Boundary-Focused Semantic Segmentation for Limited Wafer Transmission Electron Microscope Images

Yongwon Jo, Jinsoo Bae, Hansam Cho, Heejoong Roh, Kyunghye Kim, Munki Jo, Jaeung Tae, Seoung Bum Kim

Abstract: In the semiconductor industry, automated measurement through wafer transmission electron microscopy (TEM) has gained significance because of the increasing nano-scale dimensions of contemporary wafers. The application of semantic segmentation deep learning models for automated measurement is used; however, their efficacy is degraded by challenges in acquiring sufficient wafer TEM images and delineating object boundaries from lots of noises generated by electron beams. In this study, we propose a transfer learning-based semantic segmentation framework to alleviate these challenges. By leveraging transfer learning, our model addresses data scarcity issues across diverse manufacturing processes. In addition, we use a loss function that allocates more weights to boundary regions to enhance boundary recognition accuracy. We demonstrated that our framework is more efficient than simple semantic segmentation models without transfer learning through experiments in various scenarios with limited TEM images.