Journal of Applied Optics, Volume. 45, Issue 4, 804(2024)

Electric meter data detection based on global and local multi-scale context

Tianlei MA... Jun FU*, Qi MA, Zhen YANG and Xinhao LIU |Show fewer author(s)
Author Affiliations
  • School of Electrical and Information Engineering, Zhengzhou University, Zhengzhou 450001, China
  • show less
    Figures & Tables(16)
    Overall framework diagram of network
    Structure diagram of convolutional module and identity module
    Structure diagram of local multi-scale context information extraction module
    Structure diagram of global context information extraction module
    Structure diagram of spatial attention module
    Structure diagram of channel attention module
    Partial images in dataset
    Detection result images
    Curves of P-R values
    Detection results of grayscale image
    • Table 1. Structure data of feature extraction network

      View table
      View in Article

      Table 1. Structure data of feature extraction network

      输入输出
      3×3卷积层(1280, 720, 3)(640, 360, 8)
      池化层(640, 360, 8)(320, 180, 8)
      卷积模块1(320, 180, 8)(320, 180, 16)
      恒等模块1(320, 180, 16)(320, 180, 16)
      卷积模块2(320, 180, 32)(160, 90, 64)
      恒等模块2(160, 90, 64)(160, 90, 64)
      卷积模块3(160, 90, 64)(80, 45, 128)
      恒等模块3(80, 45, 128)(80, 45, 128)
    • Table 2. AP and MAP for 11 categories of models with different training ratios

      View table
      View in Article

      Table 2. AP and MAP for 11 categories of models with different training ratios

      比例类别
      0123456789小数点MAP
      5∶50.930.980.930.960.910.950.920.960.910.970.910.94
      6∶40.920.910.940.940.970.900.980.990.930.990.930.95
      7∶30.950.980.960.930.940.940.990.950.960.970.940.96
      8∶20.990.990.990.990.990.950.990.990.990.990.950.98
      9∶10.960.980.980.970.980.960.970.960.960.970.960.97
    • Table 3. Parameter settings of comparison experiment

      View table
      View in Article

      Table 3. Parameter settings of comparison experiment

      模型参数
      ITOIS[8]滤波方法(均值滤波),模板窗口(3,3),输入尺寸(30,0)
      YOLO-v3[9]输入尺寸(416,416), 主干(Darknet53), 优化器(SGD),批次大小(8)
      Faster-RCNN[14]输入尺寸(600,480), 主干(Resnet50), 优化器(Adam),批次大小(2)
      YOLO-v5[10]输入尺寸(640,640), 主干(CSPDarknet), 优化器(SGD),批次大小(8)
      ESPA-YOLOv3[17]输入尺寸(416,416), 主干(ESPAnet), 优化器(SGD),批次大小(8)
    • Table 4. AP and MAP for 11 categories of different models

      View table
      View in Article

      Table 4. AP and MAP for 11 categories of different models

      模型类别
      012345
      ITOIS[8]0.730.680.530.760.510.75
      YOLO-v3[9]0.820.710.640.610.870.80
      Faster-RCNN[14]0.740.880.760.830.840.94
      YOLO-v5[10]0.890.900.830.860.920.93
      ESPA-YOLOv3[17]0.940.910.960.970.930.89
      本文方法0.990.990.990.990.990.95
      模型类别
      6789小数点MAP
      ITOIS[8]0.620.460.710.670.390.62
      YOLO-v3[9]0.780.690.810.690.500.72
      Faster-RCNN[14]0.890.850.860.870.560.83
      YOLO-v5[10]0.860.960.930.950.710.89
      ESPA-YOLOv3[17]0.970.930.950.960.780.92
      本文方法0.990.990.990.990.950.98
    • Table 5. Parameter size and running speed of different detection models

      View table
      View in Article

      Table 5. Parameter size and running speed of different detection models

      模型参数量/MbFPS
      ITOIS[8]0.5
      YOLO-v3[9]13022
      Faster-RCNN[14]1085
      YOLO-v5[10]12231
      ESPA-YOLOv3[17]13520
      本文模型11.0343
    • Table 6. Results of ablation experiment

      View table
      View in Article

      Table 6. Results of ablation experiment

      全局上下文局部上下文MAPFPS
      0.8748
      0.9246
      0.9445
      0.9843
    Tools

    Get Citation

    Copy Citation Text

    Tianlei MA, Jun FU, Qi MA, Zhen YANG, Xinhao LIU. Electric meter data detection based on global and local multi-scale context[J]. Journal of Applied Optics, 2024, 45(4): 804

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Articles

    Received: Jul. 7, 2023

    Accepted: --

    Published Online: Oct. 21, 2024

    The Author Email: FU Jun (符俊)

    DOI:10.5768/JAO202445.0403006

    Topics