Continuous casting slab quality defect rating method and related device

Abstract

The invention belongs to the field of intellectualization of the continuous casting process of ferrous metallurgy, and discloses a continuous casting slab quality defect rating method and a related device. Processing the continuous casting billet finished product section low-power image through a trained U-Net semantic segmentation neural network to obtain a reconstructed image of a continuous casting billet effective area in the continuous casting billet finished product section low-power image; and processing the reconstructed image by using a trained defect detection model to obtain the defect type and grade of the section of the finished continuous casting billet. According to the method, the defects of the continuous casting blank can be accurately recognized, accurate monitoring of the quality of the casting blank is achieved, and the intelligent level of the continuous casting process is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent continuous casting process in iron and steel metallurgy, and specifically relates to a method and related device for rating the quality defects of continuously cast billets. Background Technology

[0002] Continuous casting technology is a crucial step in modern metallurgical production and a primary means of producing various long products, profiles, and plates. Continuously cast billets are fundamental raw materials used in the automotive, construction machinery, and home appliance industries. In addition to meeting cleanliness standards, their internal quality is subject to stringent requirements. Quality defects in continuously cast billets mainly include cracks, shrinkage cavities, porosity, and segregation, which severely affect subsequent hot working adaptability and yield.

[0003] The identification of quality defects in continuously cast billets currently relies primarily on experienced operators using conventional methods such as visual inspection, tactile inspection, staining and penetrating detection, scratch inspection, and localized heating. However, this manual inspection method is highly dependent on the operator's personal experience, skill level, and attention span, leading to significant differences in judgment standards and sensitivities among different personnel, resulting in subjective and inconsistent assessment results. Furthermore, manual inspection is inefficient for large, complex, or high-speed production billets, making it difficult to conduct a comprehensive and detailed inspection of all critical areas within a limited timeframe. This easily creates blind spots and the risk of missed inspections, and there is also a time lag in quality information feedback, allowing defective billets to enter subsequent processes, increasing rework costs. In addition, manual inspection often struggles to accurately quantify defect characteristics and establish systematic records, hindering quality traceability and analysis for improvement. In summary, the subjectivity, inefficiency, lag, and insufficient ability to identify complex defects in manual inspection have become bottlenecks restricting the stable improvement of continuously cast billet quality and the realization of intelligent production.

[0004] To this end, those skilled in the art have designed various prediction algorithms for application in metallurgical processes, as illustrated below: Chinese patent CN116452505A discloses a method for detecting and rating internal defects in continuously cast billets based on an improved YOLOv5 model. The method involves: establishing a dataset of continuously cast billet defects; building an improved YOLOv5 deep learning neural network detection model, inserting the CBAM module after the C3 structure; training the dataset using the improved YOLOv5 deep learning neural network detection model; inputting images of defects in the billet to be tested into the final improved CBAM-YOLOv5 model to obtain the detection results and target location information of the continuously cast billet and its defects; statistically comparing the length and width of the defects with the length and width of the detected billet to obtain the aspect ratio of the segregated regions for different evaluation levels; and using Lagrange interpolation to divide the whole levels into half levels.

[0005] Chinese patent CN117009586A discloses a method for tracing the quality of continuously cast billets based on decision trees and knowledge graphs. It uses a decision tree model to summarize the process parameters involved in steel production, constructs a knowledge graph of steel quality defects based on the Neo4j graph database, and designs a model for quality tracing based on this knowledge graph. This reveals the potential and influencing factors leading to quality problems, helping to improve steel quality control and production efficiency.

[0006] Chinese Patent Publication No. CN109191459A discloses an automatic identification and rating method for center segregation defects in low-magnification microstructure of continuously cast billets. The method includes: preprocessing a grayscale image of the low-magnification microstructure of the continuously cast billet, including distortion correction, white background removal, and filtering; segmenting the grayscale image to obtain a central region image, targeting the central region where the horizontal and vertical regions of the grayscale image overlap; detecting and labeling connected regions in the corresponding binary image of the central region image to separate suspected center segregation regions from crack regions within the connected regions; extracting features from suspected center segregation regions to identify and remove interfering regions; and training a BP neural network classifier model to identify and rate the central segregation regions.

[0007] Chinese patent CN118347804A discloses an automatic identification and grading system for low-magnification defects in continuously cast billets, comprising a machining module, a cleaning module, an corrosion module, a cleaning and drying module, a photographing module, a communication module, an image recognition module, and an image uploading module, which are connected sequentially. The machining module, cleaning module, corrosion module, and cleaning and drying module are connected to a first industrial control computer, while the photographing module, communication module, image recognition module, and image uploading module are connected to a second industrial control computer. The image recognition module acts as an image recognition server, using machine vision recognition technology and expert knowledge to grade the low-magnification microstructure images. The image uploading module uploads the grading results to an L2 server and an L3 quality control system.

[0008] Chinese patent CN118347804A discloses an automatic identification and grading system for low-magnification defects in continuously cast billets, comprising a machining module, a cleaning module, an corrosion module, a cleaning and drying module, a photographing module, a communication module, an image recognition module, and an image uploading module, which are connected sequentially. The machining module, cleaning module, corrosion module, and cleaning and drying module are connected to a first industrial control computer, while the photographing module, communication module, image recognition module, and image uploading module are connected to a second industrial control computer. The image recognition module acts as an image recognition server, using machine vision recognition technology and expert knowledge to grade the low-magnification microstructure images. The image uploading module uploads the grading results to an L2 server and an L3 quality control system.

[0009] Chinese patent CN118351044A discloses a method for intelligent identification of low-magnification defects in continuously cast billets based on multi-model fusion. The steps are as follows: 1) Establish an image dataset; 2) Calculate the total forward loss of a deep network; 3) Apply a knowledge graph network refined from expert knowledge to the digital image feature extraction process to assist in establishing a reference model for identifying bypass branches; 4) Calculate the first loss (Loss1) using the network prediction probability label of this reference model and the probability label calculated by the deep learning network; 5) Calculate the second loss (Loss2) using the probability label calculated by the deep learning network and the label of the original image supervision data; 6) Calculate the total loss (Total Loss) through weighted summation; 7) Iteratively calculate and update the weights of each deep neural network; 8) When the loss function value reaches its minimum, remains unchanged, or the accuracy no longer improves, the entire training and learning process is completed; 9) The finally learned deep network model parameters are applied to the actual verification process.

[0010] Chinese patent publication number 116452505 A establishes a continuous casting billet defect dataset; establishes an improved YOLOv5 deep learning neural network detection model, improves the YOLOv5 backbone network, and inserts the CBAM module after the C3 structure; trains the dataset using the improved YOLOv5 deep learning neural network detection model; inputs the defect image of the billet to be tested into the final improved CBAM-YOLOv5 model to obtain the detection results and target location information of the continuous casting billet and its defects; statistically compares the length and width of the defects with the length and width of the detected billet to obtain the length-to-width ratio of the segregated region for different evaluation levels; and uses Lagrange interpolation to divide the whole level into half levels.

[0011] However, the aforementioned existing technologies take into account the influence of surrounding factors on the continuously cast billet less, resulting in subjective factors affecting the quality defect rating of the continuously cast billet and making it insufficient in its ability to identify complex defects. Summary of the Invention

[0012] To address the problems existing in the prior art, the present invention aims to provide a method and related apparatus for rating the quality defects of continuously cast billets. The present invention can accurately identify defects in continuously cast billets, realize accurate monitoring of billet quality, and improve the level of intelligence in the continuous casting process.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: A method for rating quality defects in continuously cast billets includes the following process: Obtain a low-magnification image of the cross-section of the finished continuously cast billet; The low-magnification image of the cross-section of the finished continuous casting billet is processed by the trained U-Net semantic segmentation neural network to obtain a reconstructed image of the effective area of ​​the continuous casting billet in the low-magnification image of the cross-section of the finished continuous casting billet. The reconstructed image is processed using a pre-trained defect detection model to obtain the defect type and level of the cross-section of the continuously cast billet.

[0014] Preferably, the training process of the U-Net semantic segmentation neural network includes: In Label-me software, a polygonal annotation box is drawn on the low-magnification image of the cross-section of the continuous casting billet to select the effective area of ​​the continuous casting billet, thus obtaining the effective area selection diagram of the continuous casting billet. Using the low-magnification image of the finished continuous casting billet cross-section as input and the effective area selection map of the continuous casting billet as output, the U-Net semantic segmentation neural network is trained to obtain the trained U-Net semantic segmentation neural network.

[0015] Preferably, the training process of the U-Net semantic segmentation neural network further includes: The U-Net semantic segmentation neural network is quantitatively evaluated using three key evaluation metrics. If all three metrics meet the requirements, the evaluation is considered successful. Otherwise, the U-Net semantic segmentation neural network is retrained and quantitatively evaluated until it passes the evaluation, thus obtaining the final trained U-Net semantic segmentation neural network. When quantitatively evaluating the U-Net semantic segmentation neural network, three key evaluation metrics are used: the intersection-union ratio (IUU) is used to measure the accuracy of overlapping segments, the Dice coefficient is used to evaluate the consistency of segmentation contours, and the Hausdorff distance is used to quantify the maximum deviation of the segmentation boundary.

[0016] Preferably, the training process for the defect detection model includes: Defects are annotated on the reconstructed image to obtain a defect-annotated image; The defect detection model is trained using the defect-annotated images to obtain a trained defect detection model.

[0017] Preferably, the defect detection model adopts the YOLOv8 target detection model. When training the YOLOv8 target detection model, the defect-annotated image is input into the YOLOv8 target detection model. The YOLOv8 target detection model identifies the defects in the defect-annotated image, and then rates the defects of the continuous casting billet according to the defect type and rating criteria, and then outputs the defect type and rating result.

[0018] Preferably, when training the YOLOv8 object detection model, precision and recall are used to evaluate the accuracy of the YOLOv8 object detection model, confusion matrix is ​​used to analyze the classification error types, ROC curve and AUC value are used to measure the adaptability of the classification threshold, and Matthews correlation coefficient is used to evaluate the stability of the YOLOv8 object detection model.

[0019] Preferably, the step of annotating defects on the reconstructed image to obtain a defect-annotated image includes the following process: Defects in the reconstructed images output by the U-Net semantic segmentation neural network were labeled using Label-me software, resulting in defect-labeled images.

[0020] The present invention also provides a method for rating quality defects in continuously cast billets, for implementing the method for rating quality defects in continuously cast billets as described above, comprising: Image acquisition unit: used to acquire low-magnification images of the cross-section of the finished continuously cast billet; Image reconstruction unit: used to process the low-magnification image of the cross-section of the finished continuous casting billet using a trained U-Net semantic segmentation neural network to obtain a reconstructed image of the effective area of ​​the continuous casting billet in the low-magnification image of the cross-section of the finished continuous casting billet; Defect identification and rating unit: Used to process the reconstructed image using a trained defect detection model to obtain the defect type and rating of the cross-section of the continuous casting billet.

[0021] The present invention also provides an electronic device, comprising: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the continuous casting billet quality defect rating method of the present invention as described above.

[0022] The present invention also provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the continuous casting billet quality defect rating method of the present invention as described above.

[0023] The present invention has the following beneficial effects: This invention addresses the core problems inherent in manual inspection of continuous casting billet quality defects through a process of "acquiring low-magnification images of the finished cross-section—U-Net semantic segmentation and reconstruction of the effective region—defect detection model rating." Specifically, it addresses the issues of inconsistent judgment standards and subjective results caused by reliance on experience in manual inspection. This invention first standardizes the acquisition of low-magnification images of the finished cross-section of the continuous casting billet, then uses a trained U-Net network to accurately segment the effective region using a preset algorithm and eliminate interference. Finally, the defect detection model determines the defect type and level based on the standardized data. The entire process is algorithm-driven, completely avoiding the influence of subjective human factors and ensuring objective and stable rating results. It also addresses the low efficiency and susceptibility to errors in manual inspection of large casting billets. Visual fatigue leads to missed detections and delayed feedback. Low-magnification images of the finished continuous casting billet cross-section can cover the entire billet cross-section in one go. The processing speed of the U-Net network and defect detection model is far superior to manual methods, and it can achieve full-area scanning without blind spots. This improves detection efficiency, shortens the quality information feedback cycle, and prevents defective billets from flowing into subsequent processes, thus reducing rework costs. Simultaneously, the problem of difficulty in quantifying defect features through manual inspection is also solved. The U-Net network can simultaneously acquire quantitative parameters of the effective area, and the defect detection model can accurately calculate the area, length, and other features of defects. All detection data can be automatically stored to form a database, facilitating quality traceability and providing data support for analyzing defect causes and optimizing the continuous casting process. Furthermore, this invention deeply integrates deep learning technology with the inspection process to build a fully automated inspection system. This system can interface with the production line control system to achieve data linkage and form an intelligent closed loop through model iteration. This drives the quality inspection of continuous casting billets from a manual-dependent mode to an autonomous intelligent mode, effectively breaking through the bottleneck restricting the intelligent upgrading of continuous casting production and providing key technical support for automated production in the steel industry. Attached Figure Description

[0024] Figure 1 This is a technical roadmap for identifying and rating quality defects in continuously cast billets in an embodiment of the present invention.

[0025] Figure 2 This is an illustration of the effect of using the U-Net model to identify and eliminate the influence of the surrounding environment in continuous casting billets with and without defects in an embodiment of the present invention.

[0026] Figure 3 This is a diagram showing the results of identifying defects in continuously cast billets using the YOLOv8 model in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] This invention addresses the issue that manual inspection relies on the operator's experience and judgment. Different workers may have varying standards and sensitivities, leading to inconsistent results. When dealing with large and complex billets, it may be impossible to comprehensively inspect all areas within a short time, increasing time lag and the occurrence of errors. To accurately detect these defects, establishing a billet quality defect identification model for the continuous casting process has become an important technical means to address billet defects. This invention utilizes deep learning to accurately and in real-time detect continuous casting billet defects, providing technical support for the automation and intelligentization of steel production.

[0029] Reference Figure 1 The continuous casting billet quality defect rating method of the present invention specifically includes the following steps: Step (1): After treatment with a specific acid etching method, a low-magnification image of the continuously cast billet (i.e., a macroscopic structural image of the cross-section or longitudinal section of the continuously cast billet, see [link]) is captured using a high-definition camera. Figure 2 The low-magnification images of the continuously cast billet are collected and stored. The high-definition camera used is an industrial camera with a resolution of at least 1280×1280. Images are acquired at the cut surface of the continuously cast billet after cooling, resulting in low-magnification images of the finished product cross-section. Since the continuous casting process is conducted in a high-temperature environment, process data cannot be directly collected. Therefore, this invention only collects image data of the finished cross-section of the continuously cast billet after cooling and stores it in lossless JPEG format.

[0030] Step (2): For the low-magnification image of the finished continuous casting billet cross-section acquired in step (1), since the low-magnification image of the finished continuous casting billet cross-section taken on-site may contain factors such as conveyor rollers, background equipment, and lighting interference, it is necessary to accurately mark the main area of ​​the continuous casting billet on the low-magnification image of the finished continuous casting billet cross-section. This invention uses Label-me software to draw a polygonal annotation box on the low-magnification image of the finished continuous casting billet cross-section, completely selecting the effective area of ​​the continuous casting billet on the low-magnification image of the finished continuous casting billet cross-section, thus obtaining the effective area selection diagram of the continuous casting billet. See... Figure 2 and Figure 3 .

[0031] Step (3) uses the effective area bounding box of the continuously cast billet annotated by Label-me software in step (2) as the output of the U-Net semantic segmentation neural network, and the original image (i.e., the low-magnification image of the cross-section of the finished continuously cast billet taken in step (1)) as the input to train the U-Net semantic segmentation neural network, thus obtaining the trained U-Net semantic segmentation neural network. The input of the U-Net semantic segmentation neural network is the low-magnification image of the cross-section of the finished continuously cast billet taken in step (1), and the output is the bounding box of the effective area of ​​the continuously cast billet (denoted as the reconstructed image of the effective area of ​​the continuously cast billet). This method accurately identifies the position of the continuously cast billet through pixel-level segmentation, which can completely eliminate the interference of background (i.e., environmental factors) and ensure that only the image of the continuously cast billet body is retained in the image (i.e., the reconstructed image of the effective area of ​​the continuously cast billet) in subsequent processing.

[0032] Step (4) selects three evaluation metrics to quantitatively evaluate the reliability and stability of the U-Net semantic segmentation neural network: namely, using the intersection-union ratio to measure the accuracy of overlapping segments, using the Dice coefficient to evaluate the consistency of segments, and using the Hausdorff distance to quantify the maximum deviation of segments. These three metrics can comprehensively verify the accuracy and robustness of the model (i.e., the U-Net semantic segmentation neural network). If all three metrics meet the requirements, the evaluation is qualified, and the U-Net semantic segmentation neural network at this time can be used as the final trained U-Net semantic segmentation neural network, and then proceed to step (5); otherwise (i.e., if at least one of the three metrics does not meet the requirements), the evaluation is unqualified, and step (3) is repeated for training, and then the evaluation is repeated until the evaluation is qualified.

[0033] Step (5): Use Label-me software to annotate the defects in the reconstructed image of the effective region of the continuous casting billet output by the U-Net semantic segmentation neural network to obtain a defect-annotated image; Step (6): Use the defect-annotated images obtained in step (5) to train the YOLOv8 target detection model to obtain the trained YOLOv8 target detection model. Use the trained YOLOv8 target detection model as the defect detection model. During training, use precision and recall to evaluate the accuracy of the YOLOv8 target detection model, use confusion matrix to analyze the classification error type, use ROC curve and AUC value to measure the adaptability of classification threshold, and use Matthews correlation coefficient to comprehensively evaluate the stability of the YOLOv8 target detection model. Step (7) involves inputting the reconstructed image of the effective region of the continuously cast billet output by the U-Net semantic segmentation neural network into the improved and trained YOLOv8 object detection model. This YOLOv8 object detection model first extracts multi-scale feature information through the backbone and neck networks. Subsequently, in the detection head, it not only predicts bounding boxes through regression and outputs defect categories (such as center segregation, subcutaneous bubbles, and shrinkage cavities) through a classification branch, but also adds a dedicated rating branch. This rating branch, based on the identified defect type, calls the built-in rating rule module (whose rules originate from the national standard "YBT4002-2013 Low-Magnification Structure Defect Rating Map of Continuously Cast Steel Billet") to evaluate each detected defect, ultimately outputting the defect type, location, and corresponding rating simultaneously. All rating results are automatically recorded and reports are generated, achieving automated intelligent judgment of the continuously cast billet quality.

[0034] The trained U-Net semantic segmentation neural network and the YOLOv8 object detection model were integrated and packaged into an executable module. End-to-end testing was conducted using independently acquired on-site continuous casting billet image data to verify the accuracy of defect identification and rating results. After passing the tests, the model was deployed to the production environment, confirming that it can support online quality inspection of the continuous casting process.

[0035] Example The continuous casting billet quality defect rating method in this embodiment includes the following steps: Low-magnification images of 20# steel continuously cast billets (210mm diameter) from a steel plant over the past year were collected, totaling 330 images. The effective regions of the continuously cast billets were selected using Label-me software. A U-Net model was then trained, with the input being the collected low-magnification images of the continuously cast billets and the output being the reconstructed images of the effective regions. The reliability and accuracy of the model were evaluated using the intersection-over-union ratio (IoU), Dice coefficient, and Hausdorff regression. After meeting the criteria, defects in the reconstructed images of the effective regions of the continuously cast billets output by the U-Net semantic segmentation neural network were labeled using Label-me software, resulting in defect-labeled images. A YOLOv8 object detection model was then trained, with the input being the U-Net semantic segmentation neural network. The reconstructed image of the effective region of the continuously cast billet output by the segmentation neural network is used to label quality defects in the reconstructed image of the effective region of the continuously cast billet output by the U-Net semantic segmentation neural network using Label-me software. A YOLOv8 object detection model is obtained after model training. This trained YOLOv8 object detection model is used as the defect detection model. To incorporate continuous casting billet quality defect rating, the YOLOv8 object detection model is improved. First, multi-scale feature information is extracted through the backbone network and neck network. Then, in the detection head, in addition to predicting bounding boxes through regression and outputting defect categories (such as center segregation, subcutaneous bubbles, shrinkage cavities, etc.) through a classification branch, a dedicated rating branch is added. This rating branch calls the built-in rating rule module (whose rules are derived from the national standard "YBT 4002-2013 Low-magnification microstructure defect rating map of continuously cast steel billet") based on the identified defect type to evaluate each detected defect, and finally outputs the defect type, location, and corresponding rating simultaneously. The trained U-Net semantic segmentation neural network and the YOLOv8 object detection model are integrated and packaged into an executable module. Image data of continuously cast billets acquired independently in the field (see...) Figure 2 This includes the original image, masked image, predicted source code image, and reconstructed image. The reconstructed image can effectively identify the position of the continuously cast billet, removing the influence of scales, backgrounds, etc. End-to-end testing can eliminate the influence of environmental factors surrounding the continuously cast billet, accurately identifying and rating defects such as center cracks, shrinkage cavities, center segregation, and porosity (see...). Figure 3 This invention improves mAP by 21.3% compared to the unmodified YOLOv8 network.

[0036] In the above-described solution of this invention, addressing the issue that existing technologies fail to eliminate the influence of factors from the actual production environment, the continuous casting billet quality inspection system deployed on-site first uses high-definition cameras to acquire low-magnification images of the continuous casting billet. These raw images are directly stored on a GPU workstation equipped with a large-capacity solid-state drive. When new images are acquired, operators use Label-me software to precisely select the main area of ​​the continuous casting billet, constructing a high-quality dataset for model training. The trained U-Net segmentation model automatically removes environmental interference from the images, outputting clean images of the continuous casting billet. The processed images are then input into the YOLOv8 model for defect identification. This model strictly adheres to the YBT 4002-2013 national standard for defect classification and rating, while generating an evaluation report including indicators such as precision and recall. Finally, the integrated U-Net+YOLOv8 model package is deployed on the production line after validation with new on-site data, achieving end-to-end automated processing from raw image input to defect rating report. The continuous casting billet quality defect rating method provided by this invention can accurately identify defects in continuous casting billets, achieve accurate monitoring of billet quality, and improve the intelligence level of the continuous casting process.

[0037] Furthermore, embodiments of the present invention also provide a system for implementing the above-described continuous casting billet quality defect rating method of the present invention, the system comprising: Image acquisition unit: used to acquire low-magnification images of the cross-section of the finished continuously cast billet; Image reconstruction unit: used to process the low-magnification image of the cross-section of the finished continuous casting billet using a trained U-Net semantic segmentation neural network to obtain a reconstructed image of the effective area of ​​the continuous casting billet in the low-magnification image of the cross-section of the finished continuous casting billet; Defect identification and rating unit: Used to process the reconstructed image using a trained defect detection model to obtain the defect type and rating of the cross-section of the continuous casting billet.

[0038] The embodiments of the present invention also provide corresponding electronic devices and computer-readable storage media for implementing the solutions provided in the embodiments of the present invention.

[0039] The electronic device includes a storage device and one or more processors. The storage device stores instructions or code, and the processors execute the instructions or code to enable the device to perform the continuous casting billet quality defect rating method according to any embodiment of this application.

[0040] The storage medium stores a computer program, which, when executed by a processor, implements the continuous casting billet quality defect rating method described in any embodiment of this application.

[0041] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for rating quality defects in continuously cast billets, characterized in that, The process includes the following: Obtain a low-magnification image of the cross-section of the finished continuously cast billet; The low-magnification image of the cross-section of the finished continuous casting billet is processed by the trained U-Net semantic segmentation neural network to obtain a reconstructed image of the effective area of ​​the continuous casting billet in the low-magnification image of the cross-section of the finished continuous casting billet. The reconstructed image is processed using a pre-trained defect detection model to obtain the defect type and level of the cross-section of the continuously cast billet.

2. The method for rating quality defects in continuously cast billets according to claim 1, characterized in that, The training process of the U-Net semantic segmentation neural network includes: In Label-me software, a polygonal annotation box is drawn on the low-magnification image of the cross-section of the continuous casting billet to select the effective area of ​​the continuous casting billet, thus obtaining the effective area selection diagram of the continuous casting billet. Using the low-magnification image of the finished continuous casting billet cross-section as input and the effective area selection map of the continuous casting billet as output, the U-Net semantic segmentation neural network is trained to obtain the trained U-Net semantic segmentation neural network.

3. The method for rating quality defects in continuously cast billets according to claim 2, characterized in that, The training process of the U-Net semantic segmentation neural network also includes: The U-Net semantic segmentation neural network is quantitatively evaluated using three key evaluation metrics. If all three metrics meet the requirements, the evaluation is considered successful. Otherwise, the U-Net semantic segmentation neural network is retrained and quantitatively evaluated until it passes the evaluation, thus obtaining the final trained U-Net semantic segmentation neural network. When quantitatively evaluating the U-Net semantic segmentation neural network, three key evaluation metrics are used: the intersection-union ratio (IUU) is used to measure the accuracy of overlapping segments, the Dice coefficient is used to evaluate the consistency of segmentation contours, and the Hausdorff distance is used to quantify the maximum deviation of the segmentation boundary.

4. The method for rating quality defects in continuously cast billets according to claim 1, characterized in that, The training process for the defect detection model includes: Defects are annotated on the reconstructed image to obtain a defect-annotated image; The defect detection model is trained using the defect-annotated images to obtain a trained defect detection model.

5. The method for rating quality defects in continuously cast billets according to claim 1, characterized in that, The defect detection model adopts the YOLOv8 target detection model. When training the YOLOv8 target detection model, the defect-annotated image is input into the YOLOv8 target detection model. The YOLOv8 target detection model identifies the defects in the defect-annotated image, and then rates the defects of the continuous casting billet according to the defect type and rating criteria. Finally, it outputs the defect type and rating result.

6. The method for rating quality defects in continuously cast billets according to claim 5, characterized in that, When training the YOLOv8 object detection model, precision and recall are used to evaluate the accuracy of the YOLOv8 object detection model, confusion matrix is ​​used to analyze the classification error types, ROC curve and AUC value are used to measure the classification threshold adaptability, and Matthews correlation coefficient is used to evaluate the stability of the YOLOv8 object detection model.

7. The method for rating quality defects in continuously cast billets according to claim 4, characterized in that, The process of annotating defects on the reconstructed image to obtain a defect-annotated image includes the following steps: Defects in the reconstructed images output by the U-Net semantic segmentation neural network were labeled using Label-me software, resulting in defect-labeled images.

8. A method for rating quality defects in continuously cast billets, characterized in that, A method for implementing the continuous casting billet quality defect rating method according to any one of claims 1-7 includes: Image acquisition unit: used to acquire low-magnification images of the cross-section of the finished continuously cast billet; Image reconstruction unit: used to process the low-magnification image of the cross-section of the finished continuous casting billet using a trained U-Net semantic segmentation neural network to obtain a reconstructed image of the effective area of ​​the continuous casting billet in the low-magnification image of the cross-section of the finished continuous casting billet; Defect identification and rating unit: Used to process the reconstructed image using a trained defect detection model to obtain the defect type and rating of the cross-section of the continuous casting billet.

9. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the continuous casting billet quality defect rating method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the continuous casting billet quality defect rating method as described in any one of claims 1-7.

Images