Laser & Optoelectronics Progress, Volume. 60, Issue 14, 1410010(2023)
Lightweight Attention-Guided Network for Image Super-Resolution
Figures & Tables(10)
Fig. 3. Main structure of AG module. (a) ECA module structure; (b) SGE module structure
Fig. 5. Performance and parameters comparison of existing lightweight image super-resolution methods on Set5 dataset with the magnification of 4×
Fig. 6. Comparison of reconstruction effects of different methods on images with a magnification factor of 2 in the BSD100 dataset
Fig. 7. Comparison of reconstruction effects of different methods on images with magnification factor of 4 in Set5 dataset
Table 1. Influence of the number of AG modules on the network
View tableTable 1. Influence of the number of AG modules on the network
Parameter Dataset NAG =8 NAG =12 NAG =16 NAG =20 NAG =24 Set5 33.935 33.961 34.261 34.257 34.227 PSNR /dB Set14 29.312 30.117 30.224 29.941 29.725 BSD100 28.452 28.625 28.933 28.561 28.027 SSIM Set5 0.9105 0.9187 0.9253 0.9253 0.9236 Set14 0.8216 0.8369 0.8421 0.8297 0.8262 BSD100 0.7914 0.7963 0.8024 0.7951 0.7855 Parameters /103 376 422 456 541.2 577.3 Multi-Adds /109 47.7 61.6 75.4 89.3 103.4
Table 2. Influence of AG module structure on the network
View tableTable 2. Influence of AG module structure on the network
AG module structure Dataset Scale PSNR /dB SSIM Parameters /103 Multi-Adds /109 ECA+SE Set5 2 36.572 0.9538 429 68.5 SGE+SE Set5 2 36.937 0.9546 502 85.2 ECA+SGE Set5 2 37.584 0.9566 553 89.6 ECA+SGE+atg Set5 2 37.792 0.9594 447 83.6 ECA+SE Set14 3 29.786 0.8317 433 77.9 SGE+SE Set14 3 29.592 0.8295 505 93.3 ECA+SGE Set14 3 30.109 0.8403 526 86.1 ECA+SGE+atg Set14 3 30.224 0.8421 456 75.4 ECA+SE BSD100 4 27.267 0.7254 439 91.6 SGE+SE BSD100 4 27.294 0.7263 509 107.6 ECA+SGE BSD100 4 27.314 0.7262 506 84.2 ECA+SGE+atg BSD100 4 27.541 0.7347 470 67.9
Table 3. Test results of different methods on Set5, Set14, and BSD100 datasets
View tableTable 3. Test results of different methods on Set5, Set14, and BSD100 datasets
Method Scale Parameters /103 Multi-
Adds /109
PSNR /dB SSIM Set5 Set14 BSD100 Set5 Set14 BSD100 Bicubic 2 33.66 30.24 29.56 0.9299 0.8688 0.8431 SRCNN 57 52.7 36.34 32.45 31.36 0.9521 0.9067 0.8879 FSRCNN 12 6.0 36.94 32.63 31.53 0.9558 0.9088 0.8920 VDSR 665 612.6 37.53 33.03 31.90 0.9587 0.9124 0.8960 LapSRN 813 29.9 37.52 33.08 31.80 0.9591 0.9130 0.8950 DRCN 1774 17974 37.63 33.04 31.85 0.9588 0.9118 0.8942 DRRN 297 6796.9 37.74 33.23 32.05 0.9591 0.9136 0.8973 CARN-M 412 91.2 37.53 33.26 31.92 0.9583 0.9141 0.8960 LESRCNN 516 110 37.65 33.32 31.95 0.9586 0.9148 0.8965 FALSR-B 326 74.7 37.61 33.29 31.97 0.9585 0.9143 0.8967 FALSR-C 408 93.7 37.66 33.26 31.96 0.9586 0.9140 0.8965 ACNet 1356 501.5 37.72 33.41 32.06 0.9588 0.9160 0.8978 LAGNet 447 83.6 37.79 33.40 32.10 0.9594 0.9162 0.8991 Bicubic 3 30.39 27.55 27.21 0.8682 0.7742 0.7385 SRCNN 57 52.7 32.39 29.30 28.41 0.9033 0.8215 0.7863 FSRCNN 12 5.0 33.16 29.43 28.53 0.9140 0.8242 0.7910 VDSR 665 612.6 33.66 29.77 28.82 0.9213 0.8314 0.7976 DRCN 1774 17974 33.82 29.76 28.80 0.9226 0.8311 0.7963 DRRN 297 6796.9 34.03 29.96 28.95 0.9244 0.8349 0.8004 CARN-M 412 46.1 33.99 30.08 28.91 0.9236 0.8367 0.8000 LESRCNN 516 49.1 33.93 30.12 28.91 0.9231 0.8380 0.8005 ACNet 1541 369 34.14 30.19 28.98 0.9247 0.8398 0.8023 LAGNet 456 75.4 34.26 30.22 28.93 0.9253 0.8421 0.8024 Bicubic 4 28.42 26.00 25.96 0.8104 0.7027 0.6675 SRCNN 57 52.7 30.09 27.50 26.90 0.8530 0.7513 0.7101 FSRCNN 12 4.6 30.71 27.59 26.98 0.8657 0.7535 0.7150 VDSR 665 612.6 31.35 28.01 27.29 0.8838 0.7674 0.7251 LapSRN 813 149.4 31.54 28.19 27.32 0.8850 0.7720 0.7270 DRCN 1774 17974 31.53 28.02 27.23 0.8854 0.7670 0.7233 DRRN 297 6796.9 31.68 28.21 27.38 0.8888 0.7720 0.7284 s-LWSR16 144 8.3 31.63 27.92 27.35 0.8869 0.7701 0.7287 CARN-M 412 32.5 31.92 28.42 27.44 0.8903 0.7762 0.7304 LESRCNN 516 28.6 31.88 28.44 27.45 0.8903 0.7772 0.7313 ACNet 1784 347.9 31.83 28.46 27.48 0.8903 0.7788 0.7326 LAGNet 470 67.9 32.06 28.47 27.54 0.8912 0.7782 0.7347
Tools
Get Citation
Copy Citation Text
Zixuan Ding, Juan Zhang, Xiang Li, Xinyu Wang. Lightweight Attention-Guided Network for Image Super-Resolution[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1410010
Paper Information
Category: Image Processing
Received: Jun. 29, 2022
Accepted: Sep. 5, 2022
Published Online: Jul. 14, 2023
The Author Email: Zhang Juan (zhang-j@foxmail.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20