Scholarlink

<p>Data-driven inversion methods are gaining popularity due to their remarkable ability to learn from data rather than relying on physical assumptions. Existing deep learning full waveform inversion methods, such as FCNVMB, demonstrate excellent performance by training a U-shaped network (U-Net). However, the original U-Net lacks the ability to capture correlations between local and global information. In this paper, we embed the Vision Transformer (ViT) with hybrid attention into U-Net to address this issue. First, we design a hybrid attention mechanism (HAM) that incorporates the existing ViT’s multi-head self-attention mechanism and a local windowed attention mechanism. It captures both global and local dependencies to improve the model’s prediction capability. Second, we design a new loss function by combining a customized Canny loss and the mean square error loss. Due to the close velocity values between some strata in data, many unclear boundaries in the prediction results can be mitigated. Third, we devise a transfer learning fine-tuning scheme to address the issue of insufficient training data of the SEGSalt dataset. This scheme requires only fine-tuning to transfer the exceptional performance of the pre-trained model to other datasets. Experiments are undertaken on OpenFWI and SEGSalt dataset. Results demonstrate that UH-Net outperforms two state-of-the-art data-driven methods. The code is available at <a href="https://github.com/fansmale/uhnet">https://github.com/fansmale/uhnet</a>.</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

UH-Net: U-Net with hybrid attention for full waveform inversion

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Zheng Zhang,
  • Ziyu Qin,
  • Liyan Liu,
  • Fan Min

摘要

Data-driven inversion methods are gaining popularity due to their remarkable ability to learn from data rather than relying on physical assumptions. Existing deep learning full waveform inversion methods, such as FCNVMB, demonstrate excellent performance by training a U-shaped network (U-Net). However, the original U-Net lacks the ability to capture correlations between local and global information. In this paper, we embed the Vision Transformer (ViT) with hybrid attention into U-Net to address this issue. First, we design a hybrid attention mechanism (HAM) that incorporates the existing ViT’s multi-head self-attention mechanism and a local windowed attention mechanism. It captures both global and local dependencies to improve the model’s prediction capability. Second, we design a new loss function by combining a customized Canny loss and the mean square error loss. Due to the close velocity values between some strata in data, many unclear boundaries in the prediction results can be mitigated. Third, we devise a transfer learning fine-tuning scheme to address the issue of insufficient training data of the SEGSalt dataset. This scheme requires only fine-tuning to transfer the exceptional performance of the pre-trained model to other datasets. Experiments are undertaken on OpenFWI and SEGSalt dataset. Results demonstrate that UH-Net outperforms two state-of-the-art data-driven methods. The code is available at https://github.com/fansmale/uhnet.