherein, during weld external defect inspection, an artificial intelligence model is trained to learn the shape characteristics of predefined external defect elements (e.g., overlap, undercut, porosity, spatter, etc.) based on images collected through a vision camera, and then the area where defects exist is detected using the artificial intelligence learning model. Through the present invention, the determination of whether a welded product is normal or defective can be performed in an automated and objective manner.

[0036] In the case of spatter, it is generally mostly removed through post-processing (such as grinding); however, if the spatter is large or the weld strength is high, it may remain unremoved, making it necessary to inspect the post-processed results for any residual spatter. Furthermore, the allowable size of spatter varies depending on the intended use of the welded product. For instance, in applications where structural strength is critical, such as automotive chassis, spatter with a diameter of typically 2–3 mm or less is usually accepted as normal and cannot be removed; however, for exterior materials, spatter larger than 1 mm can affect the appearance and must be treated as a serious defect. Therefore, during vision inspection, the decision of whether to treat areas detected as spatter as final defects or to classify them as normal is... Judgment accuracy can be improved by reflecting not only the shape information of the spatter but also the actual size information of the spatter. .

[0037] The automatic external welding defect inspection device and method according to the present invention considers shape characteristics and size characteristics simultaneously for spatter and determines it as a defect only when the estimated size of the spatter is greater than or equal to a threshold set by a user.

[0038] Therefore, through the present invention Improve inspection accuracy by comprehensively determining shape and size information of external defect elements. Can make it happen.

[0039] The advantages and features of the present invention 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 invention 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 of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Meanwhile, the terms used in this specification are for describing the embodiments and are not intended to limit the present invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. The terms "comprises" and / or "comprising" as used in this specification do not exclude the presence or addition of one or more other components, steps, actions, and / or elements in addition to the mentioned components, steps, actions, and / or elements.

[0040] In describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may unnecessarily obscure the essence of the invention.

[0041] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In order to facilitate overall understanding in describing the present invention, the same reference numerals will be used for the same means regardless of the drawing number.

[0042] FIG. 2 is a block diagram showing the configuration of an automatic external welding defect inspection device according to an embodiment of the present invention. The automatic external welding defect inspection device is a device capable of automatically determining whether there is a defect by comprehensively considering the shape characteristics and size characteristics of external defect elements in a weld vision inspection.

[0043] Referring to FIG. 2, an automatic external welding defect inspection device (100, hereinafter referred to as the 'inspection device') according to one embodiment of the present invention comprises an AI model learning unit (110), an image acquisition unit (120), a defect detection unit (130), a welding quality judgment unit (140), a distance measurement unit (150), and a size estimation unit (160). The inspection device (100) illustrated in FIG. 2 is according to one embodiment, and the components of the inspection device (100) according to the present invention are not limited to the embodiment illustrated in FIG. 2 and may be added, changed, or deleted as needed. For example, if the defect detection unit (130) incorporates a pre-set AI-based external defect detection model, the AI ​​model learning unit (110) is omitted from the inspection device (100).

[0044] The AI ​​model learning unit (110) generates the external defect detection model by training an artificial intelligence model based on a learning dataset. The AI ​​model learning unit (110) can generate the external defect detection model by training a pre-trained model using a transfer learning technique based on a learning dataset.

[0045] Here, the training dataset includes multiple weld sample images and annotation files. The weld sample image data consists of images obtained by capturing each weld sample with a vision camera. The annotation files contain information regarding the presence of weld beads or external defects and the types of defects (which may be referred to as 'defect types' or 'labels'). The defect types (labels) may be classified based on the geometric characteristics of the defects. For example, defect types (labels) may have values ​​such as 'pass', 'overlap', 'undercut', 'crack', 'spatter', and 'porosity'.

[0046] And, the pre-trained model used by the AI ​​model training unit (110) may include any one of the EfficientDet model, YOLO and Faster-RCNN, or a combination thereof.

[0047] The image acquisition unit (120) takes a picture of a weld sample to acquire image data of the weld sample, or takes a picture of a weld result to acquire image data of the weld result.

[0048] The image acquisition unit (120) is composed of a combination of a camera (CMOS, CCD, etc.) for collecting real image data, an imaging lens for forming an image with the camera, and a polarizing filter for preventing light scattering. In addition, a telecentric lens for preventing image distortion or a filter capable of transmitting only light of a specific wavelength band may be additionally configured.

[0049] The defect detection unit (130) generates defect type information included in the weld sample and external defect existence area information for each defect using a pre-trained artificial intelligence-based external defect detection model based on image data of the weld sample acquired from the image acquisition unit (120). The external defect existence area information may be bounding box information. The bounding box information may be configured to include the pixel coordinates (xmin) and ymin of the x-axis of the top-left corner of the bounding box, and the pixel coordinates (xmax) and ymax of the x-axis of the bottom-right corner. The bounding box information for the external defect included in the weld sample is used by the size estimation unit (160) to generate a size conversion formula and a size correction formula.

[0050] Additionally, during the inspection process of the welded product, the defect detection unit (130) generates information on the types of external defects included in the welded product and information on the area of ​​existence of each defect (e.g., bounding box information) using a previously trained artificial intelligence-based external defect detection model based on image data of the welded product (hereinafter referred to as 'welded product image data' or 'welded product image') acquired by the image acquisition unit (120).

[0051] The defect detection unit (130) transmits information on the type of external defect generated based on the welding result image data and information on the area of ​​existence of each external defect (e.g., bounding box information) to the welding quality judgment unit (140).

[0052] Additionally, the defect detection unit (130) transmits external defect presence area information (e.g., bounding box information) generated based on welding sample or welding result image data to the size estimation unit (160).

[0053] The welding quality judgment unit (140) determines whether external defects are detected based on the external defect detection results of the defect detection unit (130) for the welding result. The welding quality judgment unit (140) can determine whether external defects are detected based on the defect type information (label) received from the defect detection unit (130). For example, the welding quality judgment unit (140) determines that there are no external defects in the welding result image if there are no detected defect types or if all labels are 'normal (pass)'. Additionally, the welding quality judgment unit (140) determines that there are external defects in the welding result image if there are defect types that are not 'normal (pass)' for the welding result image.

[0054] The welding quality judgment unit (140) determines that there are no external defects in the welding result image, and finally determines the welding quality of the welding result as 'pass'.

[0055] In addition, when the welding quality judgment unit (140) determines that there are external defects in the welding result image, it determines whether the type of external defect corresponds to a defect that requires consideration of size characteristics.

[0056] If the type of external defect detected does not correspond to a defect that requires consideration of size characteristics (e.g., spatter, porosity), the welding quality judgment unit (140) determines the quality of the welded product as defective.

[0057] If the type of external defect in the welded product corresponds to a defect for which size characteristics must be considered, the weld quality judgment unit (140) determines the quality of the welded product based on the size estimate of the external defect. For example, if the size estimate calculated by the size estimation unit (160) is greater than or equal to a set threshold, the weld quality judgment unit (140) determines the quality of the welded product to be defective.

[0058] The distance measuring unit (150) measures the distance between the image acquisition unit (120) and the sample. Additionally, the distance measuring unit (150) measures the distance between the image acquisition unit (120) and the welding result. The distance measuring unit (150) transmits the distance measurement value to the size estimation unit (160).

[0059] The distance measuring unit (150) may use various sensors. For example, to measure the distance between the image acquisition unit (120) and the sample, the distance measuring unit (150) may use any one of the following sensors: an ultrasonic distance sensing sensor, a Time of Flight (TOF) sensor using light, a Laser Imaging Detection and Ranging (LIDAR), a Laser Range Finder (LRF), or an infrared distance sensing sensor. Since an error may occur during regression analysis if the optical axis of the image acquisition unit (120) and the optical axis of the distance measuring unit (150) are misaligned, they must be installed parallel to each other.

[0060] The size estimation unit (160) generates a size transformation equation using a regression analysis algorithm (e.g., Support Vector Regression) based on bounding box information of the weld sample, a distance measurement between the image acquisition unit (120) and the weld sample, the lens focal length of the camera included in the image acquisition unit (120), and the actual size of the defect included in the weld sample.

[0061] And, the size estimation unit (160) inputs information on the external defect area of ​​the weld sample (e.g., bounding box information), a distance measurement between the image acquisition unit (120) and the weld sample, and the lens focal length of the camera included in the image acquisition unit (120) into a size conversion formula to generate an estimated size of the defect included in the weld sample, and generates a size correction formula using a regression analysis algorithm based on the estimated size of the defect in the weld sample and the actual size of the defect included in the weld sample. For example, the size estimation unit (160) can generate a size correction formula using a linear regression analysis technique.

[0062] Additionally, the size estimation unit (160) generates an estimated size value of an external defect included in a welded product (an estimated value for the actual size of the external defect) using a size conversion formula generated based on bounding box information of the external defect included in the welded product, a distance measurement value between the image acquisition unit (120) and the welded product, and the lens focal length of the camera included in the image acquisition unit (120). Then, the size estimation unit (160) inputs the estimated size value of the external defect into a size correction formula to correct the estimated size value of the external defect. The size estimation unit (160) transmits the corrected estimated size value to the weld quality judgment unit (140).

[0063] FIGS. 3a and 3b are drawings illustrating an automatic external welding defect inspection method according to an embodiment of the present invention. The automatic external welding defect inspection method is a method capable of automatically determining whether a defect exists by comprehensively considering the shape characteristics and size characteristics of external defect elements in a weld vision inspection. The automatic external welding defect inspection method according to an embodiment of the present invention includes steps S210 to S300. The automatic external welding defect inspection method according to an embodiment of the present invention can be performed by an inspection device (100).

[0064] The automatic external welding defect inspection method illustrated in FIGS. 3a and 3b is according to one embodiment, and the steps of the automatic external welding defect inspection method according to the present invention are not limited to the embodiment illustrated in FIGS. 3a and 3b and may be added, changed, or deleted as necessary. For example, if an external defect detection model, a size conversion formula, and a size correction formula are already set in the inspection device (100), steps S210 and S220 may be omitted.

[0065] Step S210 is the step of creating an external defect detection model.

[0066] The AI ​​model training unit (110) trains an artificial intelligence model based on a training dataset to generate an external defect detection model. The AI ​​model training unit (110) can generate an external defect detection model by training a pre-trained model using a transfer learning technique based on a training dataset. The pre-trained model may include any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof.

[0067] The training dataset consists of image data and annotation files of multiple weld samples. The image data is obtained by capturing each weld sample with a vision camera. The annotation file contains information on the areas where weld beads or external defects exist, as well as the types of defects (which may be referred to as 'defect types' or 'labels'). The defect types (labels) may be classified based on the geometric characteristics of the defects. For example, defect types (labels) may have values ​​such as 'pass', 'overlap', 'undercut', 'crack', 'spatter', and 'porosity'.

[0068] Information regarding the area of ​​existence of external defects refers to area information regarding defects such as spatter or pores, or weld beads, present in the corresponding weld sample. Information regarding the area of ​​existence of external defects may be bounding box information. Bounding box information may be composed of a combination of the pixel location (xmin) at the horizontal starting point of the bounding box surrounding the defect, the pixel location (ymin) at the vertical starting point, the pixel location (xmax) at the horizontal ending point, and the pixel location (ymax) at the vertical ending point. That is, bounding box information may be composed of the pixel coordinates (xmin) and ymin of the top-left corner of the bounding box, and the pixel coordinates (xmax) and ymax of the bottom-right corner. For convenience of explanation, in the following description, information regarding the area of ​​existence of external defects is bounding box information, and bounding box information is composed of the combination of the aforementioned pixel coordinates.

[0069] Various deep learning models such as the EfficientDet model, YOLO, and Faster-RCNN can be used as pre-trained models utilized by the AI ​​model training unit (110) to perform transfer learning. That is, the pre-trained model may include any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof.

[0070] The AI ​​model learning unit (110) transmits the generated defect detection model to the defect detection unit (130).

[0071] Specific details regarding the task of the AI ​​model learning unit (110) generating an external defect detection model will be described in detail later with reference to FIG. 10.

[0072] Step S220 is the step of deriving size conversion formulas and size correction formulas.

[0073] The image acquisition unit (120) captures a weld sample and acquires image data of the weld sample. The image acquisition unit (120) transmits the image data of the weld sample to the defect detection unit (130).

[0074] The defect detection unit (130) inputs image data of the weld sample into an external defect detection model to generate bounding box information for defects of the weld sample and transmits the bounding box information to the size estimation unit (160).

[0075] The distance measuring unit (150) measures the distance between the image acquisition unit (120) and the welding sample, and transmits the distance measurement value to the size estimation unit (160).

[0076] The size estimation unit (160) receives from the user the focal length (hereinafter abbreviated as 'lens focal length') of the camera lens included in the image acquisition unit (120). Additionally, the size estimation unit (160) receives from the user the actual size of the defect included in the corresponding weld sample (hereinafter referred to as 'actual defect size'). In the present invention, the size of the defect may be the area of ​​the defect or the longest length of the defect.

[0077] The size estimation unit (160) derives a size conversion formula using a regression analysis algorithm based on bounding box information, distance measurements, lens focal length, and actual defect size.

[0078] The regression analysis algorithm used by the size estimation unit (160) may include any one of linear, polynomial, exponential, logarithmic, and support vector regression (SVR), or a combination thereof. Accordingly, the size transformation formula may be configured in various forms including any one of a linear function (first-order function), a polynomial of degree 2 or higher, an exponential function, a logarithmic function, or a combination thereof.

[0079] Then, the size estimation unit (160) inputs bounding box information, distance measurements, and lens focal lengths into a size conversion formula to calculate an initial size estimate (first size estimate). Then, the size estimation unit (160) derives a size correction formula through linear regression analysis based on the initial size estimate (first size estimate) and the actual defect size.

[0080] Specific details regarding the task of the size estimation unit (160) deriving a size conversion formula and a size correction formula will be described in detail later with reference to mathematical formula 1 and FIGS. 9a to 9c.

[0081] Through steps S210 to S220, the inspection device (100) can generate an external defect detection model, a size conversion formula, and a size correction formula.

[0082] From step S230 to step S300, the inspection device (100) determines the welding quality of the welding result using a previously generated external defect detection model, a size conversion formula, and a size correction formula.

[0083] Step S230 is the step of acquiring image data of the welding result.

[0084] The image acquisition unit (120) captures the welding result and acquires image data of the welding result. FIG. 4 is an example drawing of the welding result image (A1). The image acquisition unit (120) transmits the welding result image data to the defect detection unit (130).

[0085] Step S240 is the step for detecting external defects.

[0086] The defect detection unit (130) detects external defects of the welded product using an artificial intelligence-based external defect detection model that has been trained based on the welded product image data. Specifically, the defect detection unit (130) inputs the welded product image data into the external defect detection model to generate defect type (label) and bounding box information. As described above, the defect type (label) may be distinguished according to the shape characteristics of the defect. FIG. 5 is an example diagram of the result of detecting external defects in a welded product image (A1). As shown in FIG. 5, multiple defects may be detected in a single welded product image (A1), and a bounding box (B1) is generated for each detected defect. Although not shown in FIG. 5, the defect detection unit (130) assigns a label to each defect.

[0087] The defect detection unit (130) transmits the defect type (label) and bounding box information of the detected defects (which may be multiple) to the welding quality judgment unit (140).

[0088] That is, the defect detection unit (130) detects predefined defects (e.g., overlap, undercut, crack, porosity, spatter, etc. defined by the user during learning) using an external defect detection model based on the welding result image, and provides area information (bounding box information) of the pixels where each defect exists to the welding quality judgment unit (140).

[0089] And, in the subsequent S270 step, the defect detection unit (130) can provide defect type and bounding box information to the size estimation unit (160).

[0090] The process of the defect detection unit (130) inferring using an external defect detection model and the inference result will be described later with reference to FIG. 11, FIG. 12a, and FIG. 12b.

[0091] Step S250 is a step for determining whether external defects have been detected. That is, this step is a step for determining whether external defects exist in the image of the welded product. Step S250 is a first inspection (screening inspection) step. The weld quality judgment unit (140) determines whether external defects have been detected based on defect type information (label) received from the defect detection unit (130).

[0092] For example, the welding quality judgment unit (140) determines that there are no external defects in the welding result image if there are no detected defect types or if all labels are 'normal (pass)'. Additionally, the welding quality judgment unit (140) determines that there are external defects in the welding result image if there are defect types that are not 'normal (pass)' for the welding result image.

[0093] If the welding quality judgment unit (140) determines that there are no external defects in the welding result image, it finally determines the welding quality of the welding result as 'normal' (S290). Meanwhile, if the welding quality judgment unit (140) determines that there are external defects in the welding result image, it proceeds to step S260.

[0094] Steps S260 through S280 are secondary inspection (precision inspection) processes and are performed for each defect. If multiple defects are detected in the weld result image, steps S260 through S280 may be performed repeatedly several times depending on the type of defect.

[0095] The second inspection (precision inspection) is intended for cases where external defects (e.g., spatter, pores) that require consideration of size characteristics are detected in the S240 stage.

[0096] Step S260 is a step for determining whether a detected external defect is a defect that requires judgment of defect status by considering not only shape characteristics but also size. The welding quality judgment unit (140) determines whether defect size estimation is necessary based on the type of defect detected in the welding result image.

[0097] If the type of defect detected in the welding result image is not a defect that requires the size of the defect to be considered for determination of defect status, the welding quality judgment unit (140) determines the final quality of the welding result as 'defective' (S300). For example, if the type of defect detected in the welding result image is one of overlap, undercut, or crack, and is determined to be defective based only on its shape characteristics, the welding quality judgment unit (140) determines the final quality of the welding result as 'defective' without needing to consider the size of the defect (S300).

[0098] If the type of defect detected in the welding result image is a defect that requires determining whether it is defective by considering the defect size, the welding quality judgment unit (140) proceeds to step S270.

[0099] For example, in the case of a type of defect defined by the user as one that requires consideration of size characteristics, such as spatter or porosity, the welding quality judgment unit (140) proceeds to step S270.

[0100] Step S270 is the step of estimating the actual size of the detected defect.

[0101] The size estimation unit (160) inputs bounding box information for the welding result image, a distance measurement between the image acquisition unit (120) and the welding result, and a lens focal length into a preset size conversion formula to calculate an initial size estimation value (first size estimation value). Then, the size estimation unit (160) inputs the initial size estimation value (first size estimation value) into a size correction formula to calculate a final size estimation value (second size estimation value).

[0102] Detailed information regarding step S270 is explained with reference to FIG. 3b. As illustrated in FIG. 3b, step S270 includes steps S271 through S275. This specification describes steps S271 through S275 assuming that the type of defect detected in the welded product is 'splatter'. For reference, in this embodiment, when spatter is detected among external defects, the inspection device (100) determines whether the welded product is defective by comprehensively considering shape and size characteristics.

[0103] Step S271 is the step of extracting the area where external defects exist.

[0104] The defect detection unit (130) extracts bounding boxes from the weld result image using an external defect detection model. FIG. 6 is an example drawing of the result of extracting bounding boxes (B1) from the weld result image. The defect detection unit (130) transmits bounding box information to the size estimation unit (160).

[0105] Step S272 is the step of calculating the size of the external defect area.

[0106] The size estimation unit (160) calculates the size of the bounding box (e.g., pixel area or number of diagonal pixels) based on the bounding box information.

[0107] In the present invention, the size of the external defect may be the area of ​​the external defect or the length (maximum length) of the external defect; however, in this embodiment, it is assumed that the size of the external defect is the area of ​​the external defect. It is also assumed that the size conversion formula and the size correction formula derived in step S220 are formulas for estimating the area of ​​the external defect.

[0108] According to the above assumption, the size estimation unit (160) calculates the pixel unit area of ​​each external defect based on the bounding box information of each external defect. FIG. 7 is a diagram for explaining the process of calculating the pixel area of ​​the external defect existence area. The size estimation unit (160) calculates the pixel unit area of ​​the corresponding external defect by multiplying the pixel unit horizontal length (C1) and the pixel unit vertical length (C2) of the bounding box (B1) of the external defect.

[0109] Step S273 is a step of measuring the distance between the image acquisition unit (120) and the welded product. The distance measuring unit (150) measures the distance between the image acquisition unit (120) and the welded product and transmits the distance measurement value to the size estimation unit (160).

[0110] Step S274 is the step for estimating the defect size.

[0111] The size estimation unit (160) inputs the pixel-unit area of ​​the external defect, the distance measurement between the image acquisition unit (120) and the welding result, and the preset lens focal length into a size conversion formula derived through regression analysis to calculate an initial size estimation value (first size estimation value). In this embodiment, the initial size estimation value is the initial estimation value for the actual area of ​​the external defect.

[0112] Step S275 is a step for correcting the defect size.

[0113] The size estimation unit (160) inputs the initial size estimation value (first size estimation value) into the size correction formula to calculate the final size estimation value (second size estimation value). In this embodiment, the size estimation unit (160) inputs the initial estimation value for the actual area of ​​the external defect into the size correction formula to calculate the final estimation value for the actual area of ​​the external defect.

[0114] The size estimation unit (160) transmits the final size estimate (second size estimate) of the detected external defect to the welding quality judgment unit (140).

[0115] Step S280 is a step for determining whether the estimated size of the defect is within the acceptable range.

[0116] The welding quality judgment unit (140) compares the final size estimate of the external defect (second size estimate) with a threshold (allowable size) pre-set by the user.

[0117] The welding quality judgment unit (140) finally determines that the external defect is 'defective' if the final size estimate (second size estimate) of the external defect is greater than or equal to a preset threshold. And, the welding quality judgment unit (140) finally determines that the external defect is 'normal' if the final size estimate (second size estimate) of the external defect is less than a preset threshold.

[0118] The threshold (allowable size) for the size of external defects may vary depending on the intended use of the welded product. For example, if the welded product is intended for structural use (e.g., automotive chassis), the threshold for the area of ​​external defects is 4mm, as strength is more important than appearance. 2 It can be set to . In this case, the welding quality judgment unit (140) is 4mm 2 The following spatter is considered normal, and such spatter is not removed, allowing the welded product to be used as is. As another example, if the welded product is intended for exterior use, the appearance is important; therefore, the threshold for the area of ​​external defects is 0.5mm. 2 It can be set to. In this case, 0.5mm 2 Since the above spatter can affect the appearance, the welding quality judgment unit (140) determines this spatter as a defect.

[0119] FIG. 8 is an example diagram of a welding quality judgment result of an automatic external welding defect inspection device and method according to an embodiment of the present invention. A plurality of bounding box information may be assigned to a welding result image (A1). The welding result shown in FIG. 8 has a plurality of spatter defects, and according to the final judgment of the welding quality judgment unit (140), spatter judged to be defective (D1, indicated by a black bounding box) and spatter judged to be normal (E1, indicated by a white bounding box) are distinguished.

[0120] The welding quality judgment unit (140) repeatedly performs step S280 for each defect of the welded product to be inspected, and if there are no defects finally determined to be defective, it determines the welding quality of the welded product to be 'normal' (S290).

[0121] Meanwhile, the welding quality judgment unit (140) determines the welding quality of the welded product as ‘defective’ when at least one defect is finally determined to be defective (S300).

[0122] The above-described automatic external weld defect inspection method has been explained with reference to the flowchart presented in the drawings. For simplicity of explanation, the method has been illustrated and described in a series of blocks; however, the present invention is not limited to the order of said blocks, and some blocks may occur in a different order or simultaneously with other blocks as illustrated and described herein, and various other branches, flow paths, and sequences of blocks may be implemented to achieve the same or similar results. Furthermore, not all illustrated blocks may be required for the implementation of the method described herein.

[0123] Meanwhile, in the description with reference to FIGS. 3a through 8, the steps of the automatic external welding defect inspection method according to one embodiment of the present invention may be further divided into additional steps or combined into fewer steps depending on the embodiment. In addition, some steps may be omitted as necessary, and the order between steps may be changed. Furthermore, even if other omitted details are included, the contents described in this specification may apply to the contents of FIGS. 3a through 8. Also, the contents of FIGS. 3a through 8 may apply to other parts of this specification.

[0124] FIGS. 9a to 9c are example drawings for explaining the process of deriving a size conversion formula and a size correction formula. The size estimation unit (160) can convert the pixel area of ​​an external defect into an actual size using a size conversion formula and correct it using a size correction formula.

[0125] After explaining the function of the size estimation unit (160) and the process of deriving the size conversion formula and size correction formula, the explanation will be given with reference to FIGS. 9a to 9c.

[0126] First, the process of the size estimation unit (160) deriving a size transformation equation through regression analysis is explained. In one embodiment of the present invention, the size estimation unit (160) sets bounding box information, a distance measurement between the image acquisition unit (120) and the weld sample, and the lens focal length of the image acquisition unit (120) as independent variables, and the defect size estimate as a dependent variable.

[0127] In the present invention, the size of the defect may be the area of ​​the defect or the length of the defect (maximum length).

[0128] The process of the size estimation unit (160) deriving a size conversion formula and a size correction formula based on bounding box information is explained in detail below.

[0129] The size estimation unit (160) calculates the number of pixels (diameter) of the diagonal of the bounding box using Equation 1 based on the bounding box information (xmin, ymin, xmax, ymax) of the detected defect.

[0130]

[0132] The distance measuring unit (150) measures the distance between the image acquisition unit (120) and the weld sample and transmits the distance measurement value to the size estimation unit (160). The size estimation unit (160) receives input from the user regarding the focal length (FL) of the camera lens included in the image acquisition unit (120) and the actual size of the defect included in the weld sample (in this example, the length of the defect). The size estimation unit (160) defines the number of pixels (diameter) of the diagonal of the bounding box, the distance measurement value, and the focal length as independent variables, and defines the estimated value of the actual defect size (in this example, the length of the defect) as a dependent variable, thereby deriving a size transformation equation using a regression analysis algorithm. The regression analysis algorithm used in the size estimation unit (160) may take various forms. For example, the regression analysis algorithm may include any one of linear, polynomial, exponential, logarithmic, or support vector regression (SVR), or a combination thereof.

[0133] FIG. 9a is a graph comparing the size estimate value (x-axis, predicted) and the actual defect size (ground truth) after the size estimation unit (160) derives a size transformation equation using a support vector regression algorithm and inputs test data into the derived size transformation equation to obtain a size estimate value (predicted value). In the example of FIG. 9a, the Mean Square Error (MSE) between the size estimate value and the actual defect size is 0.133 and the coefficient of determination (R2) is 0.821.

[0134] FIG. 9b is a diagram illustrating the process by which the size estimation unit (160) derives a size correction formula. The size correction formula is a formula for fitting the size estimate value to the actual value. In the example of FIG. 9b, the size estimation unit (160) derived the size correction formula using linear regression analysis based on the size estimate value and the actual defect size (ground truth). Here, the independent variable was the initial size estimate value (x-axis, predicted value) of FIG. 9a, and the dependent variable was the final size estimate value (y, hereinafter referred to as 'final size estimate value'), and the derived size correction formula is y = 1.2891x - 1.2087.

[0135] After the size correction formula is derived, the size estimation unit (160) calculates the final size estimate by inputting the initial size estimate into the size correction formula when the initial size estimate is calculated from the size conversion formula. Through the size correction formula, a final size estimate with a small error from the actual value while passing through the origin can be obtained as shown in FIG. 9c. In the example of FIG. 9a, the error (MSE) and coefficient of determination (R2) of the size conversion formula were MSE (0.133) and R2 (0.821), respectively, but when the size correction formula is additionally introduced as in the example of FIG. 9c, the effect of reducing the estimation error can be obtained as MSE (0.099) and R2 (0.866).

[0136] In the example described above, the process of the size estimation unit (160) deriving a size conversion formula and a size correction formula based on the number of diagonal pixels of the bounding box calculated based on the bounding box information was explained.

[0137] According to another example of the present invention, the size of a defect may refer to the area of ​​the defect. In this case, the size estimation unit (160) calculates the pixel-unit area of ​​the defect based on bounding box information. The size estimation unit (160) may derive a size conversion formula and a size correction formula that calculate the estimated area of ​​the defect based on the pixel-unit area of ​​the defect.

[0138] FIG. 10 is an example diagram illustrating the process of training an artificial intelligence model to generate an external defect detection model. That is, FIG. 10 is a diagram regarding a method of training an artificial intelligence model to generate an external defect detection model.

[0139] The AI ​​model training unit (110) generates an external defect detection model by training an artificial intelligence model based on a training dataset. The AI ​​model training unit (110) can generate an external defect detection model by training a pre-trained model using a transfer learning technique based on a training dataset.

[0140] A training dataset is required to train an artificial intelligence model. The training dataset consists of image data and annotation files of multiple welding samples. The image data can be collected by capturing each welding sample with a vision camera.

[0141] Furthermore, the annotation file includes information on the area of ​​existence of weld beads or external defects (e.g., bounding box information) and the type of defect (label). In other words, the annotation file must specify the type of defect subject to inspection. The aforementioned bounding box information refers to the bounding box information for defects such as spatter and porosity present in the weld sample, or for the weld bead. For defects that may occur in the weld fusion zone (joint), such as overlap, undercut, and crack, annotations can be placed centered on the weld bead; conversely, for defects existing at unspecified locations outside the joint, such as spatter, annotations can be placed centered on the location where the defect occurred.

[0142] Specifically, the bounding box information may be information including the pixel position of the horizontal starting part (xmin), the pixel position of the vertical starting part (ymin), the pixel position of the horizontal ending part (xmax), and the pixel position of the vertical ending part (ymax) of the bounding box surrounding the defect.

[0143] The AI ​​model training unit (110) can receive a training dataset from an external source and store it in an internal storage.

[0144] The AI ​​model training unit (110) can generate an external defect detection model by training a pre-trained model (G1), which is already publicly known, to be suitable for the domain through a transfer learning technique based on the training dataset (F1, F2, F3). For example, various deep learning models such as the EfficientDet model, YOLO, and Faster-RCNN can be used as the pre-trained model. That is, the pre-trained model may include any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof. The detailed structure of the pre-trained model listed above is already known, so its description is omitted.

[0145] As shown in the welding result image (F4) after inference in Fig. 10, if the learned artificial intelligence model infers that the weld sample is normal (H2) even though there are pores (H1) in the weld sample, the model may need to be updated.

[0146] When an external defect detection model generated through learning receives image data as input, it outputs information on the region of existence of a surface weld bead or external defect, an estimated label for the region, and a confidence score. For example, the external defect detection model receives image data of a weld result and outputs bounding box information for the external defect, an estimated label for the bounding box region (type of the external defect), and a confidence score. The defect detection unit (130) detects a predefined external defect using the external defect detection model based on the weld result image.

[0147] FIG. 11 is an example diagram illustrating the process of inferring external defects using an external defect detection model based on a test dataset. As described above, the defect detection unit (130) detects predefined external defects using an external defect detection model based on a welding result image. As illustrated in FIG. 11, the defect detection unit (130) can input a welding result image (F5) into an external defect detection model (G1') to obtain an inference result (H3). The inference result includes defect type ('normal' in FIG. 11) and bounding box information.

[0148] The defect detection unit (130) inputs new vision image data (image of a welded result) into a pre-trained external defect detection model. The external defect detection model outputs the location of the defect within the image (e.g., bounding box information) and the type of the defect.

[0149] FIGS. 12a and 12b are diagrams illustrating examples of inference results of an external defect detection model. FIG. 12a is an image in which the inference results are displayed on a welding result image (F5). The bounding box information in the inference results (H3) includes the pixel coordinates (xmin) and ymin of the x-axis of the top-left vertex (P1) of the bounding box, and the pixel coordinates (xmax) and ymax of the x-axis of the bottom-right vertex (P2) of the bounding box. J1 in FIG. 12b is an example of an inference result, and the inference result (J1) includes a label (defect type), a confidence score, and pixel location (coordinate) information (xmin, ymin, xmax, ymax) of the bounding box.

[0150] FIG. 13 is a block diagram showing a computer system for implementing an automatic external welding defect inspection method according to an embodiment of the present invention.

[0151] Referring to FIG. 13, a computer system (1000) may include at least one of a processor (1010), memory (1030), an input interface device (1050), an output interface device (1060), and a storage device (1040) that communicate via a bus (1070). The computer system (1000) may also further include a communication device (1020) coupled to a network. The processor (1010) may be a central processing unit (CPU) or a semiconductor device that executes instructions stored in memory (1030) or a storage device (1040). Memory (1030) and storage device (1040) may include various forms of volatile or non-volatile storage media. For example, memory may include read-only memory (ROM) and random access memory (RAM). In the embodiments of this description, memory may be located inside or outside the processor, and memory may be connected to the processor through various known means. Memory is a volatile or non-volatile storage medium of various forms, and for example, memory may include read-only memory (ROM) or random access memory (RAM).

[0152] Accordingly, embodiments of the present invention may be implemented as a method implemented on a computer or as a non-transient computer-readable medium storing computer-executable instructions. In one embodiment, when executed by a processor, the computer-readable instructions may perform a method according to at least one aspect of the present description.

[0153] The communication device (1020) can transmit or receive wired or wireless signals.

[0154] In addition, the method according to an embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and may be recorded on a computer-readable medium.

[0155] The above computer-readable medium may include program instructions, data files, data structures, etc., either individually or in combination. The program instructions recorded on the computer-readable medium may be specially designed and configured for embodiments of the present invention, or they may be known and available to a person skilled in the art of computer software. The computer-readable recording medium may include a hardware device configured to store and execute program instructions. For example, the computer-readable recording medium may be magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; ROM; RAM; flash memory, etc. The program instructions may include not only machine code, such as that generated by a compiler, but also high-level language code that can be executed by a computer through an interpreter, etc.

[0156] For reference, the components according to the embodiments of the present invention may be implemented in the form of software or hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit), and may perform certain roles.

[0157] However, 'components' are not limited to software or hardware, and each component may be configured to reside in an addressable storage medium or configured to operate one or more processors.

[0158] Accordingly, as an example, components include components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables.

[0159] Components and the functions provided within them can be combined into a smaller number of components or further separated into additional components.

[0160] At this time, it will be understood that each block of the process flow diagrams and combinations of the flow diagrams can be executed by computer program instructions. Since these computer program instructions can be loaded into the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create means to perform the functions described in the flow diagram block(s). Since these computer program instructions may be stored in computer-readable memory or may be directed toward a computer or other programmable data processing equipment to implement the functions in a specific manner, the instructions using the computer or stored in computer-readable memory may also produce a manufactured item containing the means of instruction to perform the functions described in the flow diagram block(s). Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0161] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). It should also be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order. For instance, two blocks described in succession may actually be executed substantially simultaneously, or the blocks may be executed in reverse order according to their corresponding functions.

[0162] In this embodiment, the term "part" refers to a software or hardware component, such as an FPGA or ASIC, and the "part" performs 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 configured to operate one or more processors. Thus, as an example, the "part" includes components such as software components, object-oriented software components, class components, and task components, 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 functions provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts." Furthermore, the components and "parts" may be implemented to operate one or more CPUs within a device or secure multimedia card.

[0163] Although the present invention has been described above with reference to preferred embodiments, those skilled in the art will understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the following claims. Explanation of the symbols

[0164] 100: Automatic external welding defect inspection device 110: AI model training unit 120: Video Acquisition Unit 130: Defect detection unit 140: Welding Quality Judgment Unit 150: Distance measuring unit 160: Size estimation section 1000: Computer System 1010: Processor 1020: Communication device 1030: Memory 1040: Storage device 1050: Input interface device 1060: Output interface device 1070: Bus

Claims

Claim 1 An automatic external welding defect inspection device comprising: an image acquisition unit for acquiring image data of a welding result; a defect detection unit for detecting external defects of the welding result using an artificial intelligence-based external defect detection model based on the image data of the welding result; a size estimation unit; and a welding quality judgment unit for determining the quality of the welding result based on the type and size estimation value of the external defect, wherein the defect detection unit generates the type and bounding box information of the external defect using the external defect detection model based on the image data of the welding result, and the size estimation unit generates the size estimation value of the external defect using a previously generated size conversion formula based on the bounding box information of the external defect. Claim 2 An automatic external welding defect inspection device according to claim 1, wherein the types of external defects are classified according to the shape characteristics of the external defects. Claim 3 An automatic external welding defect inspection device according to claim 1, wherein the welding quality judgment unit determines whether the type of external defect corresponds to a defect requiring consideration of size characteristics, and if the type of external defect corresponds to a defect requiring consideration of size characteristics, determines the quality of the welded product based on the size estimate value. Claim 4 delete Claim 5 An automatic external welding defect inspection device according to claim 1, wherein the defect detection unit generates bounding box information of the welding sample using the external defect detection model based on image data of the welding sample acquired by the image acquisition unit, and the size estimation unit generates the size transformation formula using a regression analysis algorithm based on the bounding box information of the welding sample, the distance measurement between the image acquisition unit and the welding sample, the focal length of the camera lens included in the image acquisition unit, and the actual size of the defect included in the welding sample. Claim 6 In paragraph 5, the above regression analysis algorithm is Support Vector Regression, an automatic external welding defect inspection device. Claim 7 In claim 5, the size estimation unit generates a size estimation value of the weld sample using the size conversion formula based on the bounding box information of the weld sample, the distance measurement value, and the focal length, generates a size correction formula using a regression analysis algorithm based on the size estimation value of the weld sample and the actual size of the defect included in the weld sample, and corrects the size estimation value of the external defect using the size correction formula, thereby forming an automatic inspection device for external weld defects. Claim 8 In claim 7, the size estimation unit generates the size correction formula using a linear regression analysis technique in an automatic external welding defect inspection device. Claim 9 An automatic welding external defect inspection device according to claim 1, further comprising an AI model learning unit, wherein the AI ​​model learning unit trains an artificial intelligence model based on a learning dataset to generate the external defect detection model. Claim 10 In claim 9, the AI ​​model learning unit generates the external defect detection model by training a pre-trained model using a transfer learning technique based on the training dataset, in an automatic welding external defect inspection device. Claim 11 An automatic external welding defect inspection device according to claim 9, wherein the training dataset comprises a plurality of welding sample images and annotation files. Claim 12 An automatic external welding defect inspection device according to claim 10, wherein the above-mentioned pre-trained model comprises any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof. Claim 13 The method comprises: a step of acquiring image data of a welded product; a step of detecting external defects of the welded product using an artificial intelligence-based external defect detection model based on the image data of the welded product; and a step of determining the quality of the welded product based on the detection result of the external defects, wherein the step of determining the quality of the welded product includes: a step of determining whether external defects are detected in the welded product; a step of determining the quality of the welded product as normal if no external defects are detected in the welded product; a step of determining whether the type of external defect corresponds to a defect requiring consideration of size characteristics if external defects are detected in the welded product; a step of determining the quality of the welded product as defective if the type of external defect does not correspond to a defect requiring consideration of size characteristics; a step of calculating an estimated size value of the external defect if the type of external defect corresponds to a defect requiring consideration of size characteristics; and a step of determining the quality of the welded product as defective if the estimated size value is greater than or equal to a set threshold, wherein the step of calculating the estimated size value includes: a step of generating bounding box information of the external defect using the external defect detection model based on the image data of the welded product. A method for automatically inspecting external welding defects, comprising the step of calculating an estimated size value using a size conversion formula generated based on bounding box information of the external defect, a distance measurement between a vision camera and the weld result, and the lens focal length of the vision camera, and correcting the estimated size value using a size correction formula generated. Claim 14 A method for automatically inspecting external welding defects according to claim 13, further comprising the step of training an artificial intelligence model based on a learning dataset to generate the external defect detection model. Claim 15 delete Claim 16 In claim 13, the type of external defect is classified according to the shape characteristics of the external defect in the automatic inspection method for external welding defects. Claim 17 delete Claim 18 A method for automatically inspecting external weld defects according to claim 13, further comprising the step of generating bounding box information of the weld sample using the external defect detection model based on image data of the weld sample, and generating the size transformation formula using a regression analysis algorithm based on the bounding box information of the weld sample, a distance measurement between the vision camera and the weld sample, the lens focal length of the vision camera, and the actual size of the defect included in the weld sample. Claim 19 A method for automatically inspecting external welding defects according to claim 13, further comprising the step of generating the external defect detection model by training a pre-trained model using a transfer learning technique based on a learning dataset. Claim 20 A method for automatically inspecting external weld defects according to claim 19, wherein the training dataset comprises a plurality of weld sample images and annotation files, and the pre-trained model comprises any one of the EfficientDet model, YOLO, and Faster-RCNN, or a combination thereof.

Images