Scholarlink

<p>A novel data augmentation method is introduced for microscopic sandstone image generation, addressing the dual challenges of capturing fine-grained details and promoting texture diversity. This approach is built on three key innovations: temporal embedding refinement (TER), which ensures consistency across diffusion steps by refining the temporal embeddings in the generative process; multi-scale edge feature fusion (MEFF), a mechanism that harmonizes global structural coherence with the preservation of intricate local details; and multi-modal data-driven guidance (MDG), which integrates diverse contextual inputs, such as text prompts and categorical labels, to steer image generation. Extensive experiments on benchmark microscopic image datasets demonstrate that the proposed method outperforms conventional augmentation techniques, delivering superior image quality and diversity. Quantitative metrics and visual analysis confirm its effectiveness in enhancing data richness and improving the generalization performance of downstream models. Improved dataset quality directly supports more accurate geophysical analyses. These analyses include porosity estimation, grain structure classification, and predictive modeling in rock mechanics. This leads to more reliable subsurface characterization and reservoir evaluation. This approach offers a robust solution for advancing model accuracy in domain-specific applications, particularly in microscopic sandstone image 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

A Data Augmentation Technique for Microscopic Sandstone Image Generation Using Diffusion Transformer

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Fengcai Huo,
  • Hongjiang Li,
  • Hongli Dong,
  • Tianren Wang,
  • Weijian Ren

摘要

A novel data augmentation method is introduced for microscopic sandstone image generation, addressing the dual challenges of capturing fine-grained details and promoting texture diversity. This approach is built on three key innovations: temporal embedding refinement (TER), which ensures consistency across diffusion steps by refining the temporal embeddings in the generative process; multi-scale edge feature fusion (MEFF), a mechanism that harmonizes global structural coherence with the preservation of intricate local details; and multi-modal data-driven guidance (MDG), which integrates diverse contextual inputs, such as text prompts and categorical labels, to steer image generation. Extensive experiments on benchmark microscopic image datasets demonstrate that the proposed method outperforms conventional augmentation techniques, delivering superior image quality and diversity. Quantitative metrics and visual analysis confirm its effectiveness in enhancing data richness and improving the generalization performance of downstream models. Improved dataset quality directly supports more accurate geophysical analyses. These analyses include porosity estimation, grain structure classification, and predictive modeling in rock mechanics. This leads to more reliable subsurface characterization and reservoir evaluation. This approach offers a robust solution for advancing model accuracy in domain-specific applications, particularly in microscopic sandstone image analysis.