Optics and Precision Engineering, Volume. 31, Issue 17, 2584(2023)
Multidimensional attention mechanism and selective feature fusion for image super-resolution reconstruction
Figures & Tables(13)
Fig. 1. Schematic diagram of network structure integrating multi-dimensional attention mechanism and selective feature fusion
Fig. 2. Schematic diagram of selective kernel feature fusion module
Fig. 3. Schematic diagram of internal principle junction of heterogeneous residual module
Fig. 4. Comparison of PSNR ablation on Urban100 dataset magnified 2 times
Fig. 5. Comparison of Img015 magnified 2 times in Urban100 dataset
Fig. 6. Comparison of Img030 magnified 3 times in Urban100 dataset
Fig. 7. Comparison of Img039 magnified 4 times in Urban100 dataset
Fig. 8. PSNR and parameters on Set5（×4）dataset of different models
Table 1. Comparison results of different methods on Urban100 data set magnified by 2 times
View tableView in ArticleTable 1. Comparison results of different methods on Urban100 data set magnified by 2 times
方法/评价指标 对称组卷积模块 互补卷积模块 选择性内核
特征融合模块
特征增强
残差模块
PSNR/dB SSIM Plan A √ √ × × 30.43 0.921 6 Plan B √ × √ × 31.26 0.915 7 Plan C √ × × √ 31.15 0.911 7 Plan D √ √ √ × 31.79 0.924 5 Plan E √ √ √ √ 32.12 0.928 8
Table 2. Comparison results of different methods on 4 Baseline datasets enlarged by 2 times
View tableView in ArticleTable 2. Comparison results of different methods on 4 Baseline datasets enlarged by 2 times
算 法 Set5 Set14 BSDS100 Urban100 PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM Bicubic［ 32 ］33.66 0.929 9 30.24 0.868 8 29.56 0.841 26.88 0.840 3 A+［ 33 ］36.54 0.954 4 32.28 0.905 6 31.21 0.886 3 29.20 0.893 8 RFL［ 34 ］36.54 0.953 7 32.26 0.904 0 31.16 0.884 0 29.11 0.890 4 DASR［ 35 ］37.22 0.951 5 32.93 0.902 9 31.76 0.890 1 30.63 0.907 9 KXNet［ 37 ］37.59 0.955 2 33.36 0.909 1 32.03 0.894 1 31.48 0.919 2 RED30［ 38 ］37.66 0.959 9 32.94 0.914 4 31.99 0.897 4 30.91 0.915 9 DnCNN［ 36 ］37.58 0.959 0 33.03 0.912 8 31.90 0.896 1 30.74 0.913 9 MemNet［ 39 ］37.78 0.959 7 33.28 0.914 2 32.08 0.897 8 31.31 0.919 5 CARN［ 19 ］37.76 0.959 0 33.52 0.916 6 31.92 0.897 8 31.51 0.931 2 CARN-M［ 19 ］37.53 0.958 3 33.26 0.914 1 32.09 0.896 0 30.83 0.923 3 EEDS+［ 15 ］37.78 0.960 9 33.21 0.915 1 31.95 0.896 3 31.92 0.928 9 MSDEPC［ 41 ］37.39 0.957 6 32.94 0.911 1 31.64 0.896 1 31.83 0.928 3 CFSRCNN［ 42 ］37.79 0.959 1 33.51 0.916 5 32.11 0.898 8 32.07 0.819 0 LESRCNN［ 22 ］37.65 0.958 6 33.32 0.914 8 31.95 0.896 4 31.45 0.920 6 LESRCNN-S［ 22 ］37.57 0.958 2 33.30 0.914 5 31.95 0.896 5 31.45 0.920 7 ACNet［ 43 ］37.72 0.958 8 33.41 0.916 0 32.06 0.897 8 31.79 0.924 5 ACNet-B［ 43 ］37.60 0.958 4 33.32 0.915 1 31.97 0.897 0 31.57 0.922 2 Ours 37.83 0.959 4 33.53 0.917 2 32.14 0.898 7 32.12 0.930 6
Table 3. Comparison results of different methods on 4 Baseline datasets enlarged by 3 times
View tableView in ArticleTable 3. Comparison results of different methods on 4 Baseline datasets enlarged by 3 times
算 法 Set5 Set14 BSDS100 Urban100 PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM Bicubic［ 32 ］30.39 0.868 2 27.55 0.774 2 27.21 0.738 5 24.46 0.734 9 A+［ 33 ］32.58 0.908 8 29.13 0.818 8 28.29 0.783 5 26.03 0.797 3 RFL［ 34 ］32.43 0.905 7 29.05 0.816 4 28.22 0.780 6 25.86 0.790 0 DASR［ 35 ］33.78 0.920 0 29.94 0.826 6 28.85 0.793 2 27.28 0.830 7 KXNet［ 37 ］34.22 0.923 8 30.27 0.834 0 29.06 0.801 0 28.00 0.845 7 RED30［ 38 ］33.82 0.929 9 29.61 0.834 1 28.93 0.799 4 27.31 0.830 3 DnCNN［ 36 ］33.75 0.929 0 29.81 0.832 1 28.85 0.798 1 27.15 0.827 6 MemNet［ 39 ］34.09 0.924 8 30.00 0.835 0 28.96 0.800 1 27.56 0.837 6 CARN［ 19 ］33.99 0.923 6 30.08 0.836 7 28.91 0.800 0 26.86 0.826 3 CARN-M［ 19 ］34.29 0.925 5 30.29 0.840 7 29.06 0.803 4 27.38 0.840 4 EEDS+［ 15 ］33.81 0.925 2 29.85 0.833 9 28.88 0.805 4 27.43 0.835 9 MSDEPC［ 41 ］33.37 0.918 4 29.62 0.827 9 29.58 0.791 8 27.23 0.819 0 CFSRCNN［ 42 ］33.24 0.925 6 30.27 0.841 0 29.03 0.803 5 28.04 0.849 6 LESRCNN［ 22 ］34.93 0.923 1 30.12 0.838 0 28.91 0.800 5 27.70 0.841 5 LESRCNN-S［ 22 ］33.05 0.923 8 30.16 0.838 4 28.94 0.801 2 27.76 0.842 4 ACNet［ 43 ］34.14 0.924 7 30.19 0.839 8 28.98 0.802 3 27.97 0.848 2 ACNet-B［ 43 ］34.07 0.924 3 30.15 0.838 6 28.97 0.801 6 27.88 0.844 7 Ours 34.35 0.926 8 30.36 0.841 7 29.11 0.804 7 28.19 0.853 1
Table 4. Comparison results of different methods on 4 Baseline datasets enlarged by 4 times
View tableView in ArticleTable 4. Comparison results of different methods on 4 Baseline datasets enlarged by 4 times
算 法 Set5 Set14 BSDS100 Urban100 PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM Bicubic［ 32 ］28.42 0.810 4 26.00 0.702 7 25.96 0.667 5 23.14 0.657 7 A+［ 33 ］30.28 0.860 3 27.32 0.749 1 26.82 0.708 7 24.32 0.718 3 RFL［ 34 ］30.14 0.854 8 27.24 0.745 1 26.82 0.705 4 24.19 0.709 6 DASR［ 35 ］31.68 0.885 4 28.26 0.766 8 27.40 0.728 6 25.49 0.762 1 KXNet［ 37 ］31.94 0.891 2 28.61 0.778 2 27.59 0.733 0 26.18 0.787 3 RED30［ 38 ］31.51 0.886 9 27.86 0.771 8 27.40 0.729 0 25.35 0.758 7 DnCNN［ 36 ］31.40 0.884 5 28.04 0.767 2 27.29 0.725 3 25.20 0.752 1 MemNet［ 39 ］31.74 0.889 3 28.26 0.772 3 27.40 0.728 1 25.50 0.763 0 CARN［ 19 ］31.92 0.890 3 28.42 0.776 2 27.44 0.730 4 25.63 0.768 8 CARN-M［ 19 ］32.13 0.893 7 28.60 0.780 6 27.58 0.734 9 26.07 0.783 7 EEDS+［ 15 ］31.53 0.886 9 28.13 0.769 8 27.35 0.726 3 25.45 0.762 9 MSDEPC［ 41 ］31.05 0.879 7 27.79 0.758 1 27.10 0.719 3 25.23 0.758 6 CFSRCNN［ 42 ］32.06 0.892 9 28.57 0.780 0 27.53 0.733 3 26.03 0.782 4 LESRCNN［ 22 ］31.88 0.890 3 28.44 0.777 2 27.45 0.731 3 25.77 0.773 2 LESRCNN-S［ 22 ］31.88 0.890 7 28.43 0.777 6 27.47 0.732 1 25.78 0.773 9 ACNet［ 43 ］31.83 0.890 3 28.46 0.778 8 27.48 0.732 6 25.93 0.779 8 ACNet-B［ 43 ］31.82 0.890 1 28.41 0.777 3 27.46 0.731 6 25.86 0.776 0 Ours 32.11 0.894 3 28.63 0.781 2 27.65 0.735 2 26.07 0.790 1
Table 5. Inference time of different models on Set5（×4）dataset
View tableView in ArticleTable 5. Inference time of different models on Set5（×4）dataset
算 法 推理时间/s CARN 0.032 0 CFSRCNN 0.029 8 MemNet 14.690 0 ACNet 0.031 5 DASR 0.072 1 KXNet 0.642 9 Ours 0.051 6
Tools
Get Citation
Copy Citation Text
Jian WEN, Jianfei SHAO, Jie LIU, Jianlong SHAO, Yuhang FENG, Rong YE. Multidimensional attention mechanism and selective feature fusion for image super-resolution reconstruction[J]. Optics and Precision Engineering, 2023, 31(17): 2584
Paper Information
Category: Information Sciences
Received: Nov. 22, 2022
Accepted: --
Published Online: Oct. 9, 2023
The Author Email: SHAO Jianfei (469365367@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20