Scholarlink

<p>Domain generalization poses significant challenges for target detection in aquaculture, as variations in illumination, water quality, and suspended particles induce substantial domain shifts, while the small size and strong camouflage of sea cucumbers further complicate detection. To address these challenges, we propose DG-YOLO, an innovative domain-generalized detection framework that achieves cross-domain robustness through restore-and-restyle data augmentation, frequency–spatial joint feature learning, and small-object perception enhancement. DG-YOLO first employs R<sup>2</sup>A-Net to generate physically constrained, style-diverse samples that simulate variations in illumination and water turbidity, thereby explicitly enriching domain diversity. The FreSpatial module disentangles structural information from domain-specific noise to extract robust, domain-invariant features. The SOEP (Small Object Enhancement Pyramid) module improves small-target perception while preserving fine-grained features, enhancing detection performance in complex backgrounds. Finally, the SlideLoss function adaptively reweights samples near decision boundaries, further strengthening the model’s discriminative capability. On the DLOU-SeaCucumber-DG dataset, DG-YOLO achieves up to 3.9% mAP improvement over existing domain generalization methods and YOLO variants, while maintaining a lightweight design. These results demonstrate both the methodological novelty and practical utility of DG-YOLO for intelligent and sustainable aquaculture management.</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

Underwater sea cucumber detection based on domain-invariant features

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Fangqun Niu,
  • Zejin Liu,
  • Xinyu Zheng,
  • Junyi Wang,
  • Dexuan Zhang,
  • Yuanshan Lin,
  • Xiaoting Zhu,
  • Shuai Yang,
  • Wei Wang,
  • Guodong Li

摘要

Domain generalization poses significant challenges for target detection in aquaculture, as variations in illumination, water quality, and suspended particles induce substantial domain shifts, while the small size and strong camouflage of sea cucumbers further complicate detection. To address these challenges, we propose DG-YOLO, an innovative domain-generalized detection framework that achieves cross-domain robustness through restore-and-restyle data augmentation, frequency–spatial joint feature learning, and small-object perception enhancement. DG-YOLO first employs R2A-Net to generate physically constrained, style-diverse samples that simulate variations in illumination and water turbidity, thereby explicitly enriching domain diversity. The FreSpatial module disentangles structural information from domain-specific noise to extract robust, domain-invariant features. The SOEP (Small Object Enhancement Pyramid) module improves small-target perception while preserving fine-grained features, enhancing detection performance in complex backgrounds. Finally, the SlideLoss function adaptively reweights samples near decision boundaries, further strengthening the model’s discriminative capability. On the DLOU-SeaCucumber-DG dataset, DG-YOLO achieves up to 3.9% mAP improvement over existing domain generalization methods and YOLO variants, while maintaining a lightweight design. These results demonstrate both the methodological novelty and practical utility of DG-YOLO for intelligent and sustainable aquaculture management.