基于多尺度与多重残差网络的图像超分辨率重建 下载: 900次
Super-Resolution Reconstruction of Images Based on Multi-Scale and Multi-Residual Network
成都理工大学信息科学与技术学院(网络安全学院), 四川 成都 610051
图 & 表
图 1. 多尺度与多重残差超分辨率重建网络结构
Fig. 1. Multi-scale and multi-residual super-resolution reconstruction network structure
下载图片 查看原文
图 2. 多尺度残差模块结构
Fig. 2. Multi-scale residual module structure
下载图片 查看原文
图 3. 每个训练周期的PSNR均值变化
Fig. 3. Average PSNR variation for each training cycle
下载图片 查看原文
图 4. 四倍放大因子下不同算法处理后的butterfly图像的超分辨率结果主观对比。(a)原图;(b) BICUBIC算法;(c) ESPCN算法;(d) SRCNN算法;(e) VDSR算法;(f) IMRSR算法
Fig. 4. Subjective comparison of super-resolution results of butterfly images processed by different algorithms under four times magnification factor. (a) GT; (b) BICUBIC algorithm; (c) ESPCN algorithm; (d) SRCNN algorithm; (e) VDSR algorithm; (f) IMRSR algorithm
下载图片 查看原文
图 5. 四倍放大因子下不同算法处理后的PPT图像的超分辨率结果主观对比。(a)原图;(b) BICUBIC算法;(c) ESPCN算法;(d) SRCNN算法;(e) VDSR算法;(f) IMRSR算法
Fig. 5. Subjective comparison of super-resolution results of PPT images processed by different algorithms under four times magnification factor. (a) GT; (b) BICUBIC algorithm; (c) ESPCN algorithm; (d) SRCNN algorithm; (e) VDSR algorithm; (f) IMRSR algorithm
下载图片 查看原文
表 1不同多尺度残差模块数对PSNR的影响
Table1. Effects of number of different multi-scale residual blocks on PSNR
Number of multi-scaleresidual block | PSNR /dB |
---|
Set5 | Set14 |
---|
7 | 31.51 | 28.18 | 11 | 31.52 | 28.14 | 151822 | 31.4631.5931.40 | 28.1628.1928.06 |
|
查看原文
表 2不同算法在不同测试集对图片进行不同超分辨率重建的PSNR平均值
Table2. Average PSNR of different algorithms for different super-resolution reconstructions on different test setsunit: dB
Dataset | Scale | BICUBIC | Self-Ex | SRCNN | ESPCN | VDSR | IMRSR |
---|
Set5 | ×2×3×4 | 33.6430.3828.42 | 36.4932.5830.31 | 36.4532.3830.19 | 36.5732.5530.31 | 37.3033.4430.98 | 37.7833.9131.59 | Set14 | ×2×3×4 | 30.0827.3825.86 | 32.2229.1627.40 | 32.3729.1527.37 | 32.4729.2727.48 | 32.9729.6927.83 | 33.2629.8828.19 | BSD100 | ×2×3×4 | 29.5927.2025.96 | 31.1828.2926.84 | 31.2528.2426.80 | 31.2928.3226.85 | 31.7728.7027.14 | 32.0028.8027.30 | Urban100 | ×2×3×4 | 26.8624.4423.13 | 29.5426.4424.79 | 29.0825.7924.20 | 29.2125.9424.28 | 30.4626.8724.95 | 31.0027.0025.15 |
|
查看原文
表 3不同算法在不同测试集对图片进行不同超分辨率重建的SSIM平均值
Table3. Average SSIM of different algorithms for different super-resolution reconstructions on different test sets
Dataset | Scale | BICUBIC | Self-Ex | SRCNN | ESPCN | VDSR | IMRSR |
---|
Set5 | ×2×3×4 | 0.93620.87920.8223 | 0.95370.90930.8619 | 0.95740.91170.8639 | 0.95850.91500.8672 | 0.96250.92760.8875 | 0.96430.93120.8957 | Set14 | ×2×3×4 | 0.88070.79250.7210 | 0.90340.81960.7518 | 0.91410.83440.7665 | 0.91530.83690.7690 | 0.91990.84560.7817 | 0.92270.84880.7892 | BSD100 | ×2×3×4 | 0.85750.75840.6856 | 0.88550.78400.7106 | 0.89740.80280.7297 | 0.89790.80430.7308 | 0.90410.81340.7413 | 0.90730.81660.7469 | Urban100 | ×2×3×4 | 0.84960.75090.6739 | 0.89670.80880.7174 | 0.89630.80070.7249 | 0.89780.80540.7287 | 0.91700.83500.7614 | 0.92350.84030.7714 |
|
查看原文
陈星宇, 张伟劲, 孙伟智, 任萍安, 欧鸥. 基于多尺度与多重残差网络的图像超分辨率重建[J]. 激光与光电子学进展, 2020, 57(18): 181009. Xingyu Chen, Weijin Zhang, Weizhi Sun, Ping'an Ren, Ou Ou. Super-Resolution Reconstruction of Images Based on Multi-Scale and Multi-Residual Network[J]. Laser & Optoelectronics Progress, 2020, 57(18): 181009.