Journal of Infrared and Millimeter Waves, Volume. 44, Issue 3, 431(2025)
Sparsity and self-similarity priors guided deep learning for blind image super-resolution
Figures & Tables(13)
Fig. 3. Structure of a neural network layer that simulates one iteration of FISTA
Fig. 5. PSNR values variation curve with iterations on Urban100 and Manga109 datasets
Table 1. The pseudocode for the DLS3P algorithm in training process
View tableView in ArticleTable 1. The pseudocode for the DLS3P algorithm in training process
DLS3P算法训练过程的伪代码 输入：低分辨率图像 DUDF模块展开层数 ，DFRG组数 ，DMFB块数 训练的总迭代次数 ，对称损失权重 循环迭代 次： 步骤1： 步骤2： 步骤3： 步骤4： 步骤5： 步骤6：反向传播更新参数 输出：训练好的网络
Table 2. Quantitative results of different algorithms on Gaussian8 dataset
View tableView in ArticleTable 2. Quantitative results of different algorithms on Gaussian8 dataset
Method Scale Set5 Set14 BSD100 Urban100 Manga109 Average PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM Bicubic 2 28.82 0.8577 26.02 0.7634 25.92 0.7310 23.14 0.7258 25.60 0.8498 25.90 0.7855 CARN［ 4 ］30.99 0.8779 28.10 0.7879 26.78 0.7286 25.27 0.7630 26.86 0.8606 27.60 0.8036 Deblur［ 31 ］+CARN［4 ］24.20 0.7496 21.12 0.6170 22.69 0.6471 18.89 0.5895 21.54 0.7946 21.69 0.6796 CARN［ 4 ］+Deblur［31 ］31.27 0.8974 29.03 0.8267 28.72 0.8033 25.62 0.7981 29.58 0.9134 28.84 0.8478 Bicubic+ZSSR［ 13 ］31.08 0.8786 28.35 0.7933 27.92 0.7632 25.25 0.7618 28.05 0.8769 28.13 0.8148 IKC［ 16 ］37.19 0.9526 32.94 0.9024 31.51 0.8790 29.85 0.8928 36.93 0.9667 33.68 0.9187 DANv1［ 19 ］37.34 0.9526 33.08 0.9041 31.76 0.8858 30.60 0.9060 37.23 0.9710 34.00 0.9239 DANv2［ 20 ］37.60 0.9544 33.44 0.9094 32.00 0.8904 31.43 0.9174 38.07 0.9734 34.51 0.9290 DCLS［ 23 ］37.63 0.9554 33.46 0.9103 32.04 0.8907 31.69 0.9202 38.31 0.9740 34.63 0.9301 DLS3P（Ours） 37.64 0.9546 33.65 0.9105 32.11 0.8916 31.90 0.9220 38.51 0.9744 34.76 0.9306 Bicubic 3 26.21 0.7766 24.01 0.6662 24.25 0.6356 21.39 0.6203 22.98 0.7576 23.77 0.6913 CARN［ 4 ］27.26 0.7855 25.06 0.6676 25.85 0.6566 22.67 0.6323 23.85 0.7620 24.94 0.7008 Deblur［ 31 ］+CARN［4 ］19.05 0.5226 17.61 0.4558 20.51 0.5331 16.72 0.5895 18.38 0.6118 18.45 0.5426 CARN［ 4 ］+Deblur［31 ］30.31 0.8562 27.57 0.7531 27.14 0.7152 24.45 0.7241 27.67 0.8592 27.43 0.7816 Bicubic+ZSSR［ 13 ］28.25 0.7989 26.15 0.6942 26.06 0.6633 23.26 0.6534 25.19 0.7914 25.78 0.7202 IKC［ 16 ］33.06 0.9146 29.38 0.8233 28.53 0.7899 24.43 0.8302 32.43 0.9316 29.57 0.8579 DANv1［ 19 ］34.04 0.9199 30.09 0.8287 28.94 0.7919 27.65 0.8352 33.16 0.9382 30.78 0.8628 DANv2［ 20 ］34.12 0.9209 30.20 0.8309 29.03 0.7948 27.83 0.8395 33.28 0.9400 30.89 0.8652 DCLS［ 23 ］34.21 0.9218 30.29 0.8329 29.07 0.7956 28.03 0.8444 33.54 0.9414 31.03 0.8672 DLS3P（Ours） 34.15 0.9219 30.31 0.8341 29.12 0.7979 28.21 0.8486 33.81 0.9429 31.12 0.8690 Bicubic 4 24.57 0.7108 22.79 0.6032 23.29 0.5786 20.35 0.5532 21.50 0.6933 22.50 0.6278 CARN［ 4 ］26.57 0.7420 24.62 0.6226 24.79 0.5963 22.17 0.5865 21.85 0.6834 24.00 0.6462 Deblur［ 31 ］+CARN［4 ］18.10 0.4843 16.59 0.3994 18.46 0.4481 15.47 0.3872 16.78 0.5371 17.08 0.4512 CARN［ 4 ］+Deblur［31 ］28.69 0.8092 26.40 0.6926 26.10 0.6528 23.46 0.6597 25.84 0.8035 26.10 0.7236 Bicubic+ZSSR［ 13 ］26.45 0.7279 24.78 0.6268 24.97 0.5989 22.11 0.5805 23.53 0.7240 24.37 0.6516 IKC［ 16 ］31.67 0.8829 28.31 0.7643 27.37 0.7192 25.33 0.7504 28.91 0.8782 28.32 0.7990 DANv1［ 19 ］31.89 0.8864 28.42 0.7687 27.51 0.7248 25.86 0.7721 30.50 0.9037 28.84 0.8111 DANv2［ 20 ］32.00 0.8885 28.50 0.7715 27.56 0.7277 25.94 0.7748 30.45 0.9037 28.89 0.8132 AdaTarget［ 17 ］31.58 0.8814 28.14 0.7626 27.43 0.7216 25.72 0.7683 29.97 0.8955 28.57 0.8059 DCLS［ 23 ］32.12 0.8890 28.54 0.7728 27.60 0.7285 26.15 0.7809 30.86 0.9086 29.05 0.8160 DLS3P（Ours） 32.13 0.8906 28.62 0.7752 27.63 0.7305 26.24 0.7849 30.94 0.9086 29.11 0.8180
Table 3. Quantitative results of different algorithms on DIV2KRK dataset
View tableView in ArticleTable 3. Quantitative results of different algorithms on DIV2KRK dataset
Method DIV2KRK PSNR SSIM Bicubic 25.33 0.679 5 EDSR［ 2 ］ （2017）25.64 0.692 8 DBPN［ 3 ］ （2018）25.58 0.691 0 RCAN［ 5 ］ （2018）25.66 0.693 6 Bicubic+ZSSR［ 13 ］ （2019）25.61 0.691 1 KernelGAN［ 9 ］+ZSSR［13 ］ （2019）26.81 0.731 6 KernelGAN［ 9 ］+SRMD［14 ］ （2019）27.51 0.726 5 KernelGAN［ 9 ］+USRNet［15 ］ （2020）20.06 0.535 9 IKC［ 16 ］ （2019）27.70 0.766 8 DANv1［ 19 ］ （2020）27.55 0.758 2 DANv2［ 20 ］ （2021）28.74 0.789 3 AdaTarget［ 17 ］ （2021）28.42 0.785 4 KOALAnet［ 18 ］ （2021）27.77 0.763 7 DCLS［ 23 ］ （2022）28.99 0.794 6 KernelDAN［ 21 ］ （2023）28.90 0.796 1 DLS3P（Ours） 29.08 0.800 7
Table 4. PSNR values variation with the number of deconvolution feature channels
View tableView in ArticleTable 4. PSNR values variation with the number of deconvolution feature channels
反卷积特征通道数 16 （DCLS） 16 32 64 Urban100 31.69 31.72 31.78 31.80
Table 5. PSNR values variation with different receptive field features
View tableView in ArticleTable 5. PSNR values variation with different receptive field features
9 9特征 15 15特征 21 21特征 Urban100 31.78 31.85 31.90
Table 6. Comparison of the parameter and computational complexity among the different models
View tableView in ArticleTable 6. Comparison of the parameter and computational complexity among the different models
方法 参数量 （M） FLOPs （G） IKC （2019） 5.29 2178.72 DANv1 （2020） 4.33 926.72 DANv2 （2021） 4.71 918.12 DCLS （2022） 19.05 DLS3P （Ours） 34.40 1224.00
Tools
Get Citation
Copy Citation Text
Sun-Yi GE, Xiao-Wei LUO, Shi-Yang FENG, Bin WANG. Sparsity and self-similarity priors guided deep learning for blind image super-resolution[J]. Journal of Infrared and Millimeter Waves, 2025, 44(3): 431
Paper Information
Category: Interdisciplinary Research on Infrared Science
Received: Oct. 22, 2024
Accepted: --
Published Online: Jul. 9, 2025
The Author Email: Bin WANG (wangbin@fudan.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20