Scholarlink

<p>As the core transmission component of the GM 6AT gearbox, the surface quality of metal gears is directly related to the overall performance of the gearbox. However, it is difficult for existing industrial defect detection methods to meet the enterprise’s demand for high-precision inspection. To address this situation, this paper proposes a new oriented detector, OGD-YOLO, which significantly improves the accuracy of defect detection by reducing the background noise. Firstly, a same scale cross-layer feature connection method combined with lightweight MLP is proposed to enrich the defect information of the deep network and improve the recognition ability of defects of small objects. Then, to further enhance the feature fusion ability of the deep network, the original feature concat module is improved. By integrating channel prior convolutional attention into the contact module, the irrelevant information within the deep network is effectively suppressed, and the feature extraction and fusion performance of the model is improved. Finally, the slim neck constructed with low-cost convolutional module VoV-GSCSP is used to replace the original network neck, which significantly reduces the computations and parameters of the model while maintaining detection accuracy. The experimental results show that OGD-YOLO achieves 78.1% mAP50 on the self-made Metal Gear dataset of 6AT Gearbox (MGDD), which is 2.6% higher than the benchmark YOLOv8s OBB and outperforms the existing mainstream object detection algorithms. It provides a more reliable and efficient solution for industrial applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

OGD-YOLO: surface defect detection for metal gears based on oriented detector

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Bufan Zhang,
  • Jinghu Yu,
  • Xingfei Zhu,
  • Qimeng Wang,
  • Fangyong Zhu

摘要

As the core transmission component of the GM 6AT gearbox, the surface quality of metal gears is directly related to the overall performance of the gearbox. However, it is difficult for existing industrial defect detection methods to meet the enterprise’s demand for high-precision inspection. To address this situation, this paper proposes a new oriented detector, OGD-YOLO, which significantly improves the accuracy of defect detection by reducing the background noise. Firstly, a same scale cross-layer feature connection method combined with lightweight MLP is proposed to enrich the defect information of the deep network and improve the recognition ability of defects of small objects. Then, to further enhance the feature fusion ability of the deep network, the original feature concat module is improved. By integrating channel prior convolutional attention into the contact module, the irrelevant information within the deep network is effectively suppressed, and the feature extraction and fusion performance of the model is improved. Finally, the slim neck constructed with low-cost convolutional module VoV-GSCSP is used to replace the original network neck, which significantly reduces the computations and parameters of the model while maintaining detection accuracy. The experimental results show that OGD-YOLO achieves 78.1% mAP50 on the self-made Metal Gear dataset of 6AT Gearbox (MGDD), which is 2.6% higher than the benchmark YOLOv8s OBB and outperforms the existing mainstream object detection algorithms. It provides a more reliable and efficient solution for industrial applications.