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

Area of Science:

• Materials Science
• Image Processing
• Microscopy

Background:

• Vibration-induced horizontal stripe artifacts degrade Scanning Electron Microscopy (SEM) image quality.
• Artifacts compromise the fidelity and quantitative analysis of SEM images.

Purpose of the Study:

• To develop a computationally efficient framework for removing horizontal stripe artifacts in SEM images.
• To improve the quality and reliability of SEM image analysis.

Main Methods:

• Development of SwinIR-KD, integrating Swin Transformer architecture with knowledge distillation.
• Implementation of a novel Horizontal Stripe Suppress Loss function.
• Utilizing a lightweight student model distilled from a pre-trained SwinIR teacher model.

Main Results:

• SwinIR-KD reduces model parameters and computational complexity by approximately 79% compared to baseline SwinIR.
• Achieved comparable or superior image restoration performance (PSNR, SSIM, FID, LPIPS).
• Demonstrated effective processing of large-scale SEM images.

Conclusions:

• SwinIR-KD offers an efficient and effective solution for SEM image artifact removal.
• The framework enhances the practical applicability of SEM imaging through a user-friendly interface.
• This approach preserves image fidelity for critical quantitative analysis.