Robust Semantic Segmentation of Wafer Transmission Electron Microscopy Image Using Multi-Task Learning with Edge Detection

Semantic segmentation for wafer transmission electron microscopy (TEM) images plays a crucial role in the automated measurement of semiconductors. However, the automated measurement of wafer TEM images presents three main challenges: difficulty in image acquisition, significant noise, and ambiguous object boundaries. While existing methods for automated measurement have used semantic segmentation algorithms, they often lead to inaccurate object boundary detection, resulting in over- or under-estimation. In this study, we propose a multi-task pre-training with masked autoencoder and virtual edge detection (MTP-MAVED) to improve object recognition and boundary detection in wafer TEM images with consideration of three challenges. Our approach includes three primary components: a pre-training phase using self-supervised representation learning to extract meaningful representations from unlabeled wafer TEM images, a multi-task learning-based fine-tuning phase incorporating both semantic segmentation and edge detection tasks, and a boundary-aware loss function to enhance boundary recognition accuracy. We demonstrate that MTP-MAVED outperforms existing methods in both object recognition and boundary detection, even with limited labeled data. This framework offers a more robust solution for addressing the complexities of wafer TEM image analysis and advances the field of automated semiconductor analysis.