Optics and Precision Engineering, Volume. 33, Issue 11, 1818(2025)

Online detection of plate weld defects incorporating triple attention mechanism

Lingyuan MENG1, Yingjun LI1,2、*, Guicong WANG1,2, Yuan LIU1, Jialong GAO1, and Peng XU1
Author Affiliations
  • 1School of Mechanical Engineering, University of Jinan, Jinan250022, China
  • 2Shandong Provincial Key Laboratory of Sensor Technology and High Precision Weighing Instruments, Jinan50004, China
  • show less
    Figures & Tables(11)
    Structure diagram of YOLO-TR network
    Schematic diagram of DySample resampling process
    Schematic diagram of structural principle of triple attention mechanism
    Structure of improved CLLAHead detection head
    Schematic diagram of prediction frame and real frame
    Detection effect of weld defects
    Comparison of accuracy curves between YOLO-TR and YOLOv5
    • Table 1. Improvement of test environment configuration for YOLOv5

      View table
      View in Article

      Table 1. Improvement of test environment configuration for YOLOv5

      软硬件版本
      CPUIntel Core i7 13700H@2.4 GHz
      内存32 GB
      显卡NVIDIA GeForce RTX 4060
      主板联想LNVNB161216
      操作系统Windows 11
      PythonPython3.8
      PycharmCommunity
      PytorchTorch2.3
      Anaconda3
    • Table 2. Dataset classification number

      View table
      View in Article

      Table 2. Dataset classification number

      数据类型烧糊烧穿错边未熔合夹渣
      训练集302236200155134
      验证集7065504543
    • Table 3. Verification results of ablation experiments

      View table
      View in Article

      Table 3. Verification results of ablation experiments

      算 法P/%R/%mAP@0.5/%mAP@0.5_0.95/%ParamGFLOPs
      YOLOv588.986.388.460.1703850816.0
      YOLOv5+Triplet Attention93.486.089.261.3733258820.5
      YOLOv5+DySample89.886.088.359.3700335915.1
      YOLOv5+CLLAHead95.082.988.261.8740868018.5
      YOLOv5+Shape-IoU91.587.188.759.6703850817.2
      YOLOv5+DySample+CLLAHead92.087.989.159.0733621221.2
      YOLOv5+Triplet Attention+CLLAHead92.387.790.359.1732238021.6
      YOLOv5+Triplet Attention+DySample89.888.088.661.2706651216.5
      YOLO-TR92.587.489.260.9735124419.1
    • Table 4. Experimental results of different algorithms

      View table
      View in Article

      Table 4. Experimental results of different algorithms

      算 法P/%R/%mAP@0.5/%mAP@0.5:0.95/%ParamGFLOPs
      YOLOv588.886.388.460.1703850816.0
      YOLOv5-transformer91.082.886.660.5702977215.6
      YOLOv3-tiny85.475.180.352.2868055212.9
      YOLOv3-spp87.677.581.451.862605644156.2
      YOLOv890.188.090.560.01112829328.5
      YOLOv1086.983.288.561.7804037824.5
      YOLOv1190.881.485.961.325912056.4
      YOLOv1286.684.584.460.225580936.3
      YOLO-TR92.587.489.260.9735124419.1
    Tools

    Get Citation

    Copy Citation Text

    Lingyuan MENG, Yingjun LI, Guicong WANG, Yuan LIU, Jialong GAO, Peng XU. Online detection of plate weld defects incorporating triple attention mechanism[J]. Optics and Precision Engineering, 2025, 33(11): 1818

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Apr. 23, 2025

    Accepted: --

    Published Online: Aug. 14, 2025

    The Author Email: Yingjun LI (me_liyj@ ujn.edu.cn)

    DOI:10.37188/OPE.20253311.1818

    Topics