Fig. 1. Generator network structure
Fig. 2. Discriminator network structure
Fig. 3. Use of channel and spatial attention modules
Fig. 4. Reconstruction effects with different values of ε coefficient
Fig. 5. Comparison of residual blocks. (a)SRGAN; (b) proposed model
Fig. 6. Variation curve of generator function loss value
Fig. 7. Variation curve of discriminant function loss value
Fig. 8. Partial enlarged comparison diagrams of the “baby” reconstruction effect of five algorithms in Set5 test set
Fig. 9. Partial enlarged comparison diagrams of the “butterfly” reconstruction effect of five algorithms in Set5 test set
Fig. 10. Partial enlarged comparison diagrams of the “pepper” reconstruction effect of five algorithms in Set14 test set
Fig. 11. Partial enlarged comparison diagrams of the “fish” reconstruction effect of five algorithms in BSDS100 test set
Fig. 12. Partial enlarged comparison diagrams of the “room” reconstruction effect of five algorithms in Urban100 test set
Fig. 13. Partial enlarged comparison diagrams of the “baby” reconstruction effect in ablation experiment in Set5 test set
Fig. 14. Partial enlarged comparison diagrams of the “butterfly” reconstruction effect in ablation experiment in Set5 test set
Fig. 15. Partial enlarged comparison diagrams of the “lenna” reconstruction effect in ablation experiment in Set14 test set
Test set | Scale | Bicubic | ESPCN | SRGAN | ESRGAN | Proposed |
---|
Set5 | 4× | 26.692 | 27.594 | 26.628 | 28.543 | 29.510 | Set14 | 4× | 24.565 | 25.186 | 24.568 | 24.532 | 26.443 | Urban100 | 4× | 21.706 | 22.300 | 22.113 | 22.792 | 23.917 | BSDS100 | 4× | 24.641 | 25.043 | 24.472 | 25.322 | 25.884 |
|
Table 1. Comparison of PSNR values of various super-resolution reconstruction methods
Test set | Scale | Bicubic | ESPCN | SRGAN | ESRGAN | Proposed |
---|
Set5 | 4× | 0.7730 | 0.7895 | 0.8023 | 0.8145 | 0.8517 | Set14 | 4× | 0.6732 | 0.6983 | 0.7019 | 0.6711 | 0.7377 | Urban100 | 4× | 0.6317 | 0.6595 | 0.6774 | 0.7050 | 0.7415 | BSDS100 | 4× | 0.6401 | 0.6702 | 0.6713 | 0.6514 | 0.7002 |
|
Table 2. Comparison of SSIM values of various super-resolution reconstruction methods
Method | Set5 | Set14 | Urban100 | BSDS100 |
---|
SRGAN | 26.628 | 24.568 | 22.113 | 24.472 | SRGAN-BN | 27.863 | 25.269 | 22.743 | 25.228 | SRGAN+CA&SA | 27.865 | 25.375 | 22.693 | 25.198 | SRGAN+Charbonnier | 28.153 | 25.724 | 22.919 | 25.404 | SRGAN+Charbonnier+CA&SA-BN | 29.510 | 26.443 | 23.917 | 25.884 |
|
Table 3. PSNR values of models with different module combinations on four test sets
Method | Set5 | Set14 | Urban | BSDS100 |
---|
SRGAN | 0.8023 | 0.7019 | 0.6774 | 0.6713 | SRGAN-BN | 0.8058 | 0.7079 | 0.6820 | 0.6728 | SRGAN+CA&SA | 0.8057 | 0.7068 | 0.6792 | 0.6754 | SRGAN+Charbonnier | 0.8203 | 0.7180 | 0.6939 | 0.6822 | SRGAN+Charbonnier+A&SA-BN | 0.8517 | 0.7377 | 0.7415 | 0.7002 |
|
Table 4. SSIM values of models with different module combinations on four test sets