Scholarlink

<p>With the advancement of marine science research, the study and monitoring of benthic organisms have gained prominence. These organisms, inhabiting the seafloor, play crucial roles in marine ecosystems. However, traditional methods for benthic organism surveys rely heavily on manual operations, which are labor-intensive and prone to errors. Moreover, the complexity of underwater environments poses significant challenges for automated detection algorithms. Current models face issues such as large parameter sizes, high computational demands, and difficulties in detecting small targets and overlapping objects. To address these challenges, we propose LUW-DETR, a lightweight algorithm based on the transformer architecture. LUW-DETR incorporates a high–low-frequency attention mechanism and dynamic group convolution module to significantly improve feature extraction and computational efficiency. Experimental results demonstrate that LUW-DETR achieves competitive performance with an mAP of 83.1% on the URPC 2020 dataset and 63.7% on the Pascal VOC 2007 dataset, while reducing parameters by 28.3% and computational cost by 29.6% compared to baseline models. These results highlight LUW-DETR’s balance between accuracy and efficiency, making it suitable for resource-constrained underwater environments. Our source code is available at:<a href="https://github.com/DingBC/LUWDETR">https://github.com/DingBC/LUWDETR</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

Lightweight LUW-DETR for efficient underwater benthic organism detection

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Ning Li,
  • Bocheng Ding,
  • Gaochao Yang,
  • Shiyun Ni,
  • Mingliang Wang

摘要

With the advancement of marine science research, the study and monitoring of benthic organisms have gained prominence. These organisms, inhabiting the seafloor, play crucial roles in marine ecosystems. However, traditional methods for benthic organism surveys rely heavily on manual operations, which are labor-intensive and prone to errors. Moreover, the complexity of underwater environments poses significant challenges for automated detection algorithms. Current models face issues such as large parameter sizes, high computational demands, and difficulties in detecting small targets and overlapping objects. To address these challenges, we propose LUW-DETR, a lightweight algorithm based on the transformer architecture. LUW-DETR incorporates a high–low-frequency attention mechanism and dynamic group convolution module to significantly improve feature extraction and computational efficiency. Experimental results demonstrate that LUW-DETR achieves competitive performance with an mAP of 83.1% on the URPC 2020 dataset and 63.7% on the Pascal VOC 2007 dataset, while reducing parameters by 28.3% and computational cost by 29.6% compared to baseline models. These results highlight LUW-DETR’s balance between accuracy and efficiency, making it suitable for resource-constrained underwater environments. Our source code is available at:https://github.com/DingBC/LUWDETR.