Deep learning denoising enables rapid SEM imaging under charging conditions for FE SEM, CD SEM, and review SEM.

In scanning electron microscopy (SEM), rapid image acquisition is essential to prevent charging on non-conductive samples. However, fast acquisition often yields noisy images due to insufficient electron signal, whereas longer exposures improve quality but increase charging risk, causing distortion or signal loss. Conventional noise reduction via multi-frame averaging is effective but often impractical under low-voltage, high-speed conditions. This study evaluates four transformer-based denoising models–Restormer, NAFNet, HINet, and CGNet–applied to SEM images acquired under charging-sensitive conditions. All models were trained on an in-house paired dataset with 2-frame inputs and 32-frame references. We report full-reference metrics (PSNR/SSIM) and perceptual metrics (LPIPS/DISTS), include a small blinded MOS assessment, and use paired per-FOV statistics with 95% confidence intervals to confirm significant gains over the 2-frame baseline. While validation results were promising, performance decreased on a held-out test set of newly acquired SEM data, underscoring the need for broader data and model generalization. Among the evaluated models, NAFNet achieved the best overall fidelity and latency–approaching the 16-frame quality band with an ∼66× speed-up over 32-frame acquisition–whereas Restormer preserved circular geometry most faithfully under distribution shift; HINet and CGNet showed more visible geometric distortions. These findings clarify the practical speed–quality benefit of two-frame + AI denoising and highlight the importance of visual/perceptual evaluation alongside conventional metrics for robust deployment in FE-SEM, CD-SEM and Review-SEM.