Scholarlink

<p>This work presents a deep learning-based pipeline for reliable quantitative analysis of Scanning Transmission Electron Microscopy (STEM) images acquired under practical low-dose conditions, in which shot noise and scan distortion can degrade the accuracy of subsequent measurements. The proposed workflow integrates four tasks in a unified analysis sequence: denoising, atomic position localization, atomic classification, and segmentation. Supervised models are trained on the physics-based TEMImageNet dataset generated by forward modeling. We employ UNet3+ for image denoising and for predicting a Gaussian map that encodes atomic center likelihoods. Atomic center positions are extracted from either the denoised image or the corresponding map, and the atoms are then grouped using Density-Based Spatial Clustering of Applications with Noise (DBSCAN). For instance segmentation, we introduce a two-stage strategy in which the predicted Gaussian map provides point prompts to Segment Anything Model&#xa0;2 (SAM2), and the Gaussian map input is used to improve the separation of closely spaced atoms. Experiments show that the denoising model improves image fidelity while preserving atomic structure information, and that the Gaussian map, rather than the denoised image, leads to more accurate separation of adjacent atoms in segmentation. Overall, the proposed pipeline is robust to low-dose noise and acquisition-related artifacts, and produces baseline outputs that can support atomic-scale deformation mapping, defect quantification, and microstructure analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Ai-assisted atomic-scale analysis in STEM images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Seunghwan Lee,
  • Taegyeom Lee,
  • Minsung Kim,
  • In Hye Kwak,
  • Tae Yeoung Lee,
  • Jae Hyuck Jang,
  • Hee-Suk Chung,
  • Sang-Chul Lee

摘要

This work presents a deep learning-based pipeline for reliable quantitative analysis of Scanning Transmission Electron Microscopy (STEM) images acquired under practical low-dose conditions, in which shot noise and scan distortion can degrade the accuracy of subsequent measurements. The proposed workflow integrates four tasks in a unified analysis sequence: denoising, atomic position localization, atomic classification, and segmentation. Supervised models are trained on the physics-based TEMImageNet dataset generated by forward modeling. We employ UNet3+ for image denoising and for predicting a Gaussian map that encodes atomic center likelihoods. Atomic center positions are extracted from either the denoised image or the corresponding map, and the atoms are then grouped using Density-Based Spatial Clustering of Applications with Noise (DBSCAN). For instance segmentation, we introduce a two-stage strategy in which the predicted Gaussian map provides point prompts to Segment Anything Model 2 (SAM2), and the Gaussian map input is used to improve the separation of closely spaced atoms. Experiments show that the denoising model improves image fidelity while preserving atomic structure information, and that the Gaussian map, rather than the denoised image, leads to more accurate separation of adjacent atoms in segmentation. Overall, the proposed pipeline is robust to low-dose noise and acquisition-related artifacts, and produces baseline outputs that can support atomic-scale deformation mapping, defect quantification, and microstructure analysis.