Implicit neural network-based coal SEM super-resolution for enhancing micro-pores measurement tasks

Area of Science:

• Materials Science
• Microscopy
• Computational Imaging

Background:

• Scanning Electron Microscopy (SEM) requires high resolution for detailed analysis.
• Prolonged radiation exposure during SEM can cause structural damage to specimens.
• Existing super-resolution (SR) methods may not adequately address radiation-induced artifacts or ensure measurement stability.

Purpose of the Study:

• To develop an interactive-interpretable super-resolution (SR) framework to mitigate radiation damage in SEM.
• To improve visual fidelity and geometric accuracy of high-resolution SEM images.
• To provide a stable and applicable solution for geometry-sensitive measurement tasks.

Main Methods:

• Integration of implicit neural representation (INR) with model-driven Half-Quadratic Splitting (HQS) optimization.
• Utilizing a local window attention mechanism within INR for contextual dependencies.
• Employing an interactive dual-branch network to decouple feature content and positional encoding.
• Unfolding the HQS algorithm into a deep, interpretable network for transparent optimization steps.

Main Results:

• The proposed SR framework significantly outperforms state-of-the-art SR algorithms in visual fidelity.
• Demonstrated applicability and stability in downstream geometry-sensitive measurement tasks.
• The method effectively addresses structural damage risks associated with prolonged radiation exposure in SEM.

Conclusions:

• The interactive-interpretable SR framework offers a robust solution for high-resolution SEM imaging under radiation.
• The integration of INR and HQS optimization enhances image quality while preserving geometric integrity.
• This approach is valuable for applications requiring precise measurements from radiation-exposed specimens.