Laser & Optoelectronics Progress, Volume. 61, Issue 8, 0812001(2024)
Lightweight Pavement Crack Detection Model Based on DeepLabv3+
Figures & Tables(14)
Fig. 3. Comparison of ordinary convolution and Ghost convolution. (a) Ordinary convolution; (b) Ghost convolution
Fig. 4. Common pooling, strip pooling and corresponding label. (a) Common pooling; (b) strip pooling;(c) corresponding label
Fig. 5. Schematic diagram of the network structure of the strip pooling module. (a) Input feature map; (b) (c) pooling feature maps;(d)(e) expand feature maps; (f) fusion feature map; (g) output feature map
Fig. 7. Shallow feature map visual diagram. (a) Original image; (b) original feature map; (c) feature map after introducing shallow feature fusion structure
Fig. 8. Visualized results of 5 models. (a) Original images; (b) ground truth; (c) traditional DeepLabv3+; (d) M-PSPNet; (e) R-PSPNet; (f) U-Net; (g) proposed method
Table 1. MobileNetv3 structure
View tableTable 1. MobileNetv3 structure
Input Operator Exp size out SE NL s 2242×3 Conv2d — 16 — HS 2 1122×16 bneck，3×3 16 16 — RE 1 1122×16 bneck，3×3 64 24 — RE 2 562×24 bneck，3×3 72 24 — RE 1 562×24 bneck，5×5 72 40 √ RE 2 282×40 bneck，5×5 120 40 √ RE 1 282×40 bneck，5×5 120 40 √ RE 1 282×40 bneck，3×3 240 80 — HS 2 142×80 bneck，3×3 200 80 — HS 1 142×80 bneck，3×3 184 80 — HS 1 142×80 bneck，3×3 184 80 — HS 1 142×80 bneck，3×3 480 112 √ HS 1 142×112 bneck，3×3 672 112 √ HS 1 142×112 bneck，5×5 672 160 √ HS 2 → 1142×160 bneck，5×5 960 160 √ HS 1 142×160 bneck，5×5 960 160 √ HS 1
Table 2. The ratio of each pixel in the training set to the total pixel
View tableTable 2. The ratio of each pixel in the training set to the total pixel
Class Background Crack Percentage /% 93.94 6.06
Table 3. Experimental environment
View tableTable 3. Experimental environment
Configuration Parameter Operating environment Ubuntu 16.04 Processor 7 Intel（R）Xeon（R）CPU E5-2680 v4 @ 2.40 GHz Graphic processing unit NVIDIA Gefore GTX Titan X Accelerating environment GPU Development platform Visual Studio Code
Table 4. Results of ablation experiment
View tableTable 4. Results of ablation experiment
Plan MobileNetv3，Ghost BCE+Dice SPM ECA SFF RIoU /% Rprecision /% Rrecall /% F1 /% FPS Params /MB 1 56.47 74.47 70.03 72.18 21.48 208.70 2 √ 55.73 73.99 69.31 71.57 57.74 14.32 3 √ √ 56.57 68.91 75.95 72.26 57.74 14.32 4 √ √ √ 56.84 69.95 75.20 72.48 52.20 14.46 5 √ √ √ √ 57.00 70.50 74.87 72.62 50.22 14.46 6 √ √ √ √ √ 57.21 70.60 75.09 72.76 47.18 14.53
Table 5. Comparison of test results using different models on the CRACK500 dataset
View tableTable 5. Comparison of test results using different models on the CRACK500 dataset
Model RIoU Rprecision Rrecall F1 FPHBN 54.51 71.23 71.65 70.56 CNNs 52.61 69.54 67.44 68.95 ACNet 54.92 68.05 74.89 69.82 Proposed method 57.21 70.60 75.09 72.76
Table 6. Comparison of test results between the proposed method and different semantic segmentation models
View tableTable 6. Comparison of test results between the proposed method and different semantic segmentation models
Model backbone RIoU /% Rprecision /% Rrecall /% F1 /% FPS Params /MB Traditional DeepLabv3+ Xception 56.47 74.47 70.03 72.18 21.48 208.70 M-PSPNet MobileNetv2 53.86 73.48 66.86 70.01 86.82 9.06 R-PSPNet ResNet50 55.37 71.74 70.81 71.27 47.09 178.17 U-Net VGG16 55.59 71.22 71.70 71.50 26.91 94.95 Proposed method MobileNetv3 57.21 70.60 75.09 72.76 47.18 14.53
Tools
Get Citation
Copy Citation Text
Xiaohua Xia, Jiangong Su, Yaoyao Wang, Yang Liu, Mingzhen Li. Lightweight Pavement Crack Detection Model Based on DeepLabv3+[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0812001
Paper Information
Category: Instrumentation, Measurement and Metrology
Received: May. 15, 2023
Accepted: Jul. 24, 2023
Published Online: Mar. 15, 2024
The Author Email: Xia Xiaohua (xhxia@chd.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20