Scholarlink

<p>Accurate segmentation in porous media is critical for analyzing multiphase fluid interactions and predicting flow behavior. However, the lack of high-quality, balanced paired training data poses a challenge for training robust deep learning models. In this paper, a novel approach that integrates neural networks with Denoising Diffusion Implicit Models (DDIMs) is presented to address class imbalance and improve semantic segmentation in digital rock analysis. Our method employs a sequential framework where an unconditional DDIM generates segmentation masks enriched with minority classes, and these masks serve as conditioning inputs for a conditional DDIM, which generates corresponding micro-CT images. These synthetic paired datasets, augmented by diffusion models, are used to train deep neural networks for multi-class segmentation of rock matrix and fluid phases. The neural network architecture, based on a U-Net framework, is trained with both real and synthetic data to significantly improve segmentation accuracy, particularly for underrepresented classes. Results demonstrate that integrating DDIM-augmented data reduces overfitting, enhances dataset diversity, and improves computational efficiency. This study contributes to advancing the application of neural networks and generative models in addressing data imbalance for semantic segmentation tasks, providing a scalable solution for complex geological datasets.</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

Mitigating data imbalance in semantic segmentation using sequential unconditional and conditional diffusion models: a case study in digital rock physics

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Alireza Kazemi,
  • Mohammad Esmaeili

摘要

Accurate segmentation in porous media is critical for analyzing multiphase fluid interactions and predicting flow behavior. However, the lack of high-quality, balanced paired training data poses a challenge for training robust deep learning models. In this paper, a novel approach that integrates neural networks with Denoising Diffusion Implicit Models (DDIMs) is presented to address class imbalance and improve semantic segmentation in digital rock analysis. Our method employs a sequential framework where an unconditional DDIM generates segmentation masks enriched with minority classes, and these masks serve as conditioning inputs for a conditional DDIM, which generates corresponding micro-CT images. These synthetic paired datasets, augmented by diffusion models, are used to train deep neural networks for multi-class segmentation of rock matrix and fluid phases. The neural network architecture, based on a U-Net framework, is trained with both real and synthetic data to significantly improve segmentation accuracy, particularly for underrepresented classes. Results demonstrate that integrating DDIM-augmented data reduces overfitting, enhances dataset diversity, and improves computational efficiency. This study contributes to advancing the application of neural networks and generative models in addressing data imbalance for semantic segmentation tasks, providing a scalable solution for complex geological datasets.