• Chinese Journal of Lasers
  • Vol. 52, Issue 12, 1202105 (2025)
Yunhao Li*, Chengtie Li, and Qiuming Li**
Author Affiliations
  • School of Control Engineering, Northeastern University at Qinhuangdao Campus, Qinhuangdao 066004, Hebei , China
  • show less
    DOI: 10.3788/CJL241390 Cite this Article Set citation alerts
    Yunhao Li, Chengtie Li, Qiuming Li. A Lightweight Real‐Time Weld Defect Classification Model[J]. Chinese Journal of Lasers, 2025, 52(12): 1202105 Copy Citation Text show less
    Schematic diagrams of models. (a) YOLOv8n-cls model; (b) TDRE-YOLO-cls model
    Fig. 1. Schematic diagrams of models. (a) YOLOv8n-cls model; (b) TDRE-YOLO-cls model
    Schematic diagram of RCR module
    Fig. 2. Schematic diagram of RCR module
    Schematic diagrams of RepConv. (a)Training multi-branch structure;(b)convolutional layer and batch normalization layer fusion;(c)inference single convolution
    Fig. 3. Schematic diagrams of RepConv. (a)Training multi-branch structure;(b)convolutional layer and batch normalization layer fusion;(c)inference single convolution
    Schematic diagram of DPSA module
    Fig. 4. Schematic diagram of DPSA module
    Schematic diagram of CSE module
    Fig. 5. Schematic diagram of CSE module
    Sample set examples. (a) Burr defect; (b) dimpling defect; (c) pore defect ; (d) no obvious defect
    Fig. 6. Sample set examples. (a) Burr defect; (b) dimpling defect; (c) pore defect ; (d) no obvious defect
    Accuracy change curves
    Fig. 7. Accuracy change curves
    Model

    Top-1 /

    %

    Weighted

    precision /%

    Weighted

    recall /%

    YOLOv8n-cls82.582.52282.776
    YOLOv11n-cls82.181.99882.245
    YOLOv8s-cls83.483.44383.929
    YOLOv11s-cls82.382.26182.752
    TDRE-YOLO-cls84.884.76084.885
    Table 1. Comparison of results of classification models on validation set
    ModelTop-1 /%Weighted precision /%Weighted recall /%Time /msSize/kB
    YOLOv8n-cls81.681.60481.7580.92907
    YOLOv11n-cls81.981.86782.2230.93124
    YOLOv8s-cls81.781.73682.0911.410028
    YOLOv11s-cls82.182.13182.5581.410784
    Mobilenetv382.482.37082.6893.96078
    Shufflenetv280.980.94481.1084.45086
    TDRE-YOLO-cls84.083.97283.9980.91393
    Table 2. Comparison of results of classification models on test set
    ModelTop-1 /%Weighted precision /%Weighted recall /%Time /msSize/kB
    Model_NoRCR82.882.79083.1590.91373
    Model_NoSPP81.581.47281.9081.01704
    Model_NoDSPA81.581.47481.9821.02570
    TDR-YOLO-cls83.683.57682.8330.91385
    TDRE-YOLO-cls84.083.97283.9980.91393
    Table 3. Comparison of ablation experimental results
    ModelPrecision /%
    Burr defectDimpling defectPore defectNo obvious defect
    YOLOv8n-cls81.64680.00083.07782.069
    YOLOv11n-cls84.68874.55683.07783.098
    YOLOv8s-cls81.95781.48180.76982.428
    YOLOv11s-cls86.25072.18984.61582.394
    TDRE-YOLO-cls85.53581.92883.96982.877
    Table 4. Precisions of classification models on four samples
    Yunhao Li, Chengtie Li, Qiuming Li. A Lightweight Real‐Time Weld Defect Classification Model[J]. Chinese Journal of Lasers, 2025, 52(12): 1202105
    Download Citation