Figure 1.
Replace the original Resblock with the RRDB structure
Figure 2.
RRDB structure
Figure 3.
Representative feature maps before and after activation
Figure 4.
The architecture diagram of ViT
Figure 5.
Comparison of CNN's local receptive field (left) and ViT's global attention weight distribution (right).
Figure 6.
The improved generator network structure
TABLE I.
PARAMETER CONFIGURATION OF THE VIT-BASE MODULE
| Module Component | Key Parameters | Main Function |
|---|---|---|
| Patch Embedding | Input Size:64×64×64 | Divide the feature map into 64 patches to reduce computational complexity. |
| Patch Size:8×8 | ||
| Output Dimension:768 | ||
| Multi-Head Attention | Number of Heads:12 | Establish global correlations to enhance complex feature modeling. |
| Dimension per Head:64 | ||
| Feed-Forward Network | MLP Structure:768→3072→768 | Perform nonlinear transformations of features to stabilize training in conjunction with LayerNorm. |
| Activation Function:GeLU | ||
| Residual Connection | Application Position:After each MSA/FFN sublayer | Prevent gradient vanishing and accelerate convergence. |
| Stacked Structure | Number of Encoder Layers:12 | Construct a deep feature transformer to improve global context modeling capabilities. |
| Total Parameters:86M |
Figure 7.
The improved generator network structure
TABLE II.
COMPARISON OF PSNR (DB) / SSIM FOR DIFFERENT METHODS ON BENCHMARK DATASETS UNDER ×2 AND ×4 SUPER-RESOLUTION
| Multiple | Model | Set5 PSNR (dB)/SSIM | Set14 PSNR (dB)/SSIM | BSD100 PSNR (dB)/SSIM | Urban100 PSNR (dB)/SSIM |
|---|---|---|---|---|---|
| ×2 | SRCNN | 36.66/0.9542 | 32.42/0.9063 | 31.36/0.8918 | 29.50/0.8946 |
| EDSR | 38.11/0.9603 | 33.92/0.9180 | 32.46/0.9015 | 32.93/0.9355 | |
| RCAN | 38.31/0.9614 | 34.15/0.9209 | 32.63/0.9027 | 33.34/0.9410 | |
| HAN | 38.34/0.9618 | 34.18/0.9214 | 32.68/0.9032 | 33.41/0.9422 | |
| NLSA | 38.35 / 0.9620 | 34.20 / 0.9217 | 32.69 / 0.9036 | 33.48 / 0.9430 | |
| SwinIR | 38.40 / 0.9624 | 34.25 / 0.9221 | 32.71 / 0.9040 | 33.57 / 0.9438 | |
| Our | 38.51 / 0.963 | 34.40 / 0.9235 | 32.85 / 0.9054 | 33.95 / 0.9465 | |
| ×4 | SRCNN | 30.49 / 0.8630 | 27.50 / 0.7510 | 26.91 / 0.7105 | 24.54 / 0.7260 |
| EDSR | 32.65 / 0.9000 | 28.95 / 0.7918 | 27.80 / 0.7449 | 26.98 / 0.8085 | |
| RCAN | 32.78 / 0.9004 | 29.06 / 0.7932 | 27.89 / 0.7468 | 27.14 / 0.8125 | |
| HAN | 32.80 / 0.9008 | 29.08 / 0.7939 | 27.91 / 0.7473 | 27.24 / 0.8140 | |
| NLSA | 32.82 / 0.9010 | 29.10 / 0.7942 | 27.93 / 0.7480 | 27.31 / 0.8153 | |
| SwinIR | 32.88 / 0.9014 | 29.14 / 0.7948 | 27.96 / 0.7488 | 27.45 / 0.8165 | |
| Our | 33.00 / 0.9025 | 29.28 / 0.7962 | 28.10 / 0.7505 | 27.73 / 0.8200 |
Figure 8.
Comparison of 4x upsampling experimental results for img062.png in the Urban100 dataset.