Scholarlink

<p>Scanning Electron Microscopy (SEM) is indispensable for characterizing perovskite thin-film microstructure. However, current analysis remains largely manual, limiting throughput and consistency. Here, we present PerovSegNet, an automated deep learning framework for SEM image segmentation that enables accurate identification of lead iodide, perovskite, and defect domains. Built upon an improved YOLOv8x architecture, our model incorporates two novel modules: the Adaptive Shuffle Dilated Convolution Block, which enhances multi-scale feature extraction, and the Separable Adaptive Downsampling module, which preserves fine-scale textures and large-scale structures for robust boundary recognition. PerovSegNet achieves a mean Average Precision of 87.25%, outperforming the baseline by 4.08% while reducing computational cost by 25.22%. Comprehensive evaluations across heterogeneous cross-laboratory datasets and magnification levels demonstrate significant improvements in robustness and scale invariance. Furthermore, digital stress tests reveal enhanced stability under image degradation. Preliminary zero-shot tests on lead-free perovskites, non-perovskite thin films, and cross-sectional imagery indicate strong feature transferability. Beyond segmentation, PerovSegNet provides quantitative grain-level metrics as reliable indicators of microstructural quality, establishing a scalable, data-driven tool for real-time process monitoring and optimization of perovskite fabrication. The source code is available at: <a href="https://github.com/wlyyj/PerovSegNet/tree/master">https://github.com/wlyyj/PerovSegNet/tree/master</a>.</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

Automated and Scalable SEM Image Analysis of Perovskite Solar Cell Materials via a Deep Segmentation Framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Lin Wang,
  • Tianxiang Hu,
  • Jianguo Pan,
  • Hao Zhang,
  • Yiqiang Zhan,
  • Xia Cai

摘要

Scanning Electron Microscopy (SEM) is indispensable for characterizing perovskite thin-film microstructure. However, current analysis remains largely manual, limiting throughput and consistency. Here, we present PerovSegNet, an automated deep learning framework for SEM image segmentation that enables accurate identification of lead iodide, perovskite, and defect domains. Built upon an improved YOLOv8x architecture, our model incorporates two novel modules: the Adaptive Shuffle Dilated Convolution Block, which enhances multi-scale feature extraction, and the Separable Adaptive Downsampling module, which preserves fine-scale textures and large-scale structures for robust boundary recognition. PerovSegNet achieves a mean Average Precision of 87.25%, outperforming the baseline by 4.08% while reducing computational cost by 25.22%. Comprehensive evaluations across heterogeneous cross-laboratory datasets and magnification levels demonstrate significant improvements in robustness and scale invariance. Furthermore, digital stress tests reveal enhanced stability under image degradation. Preliminary zero-shot tests on lead-free perovskites, non-perovskite thin films, and cross-sectional imagery indicate strong feature transferability. Beyond segmentation, PerovSegNet provides quantitative grain-level metrics as reliable indicators of microstructural quality, establishing a scalable, data-driven tool for real-time process monitoring and optimization of perovskite fabrication. The source code is available at: https://github.com/wlyyj/PerovSegNet/tree/master.