Optics and Precision Engineering, Volume. 31, Issue 22, 3371(2023)
Lightweight multi-scale difference network for remote sensing building extraction
Figures & Tables(16)
Fig. 7. Model training and validation phase metrics change trend graph
Fig. 8. Segmentation efficiency and performance of six feature processing units on the WHU dataset
Table 1. LMD-Net Network Components
View tableView in ArticleTable 1. LMD-Net Network Components
Stage Level Module type Filter Output size Encode Input 3×224×224 Level 1 BlockA 3×3/64 64×224×224 Level 2 Downsampling 3×3/128 128×112×112 BlockA 3×3/128 128×112×112 Level 3 Downsampling 3×3/256 256×56×56 MSDP 3×3/256 256×56×56 Level 4 Downsampling 3×3/512 512×28×28 BlockA 3×3/512 512×28×28 Decode Level 5 Upsampling 3×3/256 256×56×56 DIM 3×3/256 256×56×56 Level 6 Upsampling 3×3/128 128×112×112 MSDP 3×3/128 128×112×112 Level 7 Upsampling 3×3/64 64×224×224 BlockD 3×3/64 64×224×224 Output 1×1 1×224×224
Table 2. Segmentation accuracy of MSDP on WHU dataset
View tableView in ArticleTable 2. Segmentation accuracy of MSDP on WHU dataset
Group Model Position IoU Pre R Dice MSDP 1 BlockA+BlockD 无 86.84 91.34 90.31 90.80 2 Group1+MSDP 0 87.03 91.35 90.66 90.98 3 Group1+MSDP 1 87.07 91.43 90.71 91.05 4 Group1+MSDP 0、1 87.26 91.13 91.08 91.10 5 Group1+MSDP 0、2 86.89 91.20 90.32 90.84
Table 3. Experiments comparing complexity of each model
View tableView in ArticleTable 3. Experiments comparing complexity of each model
Group Model Params/M MADDs/G Flops 1 SegNet 29.44 61.36 30.7 G 2 U-Net 32.08 77.53 38.81 G 3 ENet 0.36 1.05 434.83 M 4 ESFNet 0.68 1.69 850.33 M 5 ResUNet 13.04 109.07 52.12 G 6 LMD-Net 10.42 81.68 39.23 G
Table 4. Comparison experiments of segmentation accuracy of different networks
View tableView in ArticleTable 4. Comparison experiments of segmentation accuracy of different networks
Model Iou Pre R Dice U-Net 83.61 90.09 87.57 88.78 SegNet 79.45 86.76 85.24 85.96 ENet 81.34 87.91 86.92 87.37 ESFnet 76.45 83.95 83.54 83.72 ResUNet 84.49 89.19 89.48 89.31 LMD-Net 87.72 91.76 91.04 91.38
Tools
Get Citation
Copy Citation Text
Guoyan LI, Haimiao WU, Chunhua DONG, Yi LIU. Lightweight multi-scale difference network for remote sensing building extraction[J]. Optics and Precision Engineering, 2023, 31(22): 3371
Paper Information
Category:
Received: Jun. 4, 2023
Accepted: --
Published Online: Dec. 29, 2023
The Author Email: Chunhua DONG (dch@tcu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20