Defocus-to-focus SEM Restoration on Real defocused SEM images S→R protocol (Img1)
13.6CDViT-MAE (ImgNet)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ViT-MAE (ImgNet)Pre-training Dataset=ImageNet, Architecture=ViT-MAE2026.04 | 13.6 | 1.1 | 6 | 4.9 | 6.5 | |
| ViT-MAE + MoE (ℓ1)Architecture=ViT-MAE + MoE, Loss=ℓ12026.04 | 16.1 | 1.5 | 5.4 | 3.9 | 5.7 | |
| Focused InputInput Type=Reference2026.04 | 16.3 | 0.4 | 4.2 | 3.3 | 4.2 | |
| ViT-MAE + MoE (ℓ1+TV)Architecture=ViT-MAE + MoE, Loss=ℓ1 + TV2026.04 | 16.7 | 1.7 | 4.1 | 3 | 4.5 | |
| Defocused InputInput Type=Noisy/Defocused2026.04 | 16.9 | 1.4 | 6.4 | 4.6 | 6.5 | |
| ViT-MAE (SEM)Pre-training Dataset=SEM, Architecture=ViT-MAE2026.04 | 17 | 1.6 | 5.7 | 4 | 5.8 | |
| BM3DMethod Type=Denoising2026.04 | 17.1 | 1.6 | 4.8 | 3.4 | 5.1 | |
| Richardson–LucyMethod Type=Classical Deconvolution2026.04 | 17.4 | 1.4 | 4.3 | 3 | 4.4 | |
| Noise2Noise (N2N)Architecture=UNet2026.04 | 17.8 | 1.6 | 4.4 | 3.1 | 4.7 | |
| Noise2Void (N2V)Architecture=UNet2026.04 | 17.8 | 1.6 | 4.4 | 3 | 4.6 | |
| Wiener FilterMethod Type=Classical Deconvolution2026.04 | 18 | 0.4 | 9.3 | 7.2 | 9.4 | |
| MRNArchitecture=UNet2026.04 | 19.7 | 1.4 | 4 | 2.9 | 4.1 |