• Infrared and Laser Engineering
  • Vol. 50, Issue 10, 20210011 (2021)
Xuan Wang1, Shuo Kang1, and Weidong Zhu1、2、3
Author Affiliations
  • 1School of Mechanical and Engineering, Zhejiang University, Hangzhou 310027, China
  • 2State Key Laboratory of Fluid Power and Mechatronic Systems, College of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China
  • 3Key Laboratory of Advanced Manufacturing Technology of Zhejiang Province, College of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China
  • show less
    DOI: 10.3788/IRLA20210011 Cite this Article
    Xuan Wang, Shuo Kang, Weidong Zhu. Defect detection of laminated surface in the automated fiber placement process based on improved CenterNet[J]. Infrared and Laser Engineering, 2021, 50(10): 20210011 Copy Citation Text show less
    (a) Keypoint heatmap; (b) Keypoint offset; (c) Object size
    Fig. 1. (a) Keypoint heatmap; (b) Keypoint offset; (c) Object size
    Schematic diagram of CenterNet network with ResNet-18 as backbone network
    Fig. 2. Schematic diagram of CenterNet network with ResNet-18 as backbone network
    Process of network learning fusion coefficient
    Fig. 3. Process of network learning fusion coefficient
    Structure diagram of AFP-CenterNet network
    Fig. 4. Structure diagram of AFP-CenterNet network
    (a) GT box and feasible boxes; (b) GT box heatmap obtained using a circular to filter feasible boxes; (c) GT box heatmap obtained using a ellipse to filter feasible boxes
    Fig. 5. (a) GT box and feasible boxes; (b) GT box heatmap obtained using a circular to filter feasible boxes; (c) GT box heatmap obtained using a ellipse to filter feasible boxes
    Improved bandwidth parameter solution method
    Fig. 6. Improved bandwidth parameter solution method
    (a) Data acquisition platform; (b) Infrared images of six typical kinds of AFP laminated surface defects
    Fig. 7. (a) Data acquisition platform; (b) Infrared images of six typical kinds of AFP laminated surface defects
    Graph of changes in loss value
    Fig. 8. Graph of changes in loss value
    Detection results of single frame infrared image
    Fig. 9. Detection results of single frame infrared image
    Network modelAPAP50AP75Time/msFPSMemory/MB
    CenterNet(DLA-34)0.7270.9260.7538112.377.0
    CenterNet(ResNet-101)0.6930.9040.7289310.7204.0
    CenterNet(ResNet-18)0.6260.8250.6473132.260.3
    AFP-CenterNet(MobileNetV3+ASFF)0.6940.9020.7314223.812.9
    Table 1. Detection results of CenterNet and AFP-CenterNet on AFP infrared data sets
    Network modelAP50Model memory/MB
    SSD0.80686.5
    YOLOv30.819235.0
    AFP-CenterNet0.90212.9
    Table 2. Detection results of different network models on AFP infrared data sets
    Network modelTime/msFPS
    SSD5371.8
    YOLOv34512.2
    CenterNet(ResNet-101)4252.3
    AFP-CenterNet2354.2
    Table 3. Detection speed of different network models only using CPU
    Xuan Wang, Shuo Kang, Weidong Zhu. Defect detection of laminated surface in the automated fiber placement process based on improved CenterNet[J]. Infrared and Laser Engineering, 2021, 50(10): 20210011
    Download Citation