Fig. 1. Structure of the generator
Fig. 2. Network structure after removing the BN layer
Fig. 3. Structure of the RRDB
Fig. 4. Structure of the discriminator
Fig. 5. Schematic diagram of the training process. (a) Actual training curve; (b) ideal training curve
[16] Fig. 6. Training environment of the network
Fig. 7. Interface of the MOI test system
Fig. 8. PSNR of different algorithms in the Set5 test set
Fig. 9. SSIM of different algorithms on the Set5 test set
Fig. 10. PSNR of different algorithms on the Set14 test set
Fig. 11. SSIM of different algorithms in the Set14 test set
Fig. 12. Reconstruction effects of two algorithms. (a) Original image; (b) SRGAN algorithm; (c) our algorithm
Fig. 13. Reconstruction results of 5 different algorithms. (a) Overall original image; (b) bicubic interpolation algorithm;(c) SRCNN algorithm; (d) VDSR algorithm; (e) SRResNet algorithm; (f) our algorithm; (g) partial original image
Fig. 14. Railroad track image reconstructed by 5 different algorithms. (a) Overall original image; (b) bicubic interpolation algorithm; (c) SRCNN algorithm; (d) VDSR algorithm; (e) SRResNet algorithm; (f) our algorithm; (g) partial original image
Evaluation standard | Score |
---|
No change in image quality | 5 | Slight change in image quality can be seen | 4 | Slightly hinder viewing | 3 | Hinder viewing | 2 | Seriously obstructing viewing | 1 |
|
Table 1. Evaluation standard of the image quality
Relative measurement scale | Score |
---|
Quality is the worst in this picture group | 1 | Quality is below average in this picture group | 2 | Quality is on average in this picture group | 3 | Quality is above average in this picture group | 4 | Quality is the best in this picture group | 5 |
|
Table 2. Scoring table for subjective evaluation of image quality
Algorithm | SRGAN[13] | Ours | Difference |
---|
PSNR/dB | 28.74 | 29.60 | ↑0.86 | SSIM | 0.8435 | 0.8558 | ↑0.0123 |
|
Table 3. Test results of different algorithms on the Set5 data set
Algorithm | SRGAN[13] | Ours | Difference |
---|
PSNR/dB | 25.75 | 26.44 | ↑0.69 | SSIM | 0.7370 | 0.7460 | ↑0.0090 |
|
Table 4. Test results of different algorithms on the Set14 data set
Algorithm | SRGAN[13] | Ours | Difference |
---|
PSNR/dB | 24.65 | 25.55 | ↑0.90 | SSIM | 0.6502 | 0.6549 | ↑0.0047 |
|
Table 5. Test results of different algorithms on the BSD100 data set
Data set | SRGAN(with BN) | Ours(with BN) | SRGAN(without BN) | Ours(without BN) | |
---|
PSNR /dB | SSIM | Time /s | PSNR /dB | SSIM | Time /s | PSNR /dB | SSIM | Time /s | PSNR /dB | SSIM | Time /s | |
---|
Set5 | 28.69 | 0.8415 | 0.21 | 29.02 | 0.8489 | 0.20 | 29.25 | 0.8483 | 0.20 | 29.60 | 0.8547 | 0.18 | | Set14 | 24.96 | 0.7187 | 0.43 | 26.10 | 0.7233 | 0.39 | 25.67 | 0.7203 | 0.42 | 26.41 | 0.7398 | 0.38 | | BSD100 | 24.01 | 0.6488 | 0.45 | 25.15 | 0.6511 | 0.43 | 25.10 | 0.6503 | 0.44 | 25.52 | 0.6548 | 0.41 | |
|
Table 6. Influence of BN layer on algorithm performance
Dataset | Algorithm | Bicubic | SRCNN[2] | VDSR[5] | SRResNet[13] | Ours |
---|
| PSNR /dB | 28.43 | 30.14 | 31.35 | 31.92 | 29.60 | Set5 | SSIM | 0.8211 | 0.8647 | 0.8838 | 0.8998 | 0.8558 | | MOI | 1.44 | 2.4 | 3.18 | 3.28 | 4.66 | | PSNR /dB | 25.99 | 27.18 | 28.01 | 28.39 | 26.44 | Set14 | SSIM | 0.7486 | 0.7861 | 0.7674 | 0.8116 | 0.746 | | MOI | 1.42 | 2.43 | 3.18 | 3.59 | 4.38 | | PSNR /dB | 25.96 | 26.9 | 27.29 | 27.52 | 25.55 | BSD100 | SSIM | 0.6675 | 0.7101 | 0.7251 | 0.7603 | 0.6549 | | MOI | 1.36 | 2.45 | 3.2 | 3.57 | 4.42 |
|
Table 7. Performance of different algorithms under three data sets