Acta Optica Sinica, Volume. 37, Issue 12, 1210004(2017)
Method of Rapid Image Super-Resolution Based on Deconvolution
Figures & Tables(12)
Fig. 1. Network architectures of image super-resolution methods based on deep-learning. (a) SRCNN; (b) VDSR; (c) DRCN; (d) ESPCN; (e) FSRCNN; (f) RSRD
Fig. 2. PSNR versus calculation time for different methods performing super-resolution over Set14 with magnification factor of 3
Fig. 3. Trend of mean PSNR of dataset with iteration rising under different layers. (a) Set5; (b) Set14
Fig. 4. Trend of the mean PSNR of dataset with iteration rising under different deconvolution kernel sizes. (a) Set5; (b) Set14
Fig. 5. Trend of the mean PSNR of dataset with iteration rising under different active functions. (a) Set5; (b) Set14
Fig. 8. Whole and local comparisons of foreman.bmp in Set14 processed by different methods with magnification factor of 4. (a) Real image; (b) Bicubic(29.57dB); (c) ANR(30.80 dB); (d) SRCNN(31.50 dB); (e) A+(32.20 dB); (f) ESPCN(32.02 dB); (g) FSRCNN-s(31.52 dB); (h) RSRD(32.89 dB)
Fig. 9. Whole and local comparisons of real image processed by different methods with magnification factor of 4. (a) Real image; (b) Bicubic(29.88 dB); (c) ANR(30.90 dB); (d) SRCNN(31.65 dB); (e) A+(31.30 dB); (f) ESPCN(31.85 dB); (g) FSRCNN-s(31.85 dB); (h) RSRD(32.47 dB)
Table 1. Mean PSNR of test datasets processed by all methods with different magnification factors trained by 91 images
View tableTable 1. Mean PSNR of test datasets processed by all methods with different magnification factors trained by 91 images
Dataset Scale Bicubic RFL SRCNN SRCNN-Ex SCN FSRCNN-s FSRCNN ESPCN RSRD Set5 2 33.66 36.54 36.33 36.67 36.76 36.53 36.94 36.74 37.22 Set14 2 30.23 32.26 32.15 32.35 32.48 32.22 32.54 32.47 32.83 BSD100 2 29.56 31.36 31.32 31.50 31.49 31.30 31.42 31.27 31.58 BSD200 2 29.70 31.38 31.34 31.53 31.63 31.44 31.73 31.53 31.88 Set5 3 30.39 32.43 32.45 32.83 33.04 32.55 33.06 32.55 33.15 Set14 3 27.54 29.05 29.01 29.26 29.37 29.08 29.37 29.09 29.51 BSD100 3 27.21 28.22 28.21 28.4 28.49 28.27 28.4 28.26 28.49 BSD200 3 27.26 28.25 28.27 28.47 28.54 28.32 28.55 28.34 28.58 Set5 4 28.42 30.14 30.15 30.45 30.64 30.04 30.55 30.27 30.74 Set14 4 26.00 27.24 27.21 27.44 27.62 27.12 27.50 27.39 27.69 BSD100 4 25.96 26.75 26.69 26.83 26.95 26.71 26.90 26.81 26.98 BSD200 4 25.97 26.76 26.72 26.88 26.96 26.73 26.92 26.82 26.99
Table 2. Mean PSNR of test datasets processed by all methods with different magnification factors trained by 291 images
View tableTable 2. Mean PSNR of test datasets processed by all methods with different magnification factors trained by 291 images
Dataset Scale RFL Self-Ex SRCNN SCN FSRCNN-s FSRCNN ESPCN VDSR DRCN RSRD Set5 2 36.54 36.67 36.49 36.65 36.71 37.02 36.82 37.49 37.58 37.30 Set14 2 32.28 32.35 32.22 32.29 32.36 32.67 32.58 33.00 33.01 32.96 BSD100 2 31.21 31.49 31.18 31.36 31.50 31.68 31.59 31.85 31.80 31.75 BSD200 2 37.26 31.51 31.20 31.39 31.52 31.82 31.65 31.90 31.86 32.08 Set5 3 32.58 32.58 32.58 32.75 32.81 33.19 32.87 33.62 33.77 33.42 Set14 3 29.13 29.13 29.16 29.28 29.36 29.46 29.40 29.73 29.70 29.74 BSD100 3 28.29 28.30 28.29 28.41 28.41 28.52 28.49 28.77 28.73 28.69 BSD200 3 28.31 28.32 28.34 28.48 28.49 28.64 28.56 28.80 28.76 28.80 Set5 4 30.28 30.3 30.31 30.48 30.53 30.75 30.71 31.30 31.49 31.12 Set14 4 27.32 27.33 27.40 27.49 27.50 27.62 27.51 27.95 27.97 28.00 BSD100 4 26.82 26.86 26.84 26.90 26.94 27.02 26.98 27.20 27.18 27.18 BSD200 4 26.82 26.89 26.88 26.91 29.94 27.04 26.99 27.21 27.20 27.20
Table 3. Comparisons between the proposed RSRD and other methods
View tableTable 3. Comparisons between the proposed RSRD and other methods
Method Network input Layer Deconvolution Pooling Real-time Accuracy Speed SRCNN LR + bicubic 3 No No No No. 7 No. 5 VDSR LR + bicubic 20 No No No No. 2 No. 6 DRCN LR + bicubic 5(recursive) No No No No. 1 No. 7 ESPCN LR 3 No No Yes No. 5 No. 2 FSRCNN-s LR 5 Yes No Yes No. 6 No. 1 FSRCNN LR 8 Yes No No No. 4 No. 4 RSRD(ours) LR 4 Yes Yes Yes No. 3 No. 3
Tools
Get Citation
Copy Citation Text
Chao Sun, Junwei Lü, Jianwei Li, Rongchao Qiu. Method of Rapid Image Super-Resolution Based on Deconvolution[J]. Acta Optica Sinica, 2017, 37(12): 1210004
Paper Information
Category: Image Processing
Received: Jun. 29, 2017
Accepted: --
Published Online: Sep. 6, 2018
The Author Email: Sun Chao (lemony1314@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20