Laser & Optoelectronics Progress, Volume. 59, Issue 22, 2228004(2022)
Building Segmentation Model of Remote Sensing Image Combining Multiscale Attention and Edge Supervision
Figures & Tables(13)
Fig. 6. Segmentation results of ablation experiment on WHU Building test set. (a) Original images; (b) image label;(c) baseline processing result; (d) baseline with MSA processing result; (e) baseline with EEB processing result; (f) MAE-Net processing result
Fig. 7. Segmentation results of comparative experiment on WHU Building validation set.(a) Original images; (b) image label; (c) U-Net processing result; (d) MAE-Net processing result
Fig. 8. Segmentation results of comparative experiment on Satellite Dataset Ⅱ (East Asia) test set. (a) Original images; (b) image label; (c) U-Net processing result; (d) MAE-Net processing result
Table 1. Comparison of evaluation results of ablation experiment on WHU Building test set
View tableTable 1. Comparison of evaluation results of ablation experiment on WHU Building test set
No. Baseline MSA EEB F1-score IOU 1 √ × × 0.9420 0.8904 2 √ √ × 0.9502 0.9052 3 √ × √ 0.9496 0.9041 4 √ √ √ 0.9520 0.9084
Table 2. Comparison of train and test time of ablation experiment on WHU Building test set
View tableTable 2. Comparison of train and test time of ablation experiment on WHU Building test set
No. Baseline MSA EEB Train time /h Test time per picture /s 1 √ × × 5.13 0.077 2 √ √ × 8.16 0.078 3 √ × √ 6.83 0.066 4 √ √ √ 9.13 0.089
Table 3. Accuracy results of two models for building and building edge recognition
View tableTable 3. Accuracy results of two models for building and building edge recognition
Scale U-Net MAE-Net IOU F1-score IOU F1-score Small-scale（image 1+image 2） 0.904 0.949 0.910 0.953 Large-scale（image 5+image 6） 0.918 0.957 0.927 0.963 Edge-scale（image 3+image 4） 0.438 0.609 0.545 0.706
Table 4. Experimental results of different methods on WHU building validation set
View tableTable 4. Experimental results of different methods on WHU building validation set
Method F1-score P R IOU U-Net［ 6 ］0.9135 0.9542 0.8826 0.8813 SiU-Net［ 16 ］0.9385 0.9380 0.9390 0.8840 SRINet［ 17 ］0.9423 0.9521 0.9328 0.8909 Ra-CGAN［ 1 ］0.9490 0.9510 0.9460 0.8960 DeNet［ 18 ］0.9480 0.9500 0.9460 0.9012 MAE-Net 0.9542 0.9554 0.9530 0.9124
Table 5. Experimental results of different methods with the Satellite Dataset Ⅱ (East Asia) test set
View tableTable 5. Experimental results of different methods with the Satellite Dataset Ⅱ (East Asia) test set
Method F1-score P R IOU U-Net［ 6 ］0.746 0.653 0.869 0.594 SiU-Net［ 16 ］0.758 0.725 0.796 0.611 AugU-Net［ 19 ］0.780 0.640 RSIS-MLCA［ 20 ］0.796 0.826 0.768 0.661 Ra-CGAN［ 1 ］0.812 0.852 0.775 0.677 MAE-Net 0.802 0.828 0.805 0.690
Tools
Get Citation
Copy Citation Text
Xiaoyu Yang, Xili Wang. Building Segmentation Model of Remote Sensing Image Combining Multiscale Attention and Edge Supervision[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2228004
Paper Information
Category: Remote Sensing and Sensors
Received: Aug. 17, 2021
Accepted: Oct. 13, 2021
Published Online: Oct. 13, 2022
The Author Email: Xili Wang (wangxili@snnu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20