Scholarlink

<p>Real-time estimation of spatio-temporal ocean wave fields is crucial for marine applications but challenged by the limitations of traditional sensors and the computational cost of dense vision-based reconstruction. Existing methods often rely heavily on sparse geometric data, potentially underutilizing rich image information. This paper proposed an efficient deep learning network specifically designed for real-time, sparse-to-dense wave field estimation guided by image features, through the network of Progressive Residual Incremental Multi-scale Estimation (PRIME-Net). PRIME-Net employs an encoder–decoder architecture integrating efficient modules like a Pyramid Pooling Attention Module (PPAM) and modulated decoder blocks (MDBs), along with a multi-scale residual prediction strategy to progressively refine elevation estimates from an input image and an initial sparse or low-quality height map. Experiments conducted on challenging datasets demonstrate that PRIME-Net achieves competitive or superior accuracy in reconstructing wave fields compared to state-of-the-art methods, while exhibiting significant computational efficiency suitable for real-time operation. The results validate the effectiveness of the image-guided approach and the architectural design, positioning PRIME-Net as a promising solution for the awareness of sea state.</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

PRIME-Net: an efficient progressive residual incremental multi-scale estimation network for dynamic ocean wave fields

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Feng Wang,
  • Renjie Qiao,
  • Xiaoyu Wang,
  • Haiyang Meng

摘要

Real-time estimation of spatio-temporal ocean wave fields is crucial for marine applications but challenged by the limitations of traditional sensors and the computational cost of dense vision-based reconstruction. Existing methods often rely heavily on sparse geometric data, potentially underutilizing rich image information. This paper proposed an efficient deep learning network specifically designed for real-time, sparse-to-dense wave field estimation guided by image features, through the network of Progressive Residual Incremental Multi-scale Estimation (PRIME-Net). PRIME-Net employs an encoder–decoder architecture integrating efficient modules like a Pyramid Pooling Attention Module (PPAM) and modulated decoder blocks (MDBs), along with a multi-scale residual prediction strategy to progressively refine elevation estimates from an input image and an initial sparse or low-quality height map. Experiments conducted on challenging datasets demonstrate that PRIME-Net achieves competitive or superior accuracy in reconstructing wave fields compared to state-of-the-art methods, while exhibiting significant computational efficiency suitable for real-time operation. The results validate the effectiveness of the image-guided approach and the architectural design, positioning PRIME-Net as a promising solution for the awareness of sea state.