Scholarlink

<p>To address the issues of inadequate accuracy, excessive model complexity, and insufficient real-time performance in wire rope defect identification, an enhanced RT-DETR-based algorithm for wire rope defect detection is suggested. The backbone network of the original model is optimised and substituted with FasterNet, enhancing feature information extraction whilst decreasing computational expenses. The feature fusion component is enhanced by substituting standard convolution with GhostConv and RepC3 with DRBC3, both of which augment the model's detection capability for small objects and enhance real-time detection performance. Finally, the coding section of the model is enhanced by fusing the ATM module into AIFI, which improves the connection between the model feature extraction network and the feature fusion network and improves the detection accuracy. Experiments on the wire rope defect dataset reveal that the FGDA-DETR method enhances average accuracy by 2.1% compared to the RT-DETR technique, reduces the number of parameters by 56.8%, and increases FPS by 12.5.The number of parameters in the FGDA-DETR algorithm is lower than that of common object detection algorithms such as the YOLO series and other algorithms. Compared to YOLOv3-tiny, YOLOv5m, YOLOv8m, YOLOv9m, YOLOv10m, and YOLOv11m by 28.9%, 65.6%, 66.7%, 57.4%, 43.8%, and 64.5%, respectively, and compared to Faster R-CNN, SSD, Deformable DETR, DETR, and DION, it reduces them by 79.4%, 64.2%, 78.6%, 79.3%, and 81.9%, respectively. In terms of real-time performance, except for YOLOv3-tiny, YOLOv5m, and YOLOv10m, the FPS of the FGDA-DETR algorithm surpasses that of other algorithms, with increases of 6.4, 12.5, and 8.4 compared to YOLOv8m, YOLOv9m, and YOLOv11m, respectively. Compared to Faster R-CNN, SSD, Deformable DETR, DETR, and DION, the FPS was higher by 52.6%, 56.1%, 59%, 37.1%, and 63.1%, respectively. Although YOLOv3-tiny, YOLOv5m, and YOLOv10m have slightly higher real-time performance than FGDA-DETR, YOLOv3-tiny, and YOLOv10m are inferior to FGDA-DETR in terms of average accuracy and parameter count, whilst YOLOv5m is inferior to FGDA-DETR in terms of parameter count. FGDA-DETR achieves a balance between detection accuracy, lightweight design, and real-time performance, providing an effective solution for steel rope defect detection. </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

FGDA-DETR: a lightweight and real-time enhanced algorithm for wire rope defect detection with improved accuracy and efficiency

  • 期刊论文

  • 发表日期：

    2025年7月22日

查看全文
  • Xin Li,
  • Qianqian Xu,
  • Wenqing Shen

摘要

To address the issues of inadequate accuracy, excessive model complexity, and insufficient real-time performance in wire rope defect identification, an enhanced RT-DETR-based algorithm for wire rope defect detection is suggested. The backbone network of the original model is optimised and substituted with FasterNet, enhancing feature information extraction whilst decreasing computational expenses. The feature fusion component is enhanced by substituting standard convolution with GhostConv and RepC3 with DRBC3, both of which augment the model's detection capability for small objects and enhance real-time detection performance. Finally, the coding section of the model is enhanced by fusing the ATM module into AIFI, which improves the connection between the model feature extraction network and the feature fusion network and improves the detection accuracy. Experiments on the wire rope defect dataset reveal that the FGDA-DETR method enhances average accuracy by 2.1% compared to the RT-DETR technique, reduces the number of parameters by 56.8%, and increases FPS by 12.5.The number of parameters in the FGDA-DETR algorithm is lower than that of common object detection algorithms such as the YOLO series and other algorithms. Compared to YOLOv3-tiny, YOLOv5m, YOLOv8m, YOLOv9m, YOLOv10m, and YOLOv11m by 28.9%, 65.6%, 66.7%, 57.4%, 43.8%, and 64.5%, respectively, and compared to Faster R-CNN, SSD, Deformable DETR, DETR, and DION, it reduces them by 79.4%, 64.2%, 78.6%, 79.3%, and 81.9%, respectively. In terms of real-time performance, except for YOLOv3-tiny, YOLOv5m, and YOLOv10m, the FPS of the FGDA-DETR algorithm surpasses that of other algorithms, with increases of 6.4, 12.5, and 8.4 compared to YOLOv8m, YOLOv9m, and YOLOv11m, respectively. Compared to Faster R-CNN, SSD, Deformable DETR, DETR, and DION, the FPS was higher by 52.6%, 56.1%, 59%, 37.1%, and 63.1%, respectively. Although YOLOv3-tiny, YOLOv5m, and YOLOv10m have slightly higher real-time performance than FGDA-DETR, YOLOv3-tiny, and YOLOv10m are inferior to FGDA-DETR in terms of average accuracy and parameter count, whilst YOLOv5m is inferior to FGDA-DETR in terms of parameter count. FGDA-DETR achieves a balance between detection accuracy, lightweight design, and real-time performance, providing an effective solution for steel rope defect detection.