Scholarlink

<p>Rock thin section identification is a complex task, primarily constrained by the extraction of complex minerals and the acquisition of large-scale labeled data. This paper proposes a rock thin section identification method designed for few-shot labeled data, which enables the segmentation and identification of various rock minerals with minimal labeled data. The SAM model is used for mineral particle extraction, combined with the focal loss function, transfer learning, and the integration of multiple classification models to identify thin sections. The prediction process is evaluated at multiple levels. Ultimately, the method achieved the extraction and identification of 11 minerals using only 38 labeled data samples, with an identification accuracy of 91%. This approach significantly reduces the cost of manual labeling, requiring only a small amount of labeled data and minimal training effort to identify specific mineral classes. The source code of the proposed method is available at <a href="https://github.com/Xuerenbujianhua/SAMRocks">https://github.com/Xuerenbujianhua/SAMRocks</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

Few-Shot intelligent identification of rock thin sections based on SAM

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Yuan Zhou,
  • Qing Li,
  • Zhuofeng Zhang,
  • Zhengyu Wei,
  • Qiang Du,
  • Xinlong Li

摘要

Rock thin section identification is a complex task, primarily constrained by the extraction of complex minerals and the acquisition of large-scale labeled data. This paper proposes a rock thin section identification method designed for few-shot labeled data, which enables the segmentation and identification of various rock minerals with minimal labeled data. The SAM model is used for mineral particle extraction, combined with the focal loss function, transfer learning, and the integration of multiple classification models to identify thin sections. The prediction process is evaluated at multiple levels. Ultimately, the method achieved the extraction and identification of 11 minerals using only 38 labeled data samples, with an identification accuracy of 91%. This approach significantly reduces the cost of manual labeling, requiring only a small amount of labeled data and minimal training effort to identify specific mineral classes. The source code of the proposed method is available at https://github.com/Xuerenbujianhua/SAMRocks.