Journal of Applied Optics, Volume. 46, Issue 3, 652(2025)
Lightweight pavement damage detection method based on feature mapping contribution degree
Figures & Tables(18)
Fig. 5. Channel weight visualization and activation response diagrams on YOLOv5s-MobileNetv3
Fig. 7. Adds SE attention mechanism to each convolutional layer on Neck section
Fig. 9. Count proportion of ReLU activation that is greater than set threshold (0.5)
Fig. 10. Detection results of proposed algorithm on China_Drone test set
Table 1. Damage definition and quantity statistics of China_Drone dataset in RDD2022 dataset
View tableView in ArticleTable 1. Damage definition and quantity statistics of China_Drone dataset in RDD2022 dataset
损伤类型 名称 标签数量 纵向裂纹（Longitudinal crack） D00 1426 横向裂纹（Transverse crack） D10 1263 龟状裂纹（Alligator crack） D20 293 坑洞（Potholes） D40 86 修复过路面（Repaired road） Repair 769
Table 2. Detection performance of different backbone networks
View tableView in ArticleTable 2. Detection performance of different backbone networks
模型 Parameters/M GFLOPs mAP FPS 边缘 FPS YOLOv5s 7.02 15.8 0.72 13 5 YOLOv5s-ShuffleNetV2 3.84 6.8 0.62 28 15 YOLOv5s-EfficientNet 4.09 7.7 0.63 23 19 YOLOv5s-MobileNetv3 3.37 6.0 0.70 35 26
Table 3. Information entropy statistics of different pruning standards for specific level
View tableView in ArticleTable 3. Information entropy statistics of different pruning standards for specific level
剪枝标准 层级 0 1 2 5 10 15 20 平均秩 4.47 1.25 1.43 3.97 3.91 2.91 3.00 均值 3.21 0.83 0.45 0.32 0.31 0.52 0.51 方差 2.36 1.22 1.23 0.67 0.52 0.33 0.54 中位数 1.14 0.22 0.35 0.16 0.36 0.38 0.31
Table 4. Detection performance of different pruning rates
View tableView in ArticleTable 4. Detection performance of different pruning rates
剪枝率 Parameters/M GFLOPs mAP FPS 0.1 3.08 6.1 0.643 37 0.2 2.77 4.9 0.612 40 0.5 1.97 3.1 0.546 42 0.8 0.60 1.3 0.462 50 0.0×7+0.1×16 2.11 4.6 0.676 39 0.0×7+0.3×16 1.82 2.9 0.657 44
Table 5. Visualization of different pruning rates
View tableView in ArticleTable 5. Visualization of different pruning rates
未剪枝 0.1 0.2 0.5 0.0×7+0.1×16 0.0×7+0.3×16
Table 6. Improved performance comparison of attention mechanisms
View tableView in ArticleTable 6. Improved performance comparison of attention mechanisms
Pruned-YOLOv5s -MobileNetV3 ConvSE C3SE ConvASE C3ASE Parameters/M GFLOPs mAP FPS √ 1.83 2.9 0.657 44 √ √ 2.09 4.3 0.663 26 √ √ 2.04 3.8 0.666 40 √ √ 2.09 4.3 0.668 26 √ √ 2.05 3.8 0.672 40 √ √ √ 2.07 4.1 0.686 40
Table 7. Heat map presentation of different pruning algorithms
View tableView in ArticleTable 7. Heat map presentation of different pruning algorithms
未剪枝 文献[17] 文献[25] 文献[26] 文献[27] 本文算法
Table 8. Comparison of inference time of each model
View tableView in ArticleTable 8. Comparison of inference time of each model
模型 Parameters/M mAP FPS YOLOv4-Tiny 6.06 0.66 16 YOLOx-Tiny 5.06 0.67 14 FCOS-Tiny 2.08 0.65 34 本文算法 2.06 0.686 40
Tools
Get Citation
Copy Citation Text
Yueming WANG, Yangxu WU, Zhiyu CHANG, Ping CHEN. Lightweight pavement damage detection method based on feature mapping contribution degree[J]. Journal of Applied Optics, 2025, 46(3): 652
Paper Information
Category:
Received: Jul. 22, 2024
Accepted: --
Published Online: May. 28, 2025
The Author Email: Ping CHEN (陈平)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20