Scholarlink

<p>With the increasing density of maritime traffic and the growing complexity of marine environments, accurate ship recognition plays a critical role in ensuring navigational safety. Traditional ship recognition methods often face challenges such as insufficient accuracy and poor performance in complex environments. To address these issues, we propose a novel Hierarchical Attention Compression Network (HACRNet). This method introduces a Spatial-Reduction Multi-Head Attention (SR-MSA) module that progressively reduces feature resolution, significantly lowering computational complexity and enhancing recognition efficiency. Additionally, the Cross-Layer Attention Selection Module (CLASM) dynamically integrates multi-level attention, focusing on the most discriminative regions, thus improving the model’s robustness in complex maritime conditions. To further enhance sensitivity to subtle differences between ship categories, we also design an optimized contrastive loss function. Experimental results demonstrate that, compared to state-of-the-art methods, HACRNet incurs only a 2% decrease in accuracy, while the computational cost (FLOPs) is reduced by 83%, and the frame rate (FPS) is improved by two-to-three times. This method strikes an effective balance between high-speed performance, efficiency, and accuracy. The code is publicly available online at <a href="https://github.com/runtianw/HACR.">https://github.com/runtianw/HACR.</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

HACRNet: hierarchical attention compression for high-speed fine-grained ship recognition

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Runtian Wang,
  • Renjie Qiao,
  • Wentao Zhou,
  • Kejun Wu,
  • Chengtao Cai

摘要

With the increasing density of maritime traffic and the growing complexity of marine environments, accurate ship recognition plays a critical role in ensuring navigational safety. Traditional ship recognition methods often face challenges such as insufficient accuracy and poor performance in complex environments. To address these issues, we propose a novel Hierarchical Attention Compression Network (HACRNet). This method introduces a Spatial-Reduction Multi-Head Attention (SR-MSA) module that progressively reduces feature resolution, significantly lowering computational complexity and enhancing recognition efficiency. Additionally, the Cross-Layer Attention Selection Module (CLASM) dynamically integrates multi-level attention, focusing on the most discriminative regions, thus improving the model’s robustness in complex maritime conditions. To further enhance sensitivity to subtle differences between ship categories, we also design an optimized contrastive loss function. Experimental results demonstrate that, compared to state-of-the-art methods, HACRNet incurs only a 2% decrease in accuracy, while the computational cost (FLOPs) is reduced by 83%, and the frame rate (FPS) is improved by two-to-three times. This method strikes an effective balance between high-speed performance, efficiency, and accuracy. The code is publicly available online at https://github.com/runtianw/HACR.