Scholarlink

<p>Automated classification of metamorphic rocks is crucial for geological surveys yet faces dual challenges: the scarcity of labeled thin-section samples and high intraclass heterogeneity caused by complex mineral assemblages. Traditional methods, such as transfer learning, often fail to generalize effectively as they struggle to capture the fine-grained, high-frequency petrographic textures inherent in metamorphic rocks. To address these limitations, this study repositions the task from direct identification to a <b>Generative Data Augmentation</b> strategy. We introduce a novel <b>one</b>-step <b>diffusion</b> model in <b>latent</b> space guided by <b>energy</b> distribution (<b>ODLE</b>). Unlike traditional GANs, ODLE incorporates an energy-guided mechanism to ensure that synthesized images preserve geologically meaningful features, such as mineral boundaries and structural relationships. Using a Variational Autoencoder (VAE), rock thin-section images are compressed into a latent space and reconstructed via diffusion. This process creates a diverse, realistic dataset that fills gaps in the original distribution. Our model leverages a pre-trained diffusion noise prediction model to generate high-quality images efficiently, overcoming data scarcity challenges. Experimental results demonstrate that this approach significantly improves downstream lithological classification accuracy and generates high-fidelity synthetic samples with potential applications in petrographic education and automated texture 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

Generative data augmentation for metamorphic rock thin-section classification based on one-step diffusion model

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Xingpeng Zhang,
  • Jing Xu,
  • Han Zhao,
  • Qiuli Wang,
  • Dian Qi,
  • Yan Chen,
  • Yang Yu,
  • Bin Xiao,
  • Bing Wang

摘要

Automated classification of metamorphic rocks is crucial for geological surveys yet faces dual challenges: the scarcity of labeled thin-section samples and high intraclass heterogeneity caused by complex mineral assemblages. Traditional methods, such as transfer learning, often fail to generalize effectively as they struggle to capture the fine-grained, high-frequency petrographic textures inherent in metamorphic rocks. To address these limitations, this study repositions the task from direct identification to a Generative Data Augmentation strategy. We introduce a novel one-step diffusion model in latent space guided by energy distribution (ODLE). Unlike traditional GANs, ODLE incorporates an energy-guided mechanism to ensure that synthesized images preserve geologically meaningful features, such as mineral boundaries and structural relationships. Using a Variational Autoencoder (VAE), rock thin-section images are compressed into a latent space and reconstructed via diffusion. This process creates a diverse, realistic dataset that fills gaps in the original distribution. Our model leverages a pre-trained diffusion noise prediction model to generate high-quality images efficiently, overcoming data scarcity challenges. Experimental results demonstrate that this approach significantly improves downstream lithological classification accuracy and generates high-fidelity synthetic samples with potential applications in petrographic education and automated texture analysis.