Knowledge-distilled swin transformer for efficient vibration artifact removal in SEM imaging

This study addresses the challenge of removing vibration-induced horizontal stripe artifacts in Scanning Electron Microscopy (SEM) images, which degrade image fidelity and compromise quantitative analysis. The main contribution lies in the development of SwinIR-KD, a computationally efficient image restoration framework. This framework synergistically integrates the Swin Transformer architecture with knowledge distillation and a novel Horizontal Stripe Suppress Loss. Specifically, SwinIR-KD employs a lightweight student model, distilled from a larger pre-trained SwinIR teacher model, and incorporates architectural modifications to reduce complexity. A specialized loss function, combining reconstruction, distillation, and stripe suppression components, guides the training. Experimental results on an SEM speckle image dataset demonstrate that SwinIR-KD significantly reduces model parameters and computational complexity by approximately 79% compared to the baseline SwinIR, while achieving comparable or even superior image restoration performance in terms of PSNR, SSIM, FID, and LPIPS. Furthermore, SwinIR-KD effectively processes large-scale SEM images and is complemented by a user-friendly interface for practical application.