Enhancing scanning electron microscopy imaging quality of weakly conductive samples through unsupervised learning

Area of Science:

• Materials Science
• Image Processing
• Artificial Intelligence

Background:

• Scanning electron microscopy (SEM) requires highly conductive samples for optimal imaging.
• Weakly conductive materials yield compromised SEM image quality, hindering accurate structural analysis.
• Supervised image enhancement methods are ineffective due to the lack of paired low- and high-quality SEM images for such samples.

Purpose of the Study:

• To develop an unsupervised method for enhancing SEM images of weakly conductive materials.
• To improve the accuracy of structure-related analyses in materials science.
• To overcome limitations of traditional and supervised image processing techniques in SEM analysis.

Main Methods:

• Utilized a Cycle-consistent Generative Adversarial Network (CycleGAN) for unsupervised image enhancement.
• Employed unpaired blurred (weakly conductive) and clear (well-conductive) SEM images for end-to-end training.
• Integrated an edge loss function to preserve and recover finer material details in enhanced images.

Main Results:

• The proposed unsupervised CycleGAN method effectively enhanced SEM image quality for weakly conductive samples.
• Quantitative evaluations demonstrated superior performance compared to traditional image enhancement methods.
• The method successfully recovered finer details crucial for material structure analysis.

Conclusions:

• The developed framework broadens the application of artificial intelligence in materials analysis.
• This approach offers significant implications for materials science research and advanced image restoration.
• Enables high-quality SEM imaging and analysis for previously challenging weakly conductive materials.