Optics and Precision Engineering, Volume. 32, Issue 2, 268(2024)
Design of lightweight re-parameterized remote sensing image super-resolution network
Figures & Tables(20)
Fig. 1. Architecture of Re-parameterization residual feature network（ReRFN）
Fig. 7. Results of ×4 super-resolution reconstruction for Img062 in Urban100
Fig. 8. Results of ×4 super-resolution reconstruction for Img092 in Urban100
Fig. 9. Visualization results of selected samples from the SIDD dataset by various algorithms
Table 1. Performance comparison of various algorithms in UC Merced test dataset
View tableView in ArticleTable 1. Performance comparison of various algorithms in UC Merced test dataset
算法 Scale 参数量/K M-Adds/G PSNR SSIM Bicubic ×2 - - 30.76 0.878 9 SC ×2 - - 32.77 0.916 6 SRCNN ×2 8 4.5 32.84 0.915 2 FSRCNN ×2 17 0.3 33.18 0.919 6 LGCNet ×2 193 12.7 33.48 0.923 5 DCM ×2 1 842 30.2 33.65 0.927 4 HSENet ×2 5 286 66.8 34.22 0.932 7 CNT ×2 349 4.2 33.59 0.925 5 ReFDN ×2 - - 34.06 0.930 4 ReRFN ×2 532 8.1 34.33 0.932 7 Bicubic ×3 - - 27.46 0.763 1 SC ×3 - - 28.26 0.797 1 SRCNN ×3 8 4.5 28.66 0.803 8 FSRCNN ×3 17 0.2 29.09 0.816 7 LGCNet ×3 193 12.7 29.28 0.823 8 DCM ×3 2 258 16.3 29.52 0.839 4 HSENet ×3 5 470 30.8 30.00 0.842 0 CNT ×3 349 2.6 29.44 0.831 9 ReFDN ×3 - - 29.85 0.840 5 ReRFN ×3 539 3.6 30.01 0.844 9 Bicubic ×4 - - 25.65 0.672 5 SC ×4 - - 26.51 0.715 2 SRCNN ×4 8 4.5 26.78 0.721 9 FSRCNN ×4 17 0.1 26.93 0.726 7 LGCNet ×4 193 12.7 27.02 0.733 3 DCM ×4 2 175 13.0 27.22 0.752 8 HSENet ×4 5 433 19.2 27.73 0.762 3 CNT ×4 360 1.8 27.41 0.751 2 ReFDN ×4 343 1.3 27.68 0.759 6 ReRFN ×4 548 2.1 27.69 0.762 5
Table 2. PSNR values of 21 types of remote sensing scenes for each algorithm under ×3 SR
View tableView in ArticleTable 2. PSNR values of 21 types of remote sensing scenes for each algorithm under ×3 SR
类别 Bicubic SC SRCNN FSRCNN LGCNet DCM HSENet CNT ReRFN 农业 26.86 27.23 27.47 27.61 27.66 29.06 27.64 31.79 28.13 机场 26.71 27.67 28.24 28.98 29.12 30.77 30.09 28.22 30.16 棒球场地 33.33 34.06 34.33 34.64 34.72 33.76 35.05 29.37 35.16 海滩 36.14 36.87 37.00 37.21 37.37 36.38 37.69 34.73 37.68 建筑 25.09 26.11 26.84 27.50 27.81 28.51 28.95 37.89 28.91 丛林 25.21 25.82 26.11 26.21 26.39 26.81 26.70 28.01 26.74 密集住宅 25.76 26.75 27.41 28.02 28.25 28.79 29.24 26.42 29.19 森林 27.53 28.09 28.24 28.35 28.44 28.16 28.59 28.41 28.59 高速公路 27.36 28.28 28.69 29.27 29.52 30.45 30.63 28.43 30.58 高尔夫球场 35.21 35.92 36.15 36.43 36.51 34.43 36.62 29.67 36.75 海港 21.25 22.11 22.82 23.29 23.63 26.55 24.88 36.24 24.96 路口 26.48 27.20 22.67 28.06 28.29 29.28 29.21 23.99 29.19 中型住宅区 25.68 26.54 27.06 27.58 27.76 27.21 28.55 28.42 28.60 移动房公园 22.25 23.25 23.89 24.34 24.59 26.05 25.70 27.86 25.67 立交桥 24.59 25.30 25.65 26.53 26.58 27.77 28.22 24.99 28.08 停车场 21.75 22.59 23.11 23.34 23.69 24.95 24.66 27.48 24.69 河流 28.12 28.71 28.89 29.07 29.12 28.89 29.22 23.63 29.28 跑道 29.30 30.25 30.61 31.01 31.15 32.53 31.15 29.03 30.72 稀疏住宅区 28.34 29.33 29.40 30.23 30.53 29.81 31.64 30.68 31.66 油田 29.97 30.86 31.33 31.92 32.17 29.02 32.95 31.18 32.92 网球场 29.75 30.62 30.98 31.34 31.58 30.76 32.71 32.43 32.51 平均值 27.46 28.23 28.66 29.09 29.28 29.52 30.00 29.44 30.01
Table 3. Comparison of model performance of various algorithms under ×3 SR
View tableView in ArticleTable 3. Comparison of model performance of various algorithms under ×3 SR
算法 参数量/K M-Adds/G 推理时间/s PSNR LGCNet 193 12.7 0.004 29.28 DCM 2 258 16.3 0.003 29.52 HSENet 5 470 30.8 0.059 30.00 CTN 349 2.6 0.017 29.44 ReRFN 539 3.6 0.010 30.01
Table 4. Influence of ESA and EGCB on model performance
View tableView in ArticleTable 4. Influence of ESA and EGCB on model performance
基本模块 方法 参数量/K M-adds/G PSNR/
SSIM
ReFB 方法1 486 1.99 27.78/0.763 2 方法2 525 2.06 27.86/0.766 7 方法3 509 2.03 27.84/0.764 9 方法4 548 2.09 27.90/0.766 9 方法5
（Ours）
548 2.09 27.91/0.767 9
Table 5. Impact of channel compression Rate （
r ） of EGCB module on model performanceView tableView in ArticleTable 5. Impact of channel compression Rate （
r ） of EGCB module on model performance基本
模块
通道压缩倍数r 参数量/K M-adds/G PSNR/SSIM EGCB 1 586 2.12 27.87/0.766 1 2 548 2.09 27.91 /0.767 9 3 538 2.08 27.90/0.767 0 4 535 2.07 27.89/0.767 0
Table 6. Performance analysis of the re-parameterized structural branch module
View tableView in ArticleTable 6. Performance analysis of the re-parameterized structural branch module
基本模块 3×3
卷积
扩张和压缩卷积 残差连接 PSNR/SSIM Reblock 1 × × 27.81/0.764 2 1 √ × 27.82/0.764 6 1 × √ 27.86/0.765 9 1 √ √ 27.91 /0.767 9
Table 7. Impact of the number of ReFB modules on model performance
View tableView in ArticleTable 7. Impact of the number of ReFB modules on model performance
模块 数量N 参数量/K M-adds/G PSNR/SSIM ReFB 4 382 1.46 27.85/0.764 9 5 466 1.77 27.87/0.766 0 6 548 2.09 27.91/0.767 9 7 632 2.40 27.93/0.768 2 8 715 2.70 27.89/0.766 2
Table 8. Impact of the number of stacked Reblock+GELU layers on model performance
View tableView in ArticleTable 8. Impact of the number of stacked Reblock+GELU layers on model performance
基本模块 堆叠层数N 参数量/K M-adds
/G
PSNR/SSIM ReRLFB 1 278 0.98 27.73/0.7 609 2 413 1.53 27.84/0.7 649 3 548 2.09 27.91/0.767 9 4 684 2.64 27.93/0.768 6 5 819 3.20 27.92/0.768 2
Table 9. Impact of the Multi-layer feature fusion module on model Performance
View tableView in ArticleTable 9. Impact of the Multi-layer feature fusion module on model Performance
方法 参数量/K M-adds/G PSNR/SSIM 方法1 534 2.03 27.87/0.766 1 方法2 548 2.09 27.91/0.767 9
Table 10. Performance comparison of various algorithms on benchmark dataset for ×2，×3， and ×4 SR
View tableView in ArticleTable 10. Performance comparison of various algorithms on benchmark dataset for ×2，×3， and ×4 SR
算法 尺度 参数量
/K
M-Adds
/G
Set5
PSNR/ SSIM
Set14
PSNR/ SSIM
BSD100
PSNR/ SSIM
Urban100
PSNR/ SSIM
Manga109
PSNR/ SSIM
Bicubic ×2 - - 33.66/0.929 9 30.24/0.868 8 29.56/0.843 1 26.88/0.840 3 30.80/0.933 9 SRCNN 8 52.7 36.66/0.954 2 32.42/0.906 3 31.36/0，8 879 29.50/0.894 6 35.60/0.966 3 FSRCNN 13 6.0 36.98/0.955 6 32.62/0.908 7 31.50/0.890 4 29.85/0.900 9 36.67/0.971 0 CARN 1 592 222.8 37.76/0.959 0 33.52/0.916 6 32.09/0.897 8 31.92/0.925 6 38.36/0.976 5 IMDN 694 158.8 38.00/0.960 5 33.63/0.917 7 32.19/0.899 6 32.17/0.928 3 38.88/ 0.977 4 RFDN 534 120.9 38.05/0.960 6 33.68/0.918 4 32.16/0.899 4 32.12/0.927 8 38.88/0.977 3 RLFN 527 116.1 38.07 /0.960 733.72 /0.918 7 32.22/0.900 0 32.33/0.929 9 38.89/0.977 3 DARN 589 131.6 38.04/ 0.961 0 33.63/0.918 6 32.25 /0.901 2 32.40/0.930 5 38.87/ 0.977 6 ReRFN 532 113.8 38.10 /0.960 8 33.71 /0.919 2 32.23 /0.900 2 32.38 /0.930 1 38.91 /0.977 1Bicubic ×3 - - 30.39/0.868 2 27.55/0.774 2 27.21/0.738 5 24.46/0.734 9 26.95/0.855 6 SRCNN 8 52.7 32.75/0.909 0 29.30/0.821 5 28.41/0.786 3 26.24/0.798 9 30.48/0.911 7 FSRCNN 13 5.0 33.18/0.914 0 29.37/0.824 0 28.53/0.791 0 26.43/0.808 0 31.10/0.921 0 CARN 1 592 118.8 34.29/0.925 5 30.29/0.840 7 29.06/0.803 4 28.06/0.849 3 33.50/0.944 0 IMDN 703 71.5 34.36/0.927 0 30.32/0.841 7 29.09 /0.804 628.17/0.851 9 33.61/0.944 5 RFDN 541 54.3 34.41/0.927 3 30.34/0.842 0 29.09/0.805 0 28.21/0.852 5 33.67/0.944 9 DARN 596 58.4 34.48 /0.928 6 30.41 /0.844 3 29.15/0.807 6 28.38/0.857 0 33.76 /0.945 7 ReRFN 539 51.3 34.52 /0.928 2 30.44 /0.843 6 29.14 /0.806 3 28.36 /0.855 6 33.80/0.945 9 Bicubic ×4 - - 28.42/0.810 4 26.00/0.702 7 25.96/0.667 5 23.14/0.657 7 24.89/0.786 6 SRCNN 57 52.7 30.48/0.862 8 27.49/0.750 3 26.90/0.710 1 24.52/0.722 1 27.58/0.855 5 FSRCNN 13 4.6 30.72/0.866 0 27.61/0.755 0 26.98/0.715 0 24.62/0.728 0 27.90/0.861 0 CARN 1 592 90.9 32.13/0.893 7 28.60/0.780 6 27.58/0.734 9 26.07/0.783 7 30.47/0.908 4 IMDN 715 40.9 32.21/0.894 8 28.58/0.781 1 27.56/0.735 3 26.04/0.783 8 30.45/0.907 5 RFDN 550 31.1 32.24/0.895 2 28.61/0.781 9 27.57/0.736 0 26.11/0.785 8 30.58/0.908 9 RLFN 543 29.9 32.24 /0.895 228.62/0.781 3 27.60 /0.736 426.17/0.787 7 30.60/0.909 6 DARN 606 32.9 32.24 /0.896 3 28.64 /0.783 0 27.61/0.739 0 26.25 /0.791 3 30.64 /0.910 5 ReRFN 548 29.4 32.28 /0.895 8 28.66 /0.782 6 27.61 /0.737 1 26.26 /0.790 2 30.65 /0.910 3
Table 11. Performance comparison of various algorithms on the SIDD test set
View tableView in ArticleTable 11. Performance comparison of various algorithms on the SIDD test set
算法 Params/M SIDD PSNR SSIM BM3D［ 41 ］- 25.65 0.685 KSVD［ 42 ］- 26.88 0.842 EPLL［ 43 ］- 27.11 0.870 CBDNet［ 44 ］4.34 30.78 0.754 RIDNet［ 45 ］1.49 38.71 0.914 CTN［ 36 ］0.33 38.80 0.916 ReRFN 0.53 39.13 0.954
Tools
Get Citation
Copy Citation Text
Jianbing YI, Junkuan CHEN, Feng CAO, Jun LI, Weijia XIE. Design of lightweight re-parameterized remote sensing image super-resolution network[J]. Optics and Precision Engineering, 2024, 32(2): 268
Paper Information
Category:
Received: Jul. 5, 2023
Accepted: --
Published Online: Apr. 2, 2024
The Author Email: YI Jianbing (yijianbing8@jxust.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20