Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0412004(2024)
UAV Highway Guardrail Inspection Based on Improved DeepLabV3+
Figures & Tables(13)
Table 1. Improved MobileNetv2 network structure
View tableTable 1. Improved MobileNetv2 network structure
Operator t c n s Output size Conv2d 3×3 32 1 2 3202×32 Bottleneck 1 16 1 1 3202×16 Bottleneck 6 24 2 2 1602×24 Bottleneck 6 32 3 2 802×32 Bottleneck 6 64 4 2 402×64 Bottleneck 6 96 3 1 402×96 Bottleneck 6 160 3 1 402×160 Bottleneck 6 320 1 1 402×320
Table 2. Experimental environment setting
View tableTable 2. Experimental environment setting
Name Parameter Operating system Windows 10 Processor AMD Ryzen 9 5950X Video card NVIDIA GeForce RTX 3090Ti RAM 32 G Development language Python 3.8 Development environment PyCharm 2021 Web framework PyTorch 1.7
Table 3. Comparison experiment of attentional mechanisms
View tableTable 3. Comparison experiment of attentional mechanisms
Base model Backbone network Attention model mpa /% mIoU /% Speed /（frame·s-1） DeepLabV3+ MobileNetv2 87.35 77.82 64.33 DeepLabV3+ MobileNetv2 CBAM 87.33 78.20 54.10 DeepLabV3+ MobileNetv2 ECA 87.39 78.46 66.65 DeepLabV3+ MobileNetv2 Coord 87.06 78.39 62.33 DeepLabV3+ MobileNetv2 SGE 87.79 78.56 66.74 DeepLabV3+ Xception 84.96 76.61 33.56 DeepLabV3+ Xception SGE 85.95 77.78 32.75
Table 4. Comparison of ablation experiment results
View tableTable 4. Comparison of ablation experiment results
Group MobileNetv2 Feature fusion SGE DenseASPP mpa /% mIoU /% Speed /（frame·s-1） Params /106 ① 84.96 76.61 33.56 54.71 ② √ 87.35 77.82 64.33 5.81 ③ √ √ 87.39 78.26 58.98 5.83 ④ √ √ 87.79 78.56 66.74 5.81 ⑤ √ √ 87.31 78.45 62.50 11.55 ⑥ √ √ √ 87.34 78.81 59.72 11.56 ⑦ √ √ √ √ 87.89 79.20 52.59 11.57
Table 5. Comparison of experimental results of mainstream semantic segmentation models
View tableTable 5. Comparison of experimental results of mainstream semantic segmentation models
Network Backbone network mpa /% mIoU /% Speed /（frame·s-1） Training time UNet VGG16 84.79 76.97 25.05 17 h 54 min PSPNet MobileNetv2 68.94 62.53 85.16 8 h 3 min DeepLabV3+ Xception 84.96 76.61 33.56 20 h 17 min Ours MobileNetv2 87.89 79.20 52.59 11 h 15 min
Tools
Get Citation
Copy Citation Text
Yang Wang, Dudu Guo, Qingqing Wang, Fei Zhou, Ying Qin. UAV Highway Guardrail Inspection Based on Improved DeepLabV3+[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0412004
Paper Information
Category: Instrumentation, Measurement and Metrology
Received: May. 10, 2023
Accepted: Jun. 27, 2023
Published Online: Feb. 27, 2024
The Author Email: Dudu Guo (guodd@xju.edu.cn)
CSTR:32186.14.LOP231270
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20