基于残差通道注意力和多级特征融合的图像超分辨率重建 下载: 1232次
Super-Resolution Image Reconstruction Based on Residual Channel Attention and Multilevel Feature Fusion
哈尔滨工程大学信息与通信工程学院, 黑龙江 哈尔滨 150001
图 & 表
图 1. ESPCN网络结构图
Fig. 1. Structure of ESPCN network
下载图片 查看原文
图 2. VDSR网络结构图
Fig. 2. Structure of VDSR network
下载图片 查看原文
图 3. 本文算法的网络结构图
Fig. 3. Network structure of proposed algorithm
下载图片 查看原文
图 4. 递归单元模块。(a)递归单元模块的组成;(b)残差通道注意力;(c)多级特征融合
Fig. 4. Recursive unit module. (a) Recursive unit module composition; (b) residual channel attention; (c) multilevel feature fusion
下载图片 查看原文
图 5. Set5测试集下不同层数时PSNR均值随迭代次数的变化
Fig. 5. Variation of mean PSNR with the number of iterations for different layers at Set5 test set
下载图片 查看原文
图 6. Set5测试集下不同网络结构的参数数量与PSNR均值的关系
Fig. 6. Relationship between number of parameters of different network structures and mean PSNR at Set5 test set
下载图片 查看原文
图 7. Set5测试集下不同方法的运行时间与PSNR均值之间的关系
Fig. 7. Relationship between run time of different methods and mean PSNR at Set5 test set
下载图片 查看原文
图 8. 采用不同算法恢复出来的斑马图像
Fig. 8. Comparison of zebra image recovered with different algorithms
下载图片 查看原文
图 9. 采用不同算法恢复出来的ppt图像的对比
Fig. 9. Comparison of ppt image recovered with different algorithms
下载图片 查看原文
表 1在Set5测试集下不同RCAF模型组件的PSNR均值
Table1. Means PSNR of different RCAF model components at Set 5 test set
Multilevelfeature fusion | Residualchannel attention | PSNR |
---|
√ | √ | 31.61 | × | √ | 31.35 | √ | × | 31.40 | × | × | 30.94 |
|
查看原文
表 2测试集Set5、Set14、BSD100下不同算法的PSNR均值
Table2. Mean PSNR of different algorithms at Set5, Set14, and BSD100 test sets
Test set | Scale | Bicubic | SRCNN[8] | ESPCN[10] | FSRCNN[9] | VDSR[11] | RCAF |
---|
| 2 | 33.68 | 36.19 | 36.38 | 36.45 | 37.34 | 37.62 | Set5 | 3 | 30.45 | 32.46 | 32.71 | 32.59 | 33.47 | 34.00 | | 4 | 28.46 | 30.15 | 30.29 | 30.42 | 30.78 | 31.61 | | 2 | 30.21 | 32.10 | 32.20 | 32.21 | 32.82 | 33.24 | Set14 | 3 | 27.51 | 28.99 | 29.12 | 29.12 | 29.51 | 29.87 | | 4 | 25.98 | 27.23 | 27.17 | 27.43 | 27.62 | 28.11 | | 2 | 29.43 | 30.88 | 30.93 | 31.24 | 31.51 | 31.81 | BSD100 | 3 | 27.08 | 28.06 | 28.16 | 28.25 | 28.43 | 28.72 | | 4 | 25.84 | 26.63 | 26.59 | 26.85 | 26.87 | 27.18 |
|
查看原文
表 3测试集Set5、Set14、BSD100下不同算法的SSIM均值
Table3. Mean SSIM of different algorithms at test sets Set5, Set14, and BSD100
Dataset | Scale | Bicubic | SRCNN[8] | ESPCN[10] | FSRCNN[9] | VDSR[11] | RCAF |
---|
| 2 | 0.931 | 0.955 | 0.957 | 0.957 | 0.958 | 0.963 | Set5 | 3 | 0.869 | 0.911 | 0.915 | 0.912 | 0.919 | 0.930 | | 4 | 0.810 | 0.862 | 0.863 | 0.866 | 0.875 | 0.893 | | 2 | 0.869 | 0.958 | 0.960 | 0.963 | 0.910 | 0.964 | Set14 | 3 | 0.774 | 0.884 | 0.887 | 0.892 | 0.827 | 0.897 | | 4 | 0.702 | 0.821 | 0.823 | 0.827 | 0.759 | 0.842 | | 2 | 0.844 | 0.880 | 0.883 | 0.887 | 0.892 | 0.897 | BSD100 | 3 | 0.740 | 0.776 | 0.781 | 0.780 | 0.793 | 0.797 | | 4 | 0.670 | 0.692 | 0.694 | 0.701 | 0.719 | 0.720 |
|
查看原文
席志红, 袁昆鹏. 基于残差通道注意力和多级特征融合的图像超分辨率重建[J]. 激光与光电子学进展, 2020, 57(4): 041504. Zhihong Xi, Kunpeng Yuan. Super-Resolution Image Reconstruction Based on Residual Channel Attention and Multilevel Feature Fusion[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041504.