Scholarlink

<p>In scenarios such as field geological surveys, mine explorations, and engineering construction projects where the deployment of large-scale models is inconvenient, the identification of rocks outdoors poses significant challenges. This is mainly attributed to the complexity of rock features, including color, texture, and size, which vary across different specimens. In recent years, the advancements in deep learning have opened up new possibilities for real-time rock type identification. This technology enables engineers, geologists, and construction workers to make rapid and informed decisions. In this study, a dataset was constructed by capturing images of boulders. To enrich the dataset, techniques such as image segmentation and enhancement were employed, effectively increasing the size and diversity of the dataset. Based on this dataset, a novel rock feature extraction module, namely the Simple Rock Feature Extraction (SRFE) module, was developed. This module was then integrated into a ResneXt architecture with structural improvements. By optimizing the computational structure and fine-tuning hyperparameters, the proposed model was able to accurately identify and classify both pristine rocks and those affected by lens contamination. The experimental results demonstrated remarkable performance, achieving an overall classification accuracy of 85%. Notably, in certain rock categories, the classification accuracy reached up to 93%. When compared with large-scale neural networks, the proposed model exhibited superior performance in terms of precision, recall, and F1-score. Moreover, it retained several advantages, including a compact model size, high computational speed, and low complexity, all while maintaining high levels of accuracy. In conclusion, the model proposed in this study successfully strikes a balance between accuracy and efficiency. Despite a reduction in the number of parameters compared to large-scale networks, it still maintains a high level of recognition performance. This research provides a reliable model and effective methodology for real-time lithology identification on mobile devices and offers valuable insights for future improvements in network models.</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

Investigation into the feature extraction module and lightweight network for rock image analysis

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Linrui Yang,
  • Yuguo Kang,
  • Qin Zhou,
  • Yifan Xiao,
  • Chenxi Gao,
  • Xue Ling

摘要

In scenarios such as field geological surveys, mine explorations, and engineering construction projects where the deployment of large-scale models is inconvenient, the identification of rocks outdoors poses significant challenges. This is mainly attributed to the complexity of rock features, including color, texture, and size, which vary across different specimens. In recent years, the advancements in deep learning have opened up new possibilities for real-time rock type identification. This technology enables engineers, geologists, and construction workers to make rapid and informed decisions. In this study, a dataset was constructed by capturing images of boulders. To enrich the dataset, techniques such as image segmentation and enhancement were employed, effectively increasing the size and diversity of the dataset. Based on this dataset, a novel rock feature extraction module, namely the Simple Rock Feature Extraction (SRFE) module, was developed. This module was then integrated into a ResneXt architecture with structural improvements. By optimizing the computational structure and fine-tuning hyperparameters, the proposed model was able to accurately identify and classify both pristine rocks and those affected by lens contamination. The experimental results demonstrated remarkable performance, achieving an overall classification accuracy of 85%. Notably, in certain rock categories, the classification accuracy reached up to 93%. When compared with large-scale neural networks, the proposed model exhibited superior performance in terms of precision, recall, and F1-score. Moreover, it retained several advantages, including a compact model size, high computational speed, and low complexity, all while maintaining high levels of accuracy. In conclusion, the model proposed in this study successfully strikes a balance between accuracy and efficiency. Despite a reduction in the number of parameters compared to large-scale networks, it still maintains a high level of recognition performance. This research provides a reliable model and effective methodology for real-time lithology identification on mobile devices and offers valuable insights for future improvements in network models.