Bearing steel band carbide intelligent detection rating method and system
By using a YOLOv8-based banded carbide detection model, the problem of relying on manual experience for the detection of banded carbides in bearing steel has been solved, achieving automated rating with high accuracy and consistency, which is suitable for efficient detection and quality control in modern industry.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU SUPERVISION & INSPECTION INST FOR PROD QUALITY
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-29
Smart Images
Figure CN122115338A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an intelligent detection and rating method and system for banded carbides in bearing steel, belonging to the field of banded carbides detection and rating. Background Technology
[0002] Currently, semi-quantitative analysis is commonly performed both domestically and internationally using a method that compares the sample with standard rating charts. The main standards include: Chinese National Standard GB / T 18254-2016 "High Carbon Chromium Bearing Steel": This is the most important fundamental standard, and its appendix clearly specifies the microscopic evaluation method for banded carbides in bearing steel and provides a standard rating chart (usually divided into 1-4 levels, with higher numbers indicating more severe banding). Specialized standards such as YB / T 4146-2016 "Continuously Cast Bearing Steel" also contain corresponding provisions. The core of the current semi-quantitative analysis method is: preparing high-quality longitudinal samples (parallel to the rolling direction) → etching with a suitable etchant (such as 4% nitric acid alcohol solution) → observing under a metallographic microscope → comparing the most severe field of view with the standard chart to determine the level.
[0003] The above method has the following problems:
[0004] 1. Traditional manual visual rating methods rely heavily on the experience and subjective judgment of inspectors. There may be deviations in detection and evaluation between different operators, resulting in low consistency and accuracy of results and excessive subjective dependence.
[0005] 2. Insufficient quantification: Comparative spectroscopy is a semi-quantitative method that cannot provide precise digital parameters (such as bandwidth, band gap, carbide area fraction distribution, etc.), making it difficult to meet the needs of higher-precision research and quality control. Summary of the Invention
[0006] The purpose of this invention is to provide an intelligent detection and rating method and system for bearing steel strip carbides to solve the problems of large deviations in detection and evaluation, and the need to improve the objectivity and accuracy of rating results in the existing technology.
[0007] The technical solution of this invention is:
[0008] A method for intelligent detection and rating of banded carbides in bearing steel includes the following steps:
[0009] S1. Obtain the original bearing steel image, classify and label the banded carbide regions to obtain the image training dataset;
[0010] S2. Construct a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image.
[0011] S3. After training the YOLOv8-based banded carbide detection model using the image training dataset and the loss function, the trained detection model is obtained.
[0012] S4. Input the image to be detected into the trained detection model to obtain the banded carbide image;
[0013] S5. After calculating the length, width, and sparsity of the banded carbide image, generate the corresponding grading results according to the set rules.
[0014] Further, in step S2, the preprocessing module preprocesses the input image to obtain the image to be recognized, specifically,
[0015] S21. Perform histogram equalization on the input image to obtain the histogram equalized image;
[0016] S22. After histogram equalization, the image is enhanced with a directional filter to improve the lateral features of the banded carbides. After adaptive anchor box adjustment and directional non-maximum suppression, the image is binarized to obtain the image to be identified.
[0017] Furthermore, in step S2, the image segmentation module uses either the DeeplabV3+ segmentation model or the U-net segmentation model.
[0018] Furthermore, in step S4, the loss functions adopted are cross-entropy loss function CE Loss, focal loss function FocalLoss, and auxiliary loss function Auxiliary Loss.
[0019] Further, in step S5, the sparsity calculation is specifically as follows:
[0020] 1) Length and width calculation: For the strip carbide image obtained in step S4, i.e. the binary segmentation mask of the strip carbide, the edge contour of the strip carbide is extracted using the contour extraction algorithm; the minimum bounding rectangle is fitted to the extracted edge contour of the strip carbide, and the length of the long side of the minimum bounding rectangle is recorded as the length of the strip carbide, and the length of the short side is recorded as the width of the strip carbide.
[0021] 2) Sparsity calculation: Count the total number N pixels N representing carbides (i.e., pixels with a value of 1) within the smallest bounding rectangle. carbide Simultaneously calculate the total pixel area N of the minimum bounding rectangle. total Calculate the area ratio of banded carbides P=N carbide / N total The sparsity is defined as 1-P.
[0022] A system employing the intelligent detection and rating method for banded carbides in bearing steel as described above includes an image acquisition module, a model building module, a model training module, an image detection module, and an image rating module.
[0023] Image acquisition module: acquires raw bearing steel images, performs grading and labels the banded carbide regions, and obtains the image training dataset;
[0024] Model building module: Constructs a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image.
[0025] Model training module: The YOLOv8-based banded carbide detection model is trained using an image training dataset and a loss function to obtain the trained detection model.
[0026] Image detection module: Inputs the image to be detected into the trained detection model to obtain the image of the banded carbides;
[0027] Image rating module: After calculating the length, width and sparsity of the banded carbide image, the rating result is generated according to the set rules.
[0028] The beneficial effects of this invention are: the intelligent detection and rating method and system for banded carbides in bearing steel can achieve high accuracy and objectivity in detecting and rating banded carbides in bearing steel, greatly improving detection efficiency. Using a banded carbide detection model based on YOLOv8, the target detection accuracy can reach greater than 95%, and the image segmentation accuracy can reach greater than 85%, thereby improving the objectivity and accuracy of the rating results. Attached Figure Description
[0029] Figure 1 This is a schematic flowchart of the intelligent detection and rating method for banded carbides in bearing steel according to an embodiment of the present invention;
[0030] Figure 2This is a schematic diagram of the original bearing steel image in the embodiment;
[0031] Figure 3 This is a schematic diagram of the original bearing steel image after histogram equalization in the embodiment.
[0032] Figure 4 This is a schematic diagram of the binarized image in the embodiment;
[0033] Figure 5 This is a schematic diagram of the image to be identified in the embodiment;
[0034] Figure 6 This is a schematic diagram illustrating the intelligent detection and rating system for strip carbides in bearing steel, as described in the embodiment. Detailed Implementation
[0035] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0036] The embodiment provides an intelligent detection and rating method for banded carbides in bearing steel, such as... Figure 1 This includes the following steps:
[0037] S1. Obtain the original bearing steel image, classify and label the banded carbide areas, such as... Figure 2 This yields the image training dataset.
[0038] S2. Construct a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image.
[0039] In step S2, the preprocessing module preprocesses the input image to obtain the image to be recognized. Specifically,
[0040] S21. Perform histogram equalization on the input image as follows: Figure 3 The image obtained after histogram equalization is shown below. Figure 4 ;
[0041] In step S21, histogram equalization enhances contrast by adjusting the gray-level distribution of the image, making the banded carbides in the image more prominent. The cumulative distribution function (CDF) of the image is used to redistribute the gray values, making the brightness distribution of the image more uniform, thereby improving the visual effect and detail of the image. It is particularly suitable for images with a narrow gray-level distribution, significantly enhancing the global contrast of the image and highlighting the characteristics of banded structures.
[0042] S22. After histogram equalization, a directional filter is used to enhance the lateral features of the banded carbides in the image. Adaptive anchor box adjustment and directional non-maximum suppression are then applied. Finally, the image is binarized to obtain the image to be identified, as shown below. Figure 5 .
[0043] In step S22, a Gabor filter is used as the directional filter to enhance specific directional and frequency components in the image. The existing two-dimensional Gabor filter formula is employed here. Based on the Gabor function, it consists of a sinusoidal plane wave multiplied by a Gaussian kernel function, effectively capturing directional features such as edges and textures in the image. By adjusting the filter parameters, including direction, frequency, and scale, specific directional banded features in the image are enhanced, reducing background noise interference. The Gabor filter is particularly suitable for image features with strong directionality, enhancing the lateral features of banded carbides during preprocessing and improving the accuracy of subsequent detection and segmentation.
[0044] In step S2, the YOLOv8 network is used to identify the banded carbides, with an expected accuracy of over 95%. Given that the image to be identified is a grayscale image and the objects to be identified are horizontal stripes, histogram equalization and directional filters are used for image preprocessing. Adaptive anchor box adjustment and directional non-maximum suppression are then employed to further improve accuracy. Using the YOLOv8 network enables real-time detection while maintaining high detection accuracy. YOLOv8 directly regresses the target bounding box and class probability on the input image using a convolutional neural network (CNN), resulting in fast detection speed that meets real-time detection requirements. Through multi-scale feature fusion and residual network structure, the YOLOv8 network effectively improves detection accuracy, especially in complex backgrounds, and is suitable for detecting targets of various sizes and shapes, particularly irregularly shaped targets such as banded carbides. The image segmentation module uses either the DeeplabV3+ segmentation model or the U-net segmentation model. Image segmentation is performed on the bounding boxes obtained from target detection to obtain the specific boundaries of ribbon-like objects. The expected segmentation accuracy is 85%. The DeeplabV3+ segmentation model, based on Spatial Pyramid Pooling (SPP) and an encoder-decoder structure, achieves high-precision image segmentation through multi-scale feature extraction and contextual information fusion. It performs exceptionally well in handling complex backgrounds and fine-grained features, making it highly suitable for the accurate segmentation of ribbon-like carbides. It effectively combines low-level and high-level features to achieve high-resolution, accurate segmentation. Furthermore, its simple structure and high training efficiency make it suitable for segmentation tasks in various medical and industrial images.
[0045] S3. Using the image training dataset and employing a loss function, the YOLOv8-based banded carbide detection model is trained to obtain the trained detection model.
[0046] S4. Input the image to be detected into the trained detection model to obtain the banded carbide image.
[0047] In step S4, the loss functions employed are Cross-Entropy Loss (CE Loss), Focal Loss, and Auxiliary Loss. CE Loss measures the difference between the model's predicted class probability distribution and the true distribution, effectively addressing class imbalance and improving segmentation accuracy. Focal Loss reduces the loss contribution of easily classified samples, focusing on difficult-to-classify samples, making it particularly suitable for handling class imbalance and small object segmentation tasks. Auxiliary Loss adds auxiliary task losses at different levels, improving the model's training performance and generalization ability. It is used for segmentation and detection tasks in deep neural networks, helping the model better learn multi-scale features.
[0048] S5. After calculating the length, width, and sparsity of the banded carbide image, generate the corresponding grading results according to the set rules.
[0049] In step S5, the sparsity is calculated as follows:
[0050] 1) Length and width calculation: For the banded carbide image obtained in step S4, i.e. the binary segmentation mask of the banded carbide, the edge contour of the banded carbide is extracted using a contour extraction algorithm such as findContours; the minimum bounding rectangle is fitted to the extracted edge contour of the banded carbide, and the length of the long side of the minimum bounding rectangle is recorded as the length of the banded carbide, and the length of the short side is recorded as the width of the banded carbide;
[0051] 2) Sparsity calculation: Count the total number N pixels N representing carbides (i.e., pixels with a value of 1) within the smallest bounding rectangle. carbide Simultaneously calculate the total pixel area N of the minimum bounding rectangle. total Calculate the area ratio of banded carbides P=N carbide / N total The sparsity is defined as 1-P.
[0052] In step S5, a specific example of setting the rules is as follows: Level 1: No stripes, sparsity less than 30%; Level 2: No stripes, but sparsity greater than 50%, or stripes less than 2mm wide and not penetrating; Level 2.5: Stripes of 2-3.5mm or stripes of 3.5-4mm, but sparsity less than 70%; Level 3: Stripes of 3-4.5mm, sparsity greater than 70%, or stripes of 4.5-5mm with sparsity less than 75%; Level 3.5: Stripes of 4.5-5mm with sparsity greater than 75%, or stripes of 5-7mm with sparsity less than 80%; Level 4: Stripes of 5-7mm with sparsity greater than 80%, or stripes greater than 7mm.
[0053] This intelligent detection and rating method for banded carbides in bearing steel can achieve high accuracy and objectivity in detecting and rating banded carbides in bearing steel, greatly improving detection efficiency. Employing a banded carbide detection model based on YOLOv8, it achieves a target detection accuracy of over 95% and an image segmentation accuracy of over 85%, thereby enhancing the objectivity and accuracy of the rating results.
[0054] like Figure 6 The embodiment also provides a system employing the intelligent detection and rating method for bearing steel strip carbides described in any of the above embodiments, including an image acquisition module, a model building module, a model training module, an image detection module, and an image rating module.
[0055] Image acquisition module: acquires raw bearing steel images, performs grading and labels the banded carbide regions, and obtains the image training dataset;
[0056] Model building module: Constructs a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image.
[0057] Model training module: The YOLOv8-based banded carbide detection model is trained using an image training dataset and a loss function to obtain the trained detection model.
[0058] Image detection module: Inputs the image to be detected into the trained detection model to obtain the image of the banded carbides;
[0059] Image rating module: After calculating the length, width and sparsity of the banded carbide image, the rating result is generated according to the set rules.
[0060] This intelligent detection and rating method and system for banded carbides in bearing steel has the following significant advantages compared with existing technologies:
[0061] 1) By breaking through subjective dependence and achieving objectivity and standardization of rating results, this invention completely eliminates subjective factors through advanced image processing algorithms and artificial intelligence models. It will only output a unique and objective rating result for the same sample, which completely solves the long-standing problem of "rating consistency" in the industry and greatly improves the credibility and repeatability of rating results.
[0062] 2) Improved rating precision and accuracy, capturing microscopic features difficult for the human eye to discern. This invention enables pixel-level precise analysis of metallographic images, quantitatively calculating high-precision parameters across multiple dimensions such as carbide width, area percentage, distribution uniformity, and banding continuity. The system can identify and quantify subtle tissue differences that are indistinguishable to the human eye, thus providing a more refined and accurate rating basis than traditional "image reading," and even uncovering potential quality hazards.
[0063] 3) Significantly improves detection efficiency, meeting the needs of modern industrial online inspection. This invention enables automatic continuous processing and grading of batch images, reducing the analysis time for a single sample to the second level, increasing efficiency by tens of times. This allows for seamless integration into production lines, enabling online or semi-online real-time quality monitoring and providing immediate feedback for process adjustments—something traditional manual methods cannot achieve.
[0064] 4) Achieving digitalization and in-depth data mining of the rating process. Each rating in this invention generates a complete digital report including the original image, processing diagram, multiple quantitative parameters, probability distribution diagram, etc. This structured data can be stored in a database, and subsequently, big data technology can be used to perform correlation analysis with process parameters such as smelting, rolling, and heat treatment, thereby tracing back the causes of structural defects and providing a solid data foundation for optimizing production processes and achieving predictive quality control.
[0065] 5) Reduces reliance on the experience of professional personnel, saving human resource costs. This invention integrates the experience and knowledge of experts into algorithms and models, allowing ordinary operators to operate the equipment and obtain expert-level ratings after simple training. This not only reduces reliance on highly skilled technical personnel but also significantly reduces the company's long-term human resource training costs and solves the problem of technical gaps caused by personnel turnover.
[0066] 6) Enhance product quality control and industry competitiveness. By achieving more objective, faster, and more accurate ratings, this invention helps steel companies and bearing manufacturers to more strictly control the quality of raw materials, thereby improving the fatigue life and reliability of the final bearing products from the source. This not only reduces after-sales failures and claims caused by material problems, but also strengthens the company's quality bargaining power in its supply chain, builds a high-end brand image, and ultimately enhances its core market competitiveness.
[0067] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A method for intelligent detection and rating of banded carbides in bearing steel, characterized in that: Includes the following steps, S1. Obtain the original bearing steel image, classify and label the banded carbide regions to obtain the image training dataset; S2. Construct a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image. S3. After training the YOLOv8-based banded carbide detection model using the image training dataset and the loss function, the trained detection model is obtained. S4. Input the image to be detected into the trained detection model to obtain the banded carbide image; S5. After calculating the length, width, and sparsity of the banded carbide image, generate the corresponding grading results according to the set rules.
2. The intelligent detection and rating method for banded carbides in bearing steel as described in claim 1, characterized in that: In step S2, the preprocessing module preprocesses the input image to obtain the image to be recognized. Specifically, S21. Perform histogram equalization on the input image to obtain the histogram equalized image; S22. After histogram equalization, the image is enhanced with a directional filter to improve the lateral features of the banded carbides. After adaptive anchor box adjustment and directional non-maximum suppression, the image is binarized to obtain the image to be identified.
3. The intelligent detection and rating method for banded carbides in bearing steel as described in claim 1, characterized in that: In step S2, the image segmentation module uses either the DeeplabV3+ segmentation model or the U-net segmentation model.
4. The intelligent detection and rating method for banded carbides in bearing steel as described in any one of claims 1-3, characterized in that: In step S4, the loss functions used are cross-entropy loss (CE Loss), focal loss (Focal Loss), and auxiliary loss (Auxiliary Loss).
5. The intelligent detection and rating method for banded carbides in bearing steel as described in any one of claims 1-3, characterized in that: In step S5, the sparsity is calculated as follows: 1) Length and width calculation: For the strip carbide image obtained in step S4, i.e. the binary segmentation mask of the strip carbide, the edge contour of the strip carbide is extracted using the contour extraction algorithm; the minimum bounding rectangle is fitted to the extracted edge contour of the strip carbide, and the length of the long side of the minimum bounding rectangle is recorded as the length of the strip carbide, and the length of the short side is recorded as the width of the strip carbide. 2) Sparsity calculation: Count the total number N pixels N representing carbides (i.e., pixels with a value of 1) within the smallest bounding rectangle. carbide Simultaneously calculate the total pixel area N of the minimum bounding rectangle. total Calculate the area ratio of banded carbides P=N carbide / N total The sparsity is defined as 1-P.
6. A system employing the intelligent detection and rating method for banded carbides in bearing steel according to any one of claims 1-5, characterized in that: It includes an image acquisition module, a model building module, a model training module, an image detection module, and an image classification module. Image acquisition module: acquires raw bearing steel images, performs grading and labels the banded carbide regions, and obtains the image training dataset; Model building module: Constructs a YOLOv8-based banded carbide detection model. The YOLOv8-based banded carbide detection model includes a preprocessing module, a YOLOv8 network, and an image segmentation module. The preprocessing module preprocesses the input image to obtain the image to be identified. The YOLOv8 network performs target detection on the image to be identified to obtain the detection image. The image segmentation module performs multi-scale feature extraction and context information fusion on the detection image to achieve semantic segmentation and obtain the target image. Model training module: The YOLOv8-based banded carbide detection model is trained using an image training dataset and a loss function to obtain the trained detection model. Image detection module: Inputs the image to be detected into the trained detection model to obtain the image of the banded carbides; Image rating module: After calculating the length, width and sparsity of the banded carbide image, the rating result is generated according to the set rules.