Scanning electron microscopy noise classification using deep learning

Noise plays a critical role in determining the performance and quantitative accuracy of imaging using a scanning electron microscope (SEM). In existing SEM literature, SEM noise is often assumed to follow Gaussian or Poisson statistics. However, in practice, true statistical behavior can vary depending on the operating conditions. In this work, we present a systematic analysis of SEM noise distributions using experimental images acquired under different accelerating voltages and dwell times with the in-lens and in-chamber secondary electron (SE) detectors, as well as the back-scattered electron (BSE) detector. Statistical tests, histogram analysis, and distribution fitting were employed to identify the underlying noise distributions. The results show that, at short dwell times, SE noise is best described by a Neyman Type-A distribution, whereas BSE noise is better described by a Gamma distribution. To automate the classification of these noise distributions, a deep learning framework based on convolutional neural networks (CNN) with a Convolutional Block Attention Module (CBAM) was implemented. Pretrained architectures were evaluated and compared with the proposed model. Among the pretrained models, MobileNetV2 achieved the highest accuracy of 88%, whereas the proposed CNN-CBAM model achieved the best overall performance, with F1-scores up to 0.98 and an accuracy of 97%.