• Laser & Optoelectronics Progress
  • Vol. 59, Issue 24, 2410002 (2022)
Zhongxing Duan1,*, Yuming Zhang1, and Jiahao Ma2
Author Affiliations
  • 1College of Information and Control Engineering, Xi'an University of Architecture and Technology, Xi'an 710055, Shaanxi, China
  • 2State Grid Xi'an Power Supply Company, Xi'an 710032, Shaanxi, China
  • show less
    DOI: 10.3788/LOP202259.2410002 Cite this Article Set citation alerts
    Zhongxing Duan, Yuming Zhang, Jiahao Ma. Infrared Image Recognition of Power Equipment Using Improved YOLOv4[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2410002 Copy Citation Text show less
    Original infrared images
    Fig. 1. Original infrared images
    Images enhanced by MSRCR
    Fig. 2. Images enhanced by MSRCR
    YOLOv4 network structure
    Fig. 3. YOLOv4 network structure
    Multi-scale convolution module
    Fig. 4. Multi-scale convolution module
    Infrared images of power equipment. (a) Easy to classify sample; (b) difficult to classify sample
    Fig. 5. Infrared images of power equipment. (a) Easy to classify sample; (b) difficult to classify sample
    Data annotation process
    Fig. 6. Data annotation process
    Loss decline curve
    Fig. 7. Loss decline curve
    Infrared image recognition result of the proposed method for power equipments
    Fig. 8. Infrared image recognition result of the proposed method for power equipments
    ParameterNC1C2C3θ1θ2θ3κσ
    Value315752250.330.330.3443131
    Table 1. Parameter value of MSRCR algorithm
    ParameterInsulatorBushingDisconnectorBreakerArresterVoltage transformerCurrent transformerTransformer
    Number496451418401405382334313
    Table 2. Number of devices on infrared image dataset
    Electric power equipmentAP /%Speed /(frame·s-1
    Insulator97.1571
    Bushing96.57
    Disconnector97.83
    Breaker98.23
    Arrester96.51
    Voltage transformer94.62
    Current transformer96.35
    Transformer93.24
    mAP /%96.31
    Table 3. Test results of the proposed method on different power equipments
    MethodRaw infrared imageImage enhanced by MSRCR algorithm
    mAP@0.5 /%Speed /(frame·s-1mAP@0.5 /%Speed /(frame·s-1
    Faster R-CNN93.131694.0317
    SSD90.133391.3735
    YOLOv391.546592.7168
    YOLOv492.737493.6975
    Proposed method95.126996.3171
    Table 4. Performance comparison of different methods
    Modulation parametermAP /%
    β=0.5,γ=192.63
    β=0.7,γ=192.66
    β=0.9,γ=192.82
    β=0.9,γ=294.73
    β=0.9,γ=396.31
    β=0.9,γ=495.26
    β=0.9,γ=594.67
    β=0.7,γ=396.18
    Table 5. Comparison of experimental accuracy under different parameters of Focal loss