Optics and Precision Engineering, Volume. 31, Issue 15, 2273(2023)
Image super-resolution reconstruction based on attention and wide-activated dense residual network
Figures & Tables(14)
Fig. 1. Structure of attention and wide-activated dense residual network
Fig. 3. Comparison of reconstruction effect of different combination methods on BSD100_42049
Fig. 4. Comparison of reconstruction effects of different combinations on Urban100_img002
Fig. 5. PSNR with 4× super-resolution at different channel multiples
Fig. 7. Comparison of reconstruction effects by different methods on Set5_baby
Fig. 8. Comparison of reconstruction effects by different methods on Set14_ monarch
Fig. 9. Comparison of reconstruction effects by different methods on BSD100_253027
Fig. 10. Comparison of reconstruction effects by different methods on Urban100_img091
Table 1. Effect of number of residual groups （
D ） and number of residual blocks（R ） on performance of 2× super-resolution reconstructionView tableView in ArticleTable 1. Effect of number of residual groups （
D ） and number of residual blocks（R ） on performance of 2× super-resolution reconstruction组合
方式
PSNR/dB 参数
量/M
Set5 Set14 BSD100 Urban100 R2D4 38.069 33.565 32.189 32.086 2.68 R4D2 38.029 33.538 32.153 32.069 2.52 R4D4 38.080 33.623 32.230 32.341 4.54 R4D6 37.896 33.325 31.944 31.317 6.55 R6D4 38.100 33.676 32.227 32.320 6.40
Table 2. Effect of convolution kernel size and connection mode of spatial feature transformation layer on performance of 2× super resolution reconstruction
View tableView in ArticleTable 2. Effect of convolution kernel size and connection mode of spatial feature transformation layer on performance of 2× super resolution reconstruction
连接方式 卷积核 PSNR/dB Set5 Set14 BSD100 Urban100 串联 1 38.075 33.612 32.220 32.283 串联 3 38.083 33.597 32.219 32.247 并联 1 38.080 33.623 32.230 32.341 并联 3 38.080 33.612 32.222 32.239
Table 3. Effects of attention mechanisms on 2× super-resolution reconstruction performance
View tableView in ArticleTable 3. Effects of attention mechanisms on 2× super-resolution reconstruction performance
注意力机制的选择 PSNR/dB Set5 Set14 BSD100 Urban100 × 38.038 33.572 32.202 32.058 SE 38.046 33.543 32.165 32.020 CA 38.069 33.558 32.175 32.082 CBAM 38.005 33.482 32.167 31.913 SK 38.080 33.623 32.230 32.341
Table 4. Reconstruction comparison on baseline dataset at magnifications of 2， 3 and 4
View tableView in ArticleTable 4. Reconstruction comparison on baseline dataset at magnifications of 2， 3 and 4
放大
倍数
模 型 Set5 Set14 BSD100 Urban100 PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM ×2 Bicubic［ 5 ］36.66 0.929 30.24 0.868 29.56 0.843 26.88 0.840 ESPCN［ 8 ］37.00 0.956 32.75 0.910 31.51 0.894 29.87 0.907 VDSR［ 9 ］37.54 0.959 33.03 0.912 31.90 0.896 30.76 0.914 DRCN［ 11 ］37.63 0.959 33.06 0.912 31.85 0.884 30.76 0.913 LapSRN［ 30 ］37.52 0.959 32.99 0.912 31.80 0.895 30.45 0.913 WDSR_Mini［ 23 ］37.96 0.960 33.50 0.914 32.13 0.899 31.79 0.921 IMDN［ 14 ］38.00 0.961 33.63 0.912 32.19 0.900 32.17 0.928 LatticeNet［ 15 ］38.15 0.961 33.78 0.919 32.25 0.901 32.43 0.930 WDRN 38.08 0.966 33.59 0.970 32.20 0.900 32.31 0.938 ×3 Bicubic 30.39 0.868 27.55 0.774 27.21 0.738 24.46 0.735 ESPCN 33.02 0.914 29.49 0.827 28.50 0.794 26.41 0.816 VDSR 33.65 0.921 29.78 0.831 28.82 0.798 27.14 0.828 DRCN 33.82 0.923 29.77 0.831 28.80 0.796 27.15 0.823 LapSRN 33.82 0.923 29.87 0.832 28.83 0.798 27.08 0.828 WDSR_Mini 34.375 0.925 30.31 0.840 29.06 0.803 27.97 0.847 IMDN 34.36 0.927 30.32 0.842 29.09 0.805 28.17 0.852 LatticeNet 34.53 0.928 30.39 0.842 29.15 0.806 28.33 0.854 WDRN 34.58 0.938 30.47 0.910 29.17 0.806 28.45 0.857 ×4 Bicubic 28.42 0.810 26.00 0.703 25.96 0.668 23.14 0.657 ESPCN 30.66 0.865 27.71 0.756 26.98 0.712 24.60 0.736 VDSR 31.35 0.884 28.01 0.767 27.30 0.725 25.18 0.751 DRCN 31.53 0.885 28.03 0.767 27.24 0.723 25.14 0.751 LapSRN 31.54 0.886 28.19 0.772 27.33 0.726 25.21 0.756 WDSR_Mini 32.17 0.893 28.59 0.783 27.56 0.731 25.95 0.778 IMDN 32.21 0.895 28.58 0.787 27.56 0.735 26.04 0.784 LatticeNet 32.30 0.896 28.68 0.783 27.62 0.737 26.25 0.787 WDRN 32.33 0.909 28.77 0.854 27.66 0.738 26.39 0.786
Tools
Get Citation
Copy Citation Text
Qiqi KOU, Chao LI, Deqiang CHENG, Liangliang CHEN, Haohui MA, Jianying ZHANG. Image super-resolution reconstruction based on attention and wide-activated dense residual network[J]. Optics and Precision Engineering, 2023, 31(15): 2273
Paper Information
Category: Information Sciences
Received: Nov. 1, 2022
Accepted: --
Published Online: Sep. 5, 2023
The Author Email: KOU Qiqi (kouqiqi@cumt.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20