Defect-aware data augmentation for improving machine learning-based contour extraction from SEM images

Abstract

Background Contour extraction from scanning electron microscopy (SEM)images is crucial for various fields because of its ability to provide precise surface morphology analysis. Accurate contour extraction methods are essential for quality control, failure analysis, defect detection, and the development of advanced devices.

Aim However, existing machine learning-based SEM images contour extraction methods face the problem of difficulty in obtaining enough training data, particularly the lack of effective SEM image/contour pairs. This limitation often leads to defects in the extracted contours, such as open line-ends and discontinuous lines. Our aim is to address this problem.

Approach We propose a defect-aware data augmentation pipeline that uses conditional generative adversarial networks and BicycleGAN to solve these issues. We generate defective contours by artificially introducing defects into ideal contours, which then serve as input for the generative model to generate synthetic SEM images with low contrast or weak signal in the corresponding parts.

Results These generated SEM images, paired with the original ideal contours, are then used to retrain the contour extraction model, significantly improving the model’s ability to handle previously challenging cases.

Conclusions Our experimental results demonstrate that the proposed pipeline greatly enhances contour extraction performance, resolving previously open line-end defects. We highlight the potential of defect-aware data augmentation strategies and provide a practical framework for the future incorporation of more diverse defect types.