Laser & Optoelectronics Progress, Volume. 61, Issue 8, 0828003(2024)
Lightweight Bilateral Input D-WNet Aerial Image Building Change Detection
Figures & Tables(14)
Fig. 9. Comparison of ablation results of network structure improvement experiment
Fig. 10. Feature extraction heat map of the network structure improved ablation model
Fig. 11. Building change detection results for different networks on WHU and LEVIR-CD test sets
Table 1. Comparison of experimental evaluation indexes of shape feature selection
View tableTable 1. Comparison of experimental evaluation indexes of shape feature selection
Dataset Network Input_left Input_right WHU A Former phase image Post-temporal image 98.01 97.61 97.81 95.54 B Biphasic image Canny 96.71 98.30 97.50 95.07 C Biphasic image Sobel 97.42 98.17 97.80 95.56 D-WNet Biphasic image Canny+Sobel 97.49 98.81 98.14 96.36 LEVIR-CD A Former phase image Post-temporal image 90.82 91.40 91.11 83.36 B Biphasic image Canny 89.76 91.84 90.79 82.94 C Biphasic image Sobel 89.40 91.98 90.67 83.25 D-WNet Biphasic image Canny+Sobel 90.23 92.25 91.23 83.49
Table 2. Comparison of experimental evaluation indexes of network structure improvement
View tableTable 2. Comparison of experimental evaluation indexes of network structure improvement
Dataset Network CBAM Lightweight
structure
ASPP /% /% /% /% Epoch time /s WHU D √ × × 97.96 97.90 98.03 96.02 1502 E × √ √ 96.60 98.55 97.56 95.25 438 F √ √ × 97.69 97.60 97.64 95.97 560 D-WNet √ √ √ 97.49 98.81 98.14 96.36 454 LEVIR-CD D √ × × 90.29 92.13 91.20 83.38 786 E × √ √ 88.21 92.20 90.16 82.67 235 F √ √ × 91.22 90.71 90.96 83.22 312 D-WNet √ √ √ 90.23 92.25 91.23 83.49 260
Table 3. Building change detection accuracy of different networks on WHU and LEVIR-CD test set
View tableTable 3. Building change detection accuracy of different networks on WHU and LEVIR-CD test set
Dataset Network /% /% /% /% Epoch time /s WHU U-Net 97.82 98.25 98.03 96.16 555 SENet 97.95 98.13 98.04 96.31 581 ResNet 97.75 98.04 97.89 95.88 455 DeepLabv3+ 96.77 97.56 97.16 94.49 224 D-WNet 97.49 98.81 98.14 96.36 454 LEVIR-CD U-Net 88.53 92.17 90.31 82.55 365 SENet 89.04 92.21 90.60 82.56 374 ResNet 88.91 91.73 90.30 82.14 262 DeepLabv3+ 88.06 92.08 90.03 81.86 160 D-WNet 90.23 92.25 91.23 83.49 260
Tools
Get Citation
Copy Citation Text
Fengxing Zhang, Jian Huang, Hao Li. Lightweight Bilateral Input D-WNet Aerial Image Building Change Detection[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0828003
Paper Information
Category: Remote Sensing and Sensors
Received: Jun. 8, 2023
Accepted: Jul. 24, 2023
Published Online: Mar. 15, 2024
The Author Email: Li Hao (lihao@hhu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20