Optics and Precision Engineering, Volume. 31, Issue 6, 920(2023)
Deep learning image denoising based on multi-stage supervised with Res2-Unet
Figures & Tables(22)
Fig. 14. Comparison of MSE loss and combined loss in gaussian noise training process
Table 1. Convolution kernel parameters in Res2-Unet-SE
View tableView in ArticleTable 1. Convolution kernel parameters in Res2-Unet-SE
Conv-kernel Res2Net-SE Encode 32 32 8 128 64 64 16 256 128 128 32 512 Decoder - 128 32 512 128 128 32 512 64 16 64 256
Table 2. PSNR index of different algorithms
View tableView in ArticleTable 2. PSNR index of different algorithms
Method Cameraman House Peppers Starfish Butterfly Airplane Parrot Average BM3D 31.88 34.85 32.75 31.17 31.96 31.15 31.27 32.14 WNNM 32.18 35.15 32.97 32.72 29.85 31.40 31.61 32.26 DNCNN 32.59 34.99 33.24 32.13 33.25 31.67 31.88 32.82 FFDNET 32.37 35.05 33.01 31.95 32.92 31.55 31.79 32.65 DURN 32.41 35.02 33.12 32.10 33.34 31.59 31.97 32.79 MPRNET 32.59 35.13 33.17 32.07 33.54 31.74 31.85 32.87 MSDR 32.64 35.01 32.90 32.23 33.45 31.69 32.04 32.85 BM3D 29.39 32.85 30.20 28.62 29.34 28.52 28.81 29.67 WNNM 29.64 33.22 30.40 29.03 29.85 28.69 29.12 29.99 DNCNN 30.00 33.49 31.09 29.87 30.68 29.24 29.53 30.55 FFDNET 30.05 33.26 30.72 29.28 30.29 29.01 29.42 30.29 DURN 29.77 32.84 30.30 29.19 30.10 28.91 29.11 30.03 MPRNET 30.23 33.51 30.90 29.82 30.58 29.37 29.71 30.58 MSDR 30.21 33.55 30.94 29.97 30.73 29.25 29.59 30.60 BM3D 26.35 29.60 26.76 25.10 25.91 25.39 25.86 26.42 WNNM 26.45 30.33 26.95 25.44 26.32 25.42 26.14 26.72 DNCNN 27.26 29.96 27.35 25.64 26.83 26.83 25.83 27.10 FFDNET 27.24 30.36 27.41 25.68 26.92 25.79 26.57 27.13 DURN 26.86 30.01 27.01 25.65 26.70 25.81 26.14 26.88 MPRNET 27.29 30.25 27.41 25.66 27.09 26.24 26.59 27.21 MSDR 27.37 30.30 27.59 26.02 27.01 25.93 26.67 27.27
Table 3. SSIM index of different algorithms
View tableView in ArticleTable 3. SSIM index of different algorithms
Method Cameraman House Peppers Starfish Butterfly Airplane Parrot Average BM3D 0.776 0 0.806 2 0.789 9 0.751 7 0.821 6 0.774 7 0.777 2 0.785 3 WNNM 0.784 6 0.822 5 0.800 5 0.759 4 0.834 6 0.784 4 0.784 7 0.795 8 DNCNN 0.807 7 0.818 4 0.809 0 0.772 2 0.851 3 0.797 7 0.795 3 0.807 3 FFDNET 0.813 8 0.827 3 0.816 5 0.775 1 0.858 4 0.799 7 0.880 4 0.813 0 DURN 0.797 4 0.822 1 0.805 0 0.771 9 0.844 4 0.794 0 0.783 0 0.802 5 MPRNET 0.812 3 0.824 5 0.813 6 0.774 7 0.861 3 0.798 6 0.799 5 0.812 1 MSDR 0.814 6 0.829 9 0.814 5 0.779 4 0.860 1 0.799 2 0.801 1 0.814 1
Table 3. SSIM index of different algorithms
View tableView in ArticleTable 3. SSIM index of different algorithms
Method Cameraman House Peppers Starfish Butterfly Airplane Parrot Average BM3D 0.900 2 0.886 4 0.906 8 0.901 8 0.938 9 0.901 5 0.894 8 0.904 3 WNNM 0.903 0 0.890 1 0.909 3 0.907 7 0.942 0 0.902 2 0.851 9 0.900 8 DNCNN 0.913 1 0.885 6 0.912 1 0.914 6 0.950 1 0.907 7 0.904 9 0.912 5 FFDNET 0.911 8 0.887 6 0.911 2 0.912 6 0.949 0 0.907 3 0.904 5 0.912 0 DURN 0.909 7 0.886 2 0.910 1 0.906 5 0.934 7 0.905 4 0.902 1 0.907 8 MPRNET 0.913 3 0.887 1 0.911 9 0.913 5 0.952 2 0.918 4 0.904 8 0.914 4 MSDR 0.914 4 0.890 1 0.911 9 0.914 4 0.950 5 0.908 3 0.905 2 0.913 6 BM3D 0.851 3 0.857 7 0.868 3 0.853 2 0.903 5 0.857 4 0.848 9 0.862 9 WNNM 0.859 4 0.860 2 0.872 7 0.857 6 0.908 4 0.861 6 0.851 9 0.867 4 DNCNN 0.840 8 0.852 4 0.875 1 0.853 9 0.916 5 0.854 0 0.839 3 0.861 7 FFDNET 0.875 5 0.862 1 0.874 9 0.865 9 0.920 5 0.870 3 0.862 4 0.875 9 DURN 0.866 7 0.859 1 0.874 9 0.865 7 0.915 2 0.864 9 0.853 5 0.871 4 MPRNET 0.875 9 0.862 7 0.879 6 0.865 4 0.921 1 0.870 3 0.861 8 0.876 7 MSDR 0.876 9 0.863 0 0.874 7 0.866 1 0.922 9 0.871 9 0.867 8 0.877 6
Table 4. PSNR/SSIM index of different algorithms
View tableView in ArticleTable 4. PSNR/SSIM index of different algorithms
Method 指标 场景1 场景2 场景3 Average BM3D PSNR/dB 28.71 30.29 35.89 31.63 SSIM 0.854 6 0.726 2 0.868 9 0.816 6 DNCNN PSNR/dB 25.19 34.11 37.63 32.31 SSIM 0.647 6 0.883 9 0.889 3 0.806 9 DURN PSNR/dB 30.19 36.84 38.80 35.27 SSIM 0.895 7 0.926 4 0.911 0 0.911 0 FFDNET PSNR/dB 31.18 36.45 40.19 35.94 SSIM 0.883 1 0.922 8 0.933 9 0.913 3 CBDNET PSNR/dB 31.28 36.97 40.88 36.38 SSIM 0.894 3 0.934 3 0.946 4 0.925 0 MPRNET PSNR/dB 31.39 37.49 41.74 36.87 SSIM 0.893 4 0.941 1 0.955 9 0.930 1 MSDR PSNR/dB 31.52 37.54 41.52 36.86 SSIM 0.895 9 0.942 0 0.953 1 0.930 3
Table 5. PSNR index of different measure dimensions
View tableView in ArticleTable 5. PSNR index of different measure dimensions
S 2 3 4 5 PSNR/dB 26.94 27.12 27.27 27.25 SSIM 0.803 1 0.812 8 0.814 1 0.813 6
Table 6. PSNR index of different module
View tableView in ArticleTable 6. PSNR index of different module
Method No SAM，NO SE SE SAM SE，SAM PSNR/dB 35.85 37.11 37.27 37.54 SSIM 0.922 9 0.936 4 0.938 5 0.941 7
Table 7. PSNR/SSIM and operation time of different modules
View tableView in ArticleTable 7. PSNR/SSIM and operation time of different modules
Index Method 3×3conv+single-stage 3×3conv+single-stage+Res2Net 3×3conv+muti-stage 3×3conv+muti-stage+Res2Net PSNR/dB 25.27 27.12 26.47 27.27 SSIM 0.773 9 0.813 2 0.792 5 0.814 1 Time/ms 169 194 239 332
Tools
Get Citation
Copy Citation Text
Yan LIU, Gang CHEN, Chunyu YU, Shiyun WANG, Bin SUN. Deep learning image denoising based on multi-stage supervised with Res2-Unet[J]. Optics and Precision Engineering, 2023, 31(6): 920
Paper Information
Category: Information Sciences
Received: Jun. 30, 2022
Accepted: --
Published Online: Apr. 4, 2023
The Author Email: YU Chunyu (yucy@njupt.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20