Scholarlink

<p>Traditional rock photomicrograph lithology identification methods are inefficient and subjective, and existing deep - learning models face accuracy bottlenecks. To address these issues, this study built a diverse rock photomicrograph image dataset. Then, an improved network model FSDS (Fasternet with DualSE Module and ScConv) based on Fasternet was proposed for efficient and accurate lithology identification. FSDS innovatively incorporates the DualSE module, strengthening the model’s perception of key features through spatial and channel attention. The ScConv module is also embedded to reduce feature redundancy without heavy computational load. The model structure is adjusted for better resource utilization. Tests on the self - built dataset show that FSDS has a reasonable number of parameters. Its average recognition accuracy reaches 83%, 5% higher than Fasternet and similar models, with only 2.6&#xa0;M floating - point operations. This proves FSDS’s excellent performance and offers strong technical support for geological rock 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

Research on intelligent classification of limestone photomicrographs based on the improved FasterNet architecture

  • 期刊论文

  • 发表日期：

    2025年10月21日

查看全文
  • Tingting Wang,
  • Yuanchun Niu,
  • Wanchun Zhao,
  • Ranjith Pathegama Gamage,
  • Isha Ahmad

摘要

Traditional rock photomicrograph lithology identification methods are inefficient and subjective, and existing deep - learning models face accuracy bottlenecks. To address these issues, this study built a diverse rock photomicrograph image dataset. Then, an improved network model FSDS (Fasternet with DualSE Module and ScConv) based on Fasternet was proposed for efficient and accurate lithology identification. FSDS innovatively incorporates the DualSE module, strengthening the model’s perception of key features through spatial and channel attention. The ScConv module is also embedded to reduce feature redundancy without heavy computational load. The model structure is adjusted for better resource utilization. Tests on the self - built dataset show that FSDS has a reasonable number of parameters. Its average recognition accuracy reaches 83%, 5% higher than Fasternet and similar models, with only 2.6 M floating - point operations. This proves FSDS’s excellent performance and offers strong technical support for geological rock analysis.