A generic deep-learning-based defect segmentation model for electron micrographs for automatic defect inspection

Abstract

Defect inspection is an important part in the semiconductor manufacturing. This task is tedious and time consuming if done manually. Therefore, reliably automating this task is a major challenge for many semiconductor manufacturers. In the recent years, deep-learning methods for object detection have demonstrated ever better performances. However, most of the publicly available models are trained on natural images and objects. Hence, most of them needs a long and data greedy training step to be used on industrial Transmission Electron Microscopes or Scanning Electron Microscope images. In this context, we propose a deep-learning based model to detect and segment defects in electron micrographs. Using SmartDef3 from Pollen Metrology, we annotated defects on images from several industrial applications. We split them in a training and validation dataset, with which an Instance segmentation model with state-of-the-art backbone is trained. The model is then evaluated on different use cases. Competitive performances on new data in terms of detection rates and segmentation quality are demonstrated and discussed. Furthermore, the model showed a relevant defect detection rate even on images that are not in the semiconductor domain, providing an interesting tool for defects detection on a new use case without new training data. This shows how deep learning strategies can help save time and costs by automation of defects inspections. Furthermore, advanced metrological analysis of the defect can be simultaneously obtained that help optimizing the manufacturing processes and reduce defect production rate.