Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 5, 666(2023)

Rail surface crack detection algorithm based on improved YOLOv5s

Miao-sen ZHOU1,2, Quan-wu TANG1,2、*, Tian-tian SHI1,2, Tong-lan LUO1,2, Ze-xin ZHANG1,2, and Yong-xia XUE1,2
Author Affiliations
  • 1School of Physics and Electronic-Electrical Engineering,Ningxia University,Yinchuan 750021,China
  • 2Ningxia Key Laboratory of Intelligent Sensing for Desert Information,Yinchuan 750021,China
  • show less
    Figures & Tables(21)
    Sample images of three different levels of sleeper surface cracks.(a)sneg;(b)mneg;(c)lneg.
    YOLOv5s-6.1 overall structure
    Flowchart of height measurement
    C3 module
    Resunit structure
    C3_X and C3_X_F structure
    SPPF structure
    Feature pyramid network(FPN)and path aggregation network(PAN)
    CA attention mechanism
    C3CA structure diagram
    CA added location
    Two cross scale connection optimizations implemented by BiFPN
    Repeated three-layer BIFPN structure
    Comparison curves of two models training box_Loss
    Comparison of detection results before and after improvement.(a)Original drawing;(b)YOLOv5s detection effect drawing;(c)YOLOv5s-CBE detection effect drawing.
    • Table 1. Number of dataset label categories

      View table
      View in Article

      Table 1. Number of dataset label categories

      类别名数量/个
      sneg6 434
      mneg2 568
      lneg1 376
    • Table 2. Verification experiment of fusion attention mechanism module

      View table
      View in Article

      Table 2. Verification experiment of fusion attention mechanism module

      ModelClassesPrecision/%Recall/%mAP@ 0.5/%Weight/MBParameters/MGFLOPs
      YOLOv5s

      All

      sneg

      mneg

      lneg

      68.9

      71.6

      69.6

      65.4

      73.5

      75.5

      66.1

      78.9

      77.0

      77.7

      73.8

      79.5

      14.57.0215.8
      YOLOv5s-CA

      All

      sneg

      mneg

      lneg

      65.8

      70.0

      67.9

      59.4

      75.0

      80.1

      66.9

      78.0

      77.8

      82.4

      76.9

      74.1

      14.57.0415.8
      YOLOv5s-C3CA

      All

      sneg

      mneg

      lneg

      69.4

      71.5

      70.6

      66.0

      67.4

      73.6

      64.4

      64.1

      77.2

      81.6

      77.2

      72.7

      13.36.4214.0
      YOLOv5s-CC

      All

      sneg

      mneg

      lneg

      67.6

      73.5

      65.9

      63.5

      75.1

      77.2

      69.1

      78.9

      78.4

      82.3

      73.8

      79.0

      13.46.4414.0
    • Table 3. Partial validation experiments of improved Neck

      View table
      View in Article

      Table 3. Partial validation experiments of improved Neck

      ModelClassesPrecision/%Recall/%mAP@ 0.5/%Weight/MBParameters/MGFLOPs
      YOLOv5s

      All

      sneg

      mneg

      lneg

      68.9

      71.6

      69.6

      65.4

      73.5

      75.5

      66.1

      78.9

      77.0

      77.7

      73.8

      79.5

      14.57.0215.8
      YOLOv5s-B

      All

      sneg

      mneg

      lneg

      72.3

      73.6

      72.7

      70.4

      70.4

      76.4

      60.6

      74.3

      78.7

      81.2

      75.7

      79.2

      14.67.0816.0
    • Table 4. Partial verification experiment of improved loss function

      View table
      View in Article

      Table 4. Partial verification experiment of improved loss function

      ModelClassesPrecision/%Recall/%mAP@ 0.5/%Weight/MBParameters/MGFLOPs
      YOLOv5s

      All

      sneg

      mneg

      lneg

      68.9

      71.6

      69.6

      65.4

      73.5

      75.5

      66.1

      78.9

      77.0

      77.7

      73.8

      79.5

      14.57.0215.8
      YOLOv5s-E

      All

      sneg

      mneg

      lneg

      65.2

      68.4

      65.5

      61.7

      77.3

      79.5

      70.8

      81.7

      78.2

      81.1

      74.8

      78.5

      14.57.0215.8
    • Table 5. Ablation experiment

      View table
      View in Article

      Table 5. Ablation experiment

      ModelCBEPrecision/%Recall/%mAP@ 0.5/%Weight/MBParameters/MGFLOPs
      YOLOv5s68.973.577.014.57.0215.8
      YOLOv5s-C67.675.178.413.46.4414.0
      YOLOv5s-B72.370.478.714.67.0816.0
      YOLOv5s-E65.277.378.214.57.0215.8
      YOLOv5s-CB70.073.880.013.56.5114.2
      YOLOv5s-CE72.969.079.013.46.4414.0
      YOLOv5s-BE63.580.079.614.67.0816.0
      YOLOv5s-CBE67.976.280.713.56.5114.2
    • Table 6. Comparative experiment

      View table
      View in Article

      Table 6. Comparative experiment

      ModelPrecision/%Recall/%mAP@ 0.5/%Weight/MBParameters/MGFLOPs
      Faster R-CNN64.478.079.3108137.10370.21
      SSD80.451.066.891.626.2862.74
      YOLOX63.282.680.034.354.21165.01
      YOLOv464.464.470.724464.3660.52
      CenterNet64.571.577.612432.6670.22
      YOLOv5s-CBE67.976.280.713.56.5114.2
    Tools

    Get Citation

    Copy Citation Text

    Miao-sen ZHOU, Quan-wu TANG, Tian-tian SHI, Tong-lan LUO, Ze-xin ZHANG, Yong-xia XUE. Rail surface crack detection algorithm based on improved YOLOv5s[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(5): 666

    Download Citation

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

    Category: Research Articles

    Received: Aug. 13, 2022

    Accepted: --

    Published Online: Jul. 4, 2023

    The Author Email: Quan-wu TANG (tangqw@nxu.edu.cn)

    DOI:10.37188/CJLCD.2022-0267

    Topics