• Optics and Precision Engineering
  • Vol. 31, Issue 3, 404 (2023)
Jian QIAO1,2, Nengda CHEN1, Yanxiong WU3, Yang WU1, and Jingwei YANG1,*
Author Affiliations
  • 1School of Electrical and Mechanical Engineering and Automation, Foshan University, Foshan528000, China
  • 2Ji Hua Laboratory, Foshan5800, China
  • 3School of Physics and Optoelectronic Engineering, Foshan University, Foshan528000, China
  • show less
    DOI: 10.37188/OPE.20233103.0404 Cite this Article
    Jian QIAO, Nengda CHEN, Yanxiong WU, Yang WU, Jingwei YANG. Defect detection of cylindrical surface of metal pot combining attention mechanism[J]. Optics and Precision Engineering, 2023, 31(3): 404 Copy Citation Text show less
    Structure of BiYOLOX network
    Fig. 1. Structure of BiYOLOX network
    Multiscale feature learning based on attention mechanism
    Fig. 2. Multiscale feature learning based on attention mechanism
    SAM module based on dilated convolution
    Fig. 3. SAM module based on dilated convolution
    Variation curves of sinusoidal attenuation factor
    Fig. 4. Variation curves of sinusoidal attenuation factor
    Example of classification loss function curves
    Fig. 5. Example of classification loss function curves
    Image acquisition and detection system
    Fig. 6. Image acquisition and detection system
    Samples of various defects on cylindrical surface of metal pot
    Fig. 7. Samples of various defects on cylindrical surface of metal pot
    Target quantity distribution of various labels
    Fig. 8. Target quantity distribution of various labels
    Statistics of defect recall rate of each category under different θ settings
    Fig. 9. Statistics of defect recall rate of each category under different θ settings
    Detection results of cylindrical surface defects of highlight reflective metal pot
    Fig. 10. Detection results of cylindrical surface defects of highlight reflective metal pot
    Super parameterValue
    Batch size4
    Cosine scheduler0.000 1
    Weight decay0.000 01
    NMS0.05
    Epoch300
    Table 1. Model super parameter setting
    Neck

    Params

    /M

    FLOPs

    /G

    FPS

    /(frame∙s-1

    mAP0.5

    /%

    PANet9.029.97728.0086.52
    BiFPN6.425.92432.2090.06
    Table 2. Performance comparison of feature fusion network
    Reg_LossFPS/(frame∙s-1mAP0.5/%mAP0.75/%
    IoU32.1085.2036.20
    GIoU32.0086.1138.47
    DIoU32.0088.9435.29
    CIoU32.0089.3039.63
    Table 3. Comparison of detection effect of regression loss functions
    Obj_LossαγθmAP0.5/%mAP0.75/%
    FL0.751-90.0234.99
    FL0.752-89.3033.63
    SFL0.7510.25π90.1438.95
    SFL0.7510.5π91.3736.56
    SFL0.7510.75π90.9135.31
    SFL0.751π90.5835.01
    Table 4. Comparison of detection effect of classification loss functions
    IndexLocationFNF

    FPS/

    (frame∙s-1

    mAP0.5

    /%

    mAP0.75

    /%

    Add1Add2Add3Add4
    10000132.2289.5733.28
    21100129.9889.6635.68
    30011130.8490.9238.21
    40111129.7189.6934.57
    51111127.7089.5935.98
    61111027.7888.4633.52
    Table 5. Performance comparison of feature fusion networks with Di_CBAM introduced at different locations
    IndexBiFPNDi_CBAMSFLFPS/(frame∙s-1mAP0.5/%mAP0.75/%
    100028.0086.5239.07
    210032.2090.0639.53
    301027.6088.1339.54
    400128.0086.8938.86
    511030.8490.3839.65
    610132.0091.3736.56
    701127.6088.2638.97
    811130.8490.9238.21
    Table 6. Comparison of impact of improvement strategies on model performance
    ModelParams/MFPS/(frame∙s-1mAP0.5/%mAP0.75/%
    Light-YOLOv38.628.7086.1235.92
    Light-YOLOv57.329.0687.3537.48
    YOLOX9.028.0086.5239.07
    BiYOLOX(Di_CBAM+SFL)6.430.8490.9238.21
    Table 7. Performance comparison of lightweight improved models
    Jian QIAO, Nengda CHEN, Yanxiong WU, Yang WU, Jingwei YANG. Defect detection of cylindrical surface of metal pot combining attention mechanism[J]. Optics and Precision Engineering, 2023, 31(3): 404
    Download Citation