Scholarlink

<p>Mineral image identification has been extensively studied recently, but the accuracy still needs to be increased. Although hardness has been added to increase the identification accuracy, more features like streak and luster have not been not used. Therefore, this paper integrates multiple modals to enhance mineral identification accuracy. However, practical applications often face challenges due to the limited availability of complete modals, most cases provide only mineral images. So, transferring knowledge from multimodal models to single-modal models using a generalized distillation strategy is presented in here. The effectiveness of the multimodal and the generalized distillation strategy were demonstrated on a mineral image dataset with 36 mineral types. The experimental results showed that the full-modal model was nearly 10% higher in Top-1 accuracy than single-modal model and ResNet50 after the distillation from the Swin Transformer with full modals had a 4% Top-1 accuracy increase.</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

Mineral Identification Based on Multimodal and Knowledge Distillation

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Hongxiang Liao,
  • Xiaohui Ji,
  • Mei Yang,
  • Mingyue He,
  • Guocheng Lv,
  • Min Liu,
  • Zhaochong Zhang,
  • Shan Zeng,
  • Jiaxue Li,
  • Yuzhu Wang

摘要

Mineral image identification has been extensively studied recently, but the accuracy still needs to be increased. Although hardness has been added to increase the identification accuracy, more features like streak and luster have not been not used. Therefore, this paper integrates multiple modals to enhance mineral identification accuracy. However, practical applications often face challenges due to the limited availability of complete modals, most cases provide only mineral images. So, transferring knowledge from multimodal models to single-modal models using a generalized distillation strategy is presented in here. The effectiveness of the multimodal and the generalized distillation strategy were demonstrated on a mineral image dataset with 36 mineral types. The experimental results showed that the full-modal model was nearly 10% higher in Top-1 accuracy than single-modal model and ResNet50 after the distillation from the Swin Transformer with full modals had a 4% Top-1 accuracy increase.