• Laser & Optoelectronics Progress
  • Vol. 59, Issue 24, 2410008 (2022)
Lirong Li1,2,*, Peng Chen1, Yunliang Zhang1, Kai Zhang1..., Wei Xiong1,2 and Pengcheng Gong1,2|Show fewer author(s)
Author Affiliations
  • 1School of Electrical and Electronic Engineering, Hubei University of Technology, Hubei 430064, Wuhan, China
  • 2Hubei Engineering Research Center of New Energy and Power Grid Equipment Safety Monitoring, Hubei 430064, Wuhan, China
  • show less
    DOI: 10.3788/LOP202259.2410008 Cite this Article Set citation alerts
    Lirong Li, Peng Chen, Yunliang Zhang, Kai Zhang, Wei Xiong, Pengcheng Gong. Insulator Defect Detection Based on Multi-Scale Feature Coding and Dual Attention Fusion[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2410008 Copy Citation Text show less
    CenterNet structure
    Fig. 1. CenterNet structure
    Structure of proposed algorithm
    Fig. 2. Structure of proposed algorithm
    Schematic diagrams of Bottleneck block and Res2Net Module. (a) Bottleneck block; (b) Res2Net module
    Fig. 3. Schematic diagrams of Bottleneck block and Res2Net Module. (a) Bottleneck block; (b) Res2Net module
    ASPP module
    Fig. 4. ASPP module
    Schematic diagram of dual attention fusion
    Fig. 5. Schematic diagram of dual attention fusion
    Loss curve
    Fig. 6. Loss curve
    Sample dataset
    Fig. 7. Sample dataset
    Example of real box labels
    Fig. 8. Example of real box labels
    Visualization of detection results. (a) Visualization of insulator detection results; (b) visualization of test results of insulators and defective insulators
    Fig. 9. Visualization of detection results. (a) Visualization of insulator detection results; (b) visualization of test results of insulators and defective insulators
    Real categoryPredictive valueDefinition
    11TP
    10FP
    01TN
    00FN
    Table 1. Definitions of TP, TN, FP, and FN

    Algorithm framework

    Backbone network

    AP50

    P

    R

    mAP

    Normal

    Defect

    Normal

    Defect

    Normal

    Defect

    CenterNet

    ResNet18

    80.80

    14.95

    96.86

    50.00

    41.48

    0.79

    47.87

    ResNet50

    94.72

    76.46

    94.20

    89.47

    84.08

    53.54

    85.59

    ResNet101

    91.65

    62.70

    98.86

    84.00

    77.80

    33.07

    77.17

    DLANet34

    93.61

    62.61

    99.41

    81.40

    75.56

    27.56

    78.11

    Res2Net50

    94.26

    82.29

    95.93

    91.09

    84.53

    72.44

    88.27

    Table 2. Experimental results of different backbone networks
    No.Res2Net50ASPPECASEPR
    NormalDefectNormalDefect
    198.6892.2283.6365.35
    297.1692.5284.5377.95
    395.7394.5085.4381.10
    497.2696.3687.4483.46
    Table 3. Ablation experiment
    No.Res2Net50ASPPECASEAP50F1mAP /%
    Normal /%Defect /%NormalDefect
    193.2986.240.910.7689.77
    295.3388.580.900.8591.96
    394.7392.270.900.8793.50
    495.8894.810.920.8995.35
    Table 4. Ablation experiment
    The first branchThe second branchAP50PRmAP
    NormalDefectNormalDefectNormalDefect
    ECAECA95.2889.7097.7493.3387.4477.1792.49
    ECASE95.8894.8197.2696.3687.4483.4695.35
    Table 5. Comparison experiment of adding different attention on two branches

    Algorithm framework (backbone)

    AP50

    P

    R

    mAP

    FPS

    Normal

    Defect

    Normal

    Defect

    Normal

    Defect

    SSD(VGG16)

    91.02

    88.87

    93.75

    92.86

    87.5

    88.67

    89.95

    58.62

    RetinaNet(ResNet50)

    88.70

    86.38

    95.91

    88.03

    78.92

    84.1

    87.54

    44.30

    FasterRCNN(VGG16)

    96.25

    70.49

    73.47

    42.91

    96.86

    95.28

    83.37

    29.16

    YOLOv3(Darknet53)

    93.90

    87.45

    91.61

    91.27

    93.05

    90.55

    90.68

    75.63

    CenterNet(ResNet50)

    94.72

    76.46

    94.20

    89.47

    84.08

    53.54

    85.59

    97.51

    Proposed algorithm

    95.88

    94.81

    97.26

    96.36

    87.44

    83.46

    95.35

    65.95

    Table 6. Comparison of different algorithms
    Lirong Li, Peng Chen, Yunliang Zhang, Kai Zhang, Wei Xiong, Pengcheng Gong. Insulator Defect Detection Based on Multi-Scale Feature Coding and Dual Attention Fusion[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2410008
    Download Citation