• Opto-Electronic Engineering
  • Vol. 51, Issue 10, 240170 (2024)
Xiaoyan Wang1, Xiyu Wang2, Jie Li3、*, Wenhui Liang2, Jianhong Mou2, and Churan Bi1
Author Affiliations
  • 1School of Statistics and Data Science,Beijing Wuzi University,Beijing 101149,China
  • 2School of Information,Beijing Wuzi University,Beijing 101149,China
  • 3School of Mechanical-electronic and Vehicle Engineering,Beijing University of Civil Engineering and Architecture,Beijing 102616,China
  • show less
    DOI: 10.12086/oee.2024.240170 Cite this Article
    Xiaoyan Wang, Xiyu Wang, Jie Li, Wenhui Liang, Jianhong Mou, Churan Bi. MAS-YOLOv8n road damage detection algorithm from the perspective of drones[J]. Opto-Electronic Engineering, 2024, 51(10): 240170 Copy Citation Text show less
    YOLOv8 model structure
    Fig. 1. YOLOv8 model structure
    Multi-branch hybrid attention mechanism (MBMA module) structure
    Fig. 2. Multi-branch hybrid attention mechanism (MBMA module) structure
    C2f structure improvement. (a) Original C2f structure; (b) Improved C2f structure
    Fig. 3. C2f structure improvement. (a) Original C2f structure; (b) Improved C2f structure
    Example of bounding box regression
    Fig. 4. Example of bounding box regression
    Schematic diagram of ShapeIoU calculation
    Fig. 5. Schematic diagram of ShapeIoU calculation
    Examples of road damage types
    Fig. 6. Examples of road damage types
    YOLOv8n confusion matrix
    Fig. 7. YOLOv8n confusion matrix
    Confusion matrix of the model in this article
    Fig. 8. Confusion matrix of the model in this article
    Example of detection results
    Fig. 9. Example of detection results
    Heat map of visual features of attention mechanism
    Fig. 10. Heat map of visual features of attention mechanism
    Experimental results before and after improvement of the label allocation algorithm. (a) Comparison of Loss value changes; (b) Comparison of mAP changes
    Fig. 11. Experimental results before and after improvement of the label allocation algorithm. (a) Comparison of Loss value changes; (b) Comparison of mAP changes
    Damage typeDetailClass nameNumber of China-DroneNumber of dataset1Number of dataset2
    CrackLongitudinal crackD00142639952678
    Lateral crackD10126339791096
    Alligator crackD202936199641
    Other corruptionRutting,bump,pothole,separationD40862243235
    Crosswalk blurD43736
    White line blurD443995
    Special signsManhole coverD503553
    RepairRepair769277
    Table 1. Dataset road damage details
    CategoryEnvironment condition
    CPUAMD Ryzen 7 5800X 8-Core Processor
    GPUNVIDIA GeForce RTX 3060
    Graphics memory12 G
    Operating systemUbuntu 22.04
    CUDA versionCUDA 12.0
    Scripting languagePython
    Table 2. Experimental environment configuration
    ModelChina-Drone mAP@0.5/%Dataset1 mAP@0.5/%Dataset2 mAP@0.5/%Parameter/MModel volume/MB
    YOLOv5n64.764.092.22.55.03
    YOLOv8n68.564.793.63.05.96
    YOLOv10n62.461.891.42.75.51
    GOLD-YOLO66.165.994.57.211.99
    Faster-RCNN67.866.494.734.6310.24
    TOOD69.065.694.928.3243.95
    RTMDet-Tiny65.664.193.04.477.76
    RT-DETR68.267.287.520.0308
    MAS-YOLOv8n71.667.395.33.25.96
    Table 3. Comparative experimental results
    Attention mechanismChina-Drone mAP@0.5/%Dataset1 mAP@0.5/%Dataset2 mAP@0.5/%Parameter/M
    68.564.793.63.0
    SE69.164.193.53.1
    CMBA67.465.594.53.2
    CA68.865.794.53.2
    MBMA70.766.794.83.2
    Table 4. Verification results of attention mechanism
    ModelChina-Drone mAP@0.5/%Dataset1 mAP@0.5/%Dataset2 mAP@0.5/%Parameter/MGFLOPSModel volume/MBFPS
    1YOLOv8n68.564.793.63.08.15.96137
    2+MBMA70.766.794.83.28.15.96116
    3+ShapeIoU70.967.095.03.08.15.96135
    4MAS-YOLOv8n71.667.395.33.28.15.96114
    Table 5. Results of the ablation experiment
    Xiaoyan Wang, Xiyu Wang, Jie Li, Wenhui Liang, Jianhong Mou, Churan Bi. MAS-YOLOv8n road damage detection algorithm from the perspective of drones[J]. Opto-Electronic Engineering, 2024, 51(10): 240170
    Download Citation