Scholarlink

<p>Detection and removal of floating objects are crucial for water pollution control and the development of sustainable aquatic ecosystems. While unmanned platforms offer a promising solution, their limited computational resources, coupled with complex background interference (e.g., illumination, reflection), make accurate and efficient detection of small objects a major challenge. To overcome this challenge, we propose the SG-YOLO model, a novel lightweight model that excels in complex scenes and small object detection. Our approach incorporates three key innovations to address the core challenges: A lightweight cross-channel architecture combining heavily parameterized convolutions and an attention mechanism to enhance feature representation and robustness to complex interferences. The introduction of SPD-Conv and a specially designed SG-C2f module to comprehensively preserve and refine the feature details of small objects and prevent them from being lost in deep networks. A detection head with multi-scale feature fusion to enhance scale invariance and improve the accuracy and efficiency of small object detection. Extensive experiments on the IWHR_AI_Lable_Floater_V1 and FLOW-IMG datasets demonstrate that SG-YOLO achieves mAP@0.5 accuracy of 91.8 and 84.2%, respectively, improving upon YOLOv8 by 2.4 and 5.0%, while maintaining a parameter size of only 3.3&#xa0;M and achieving a frame rate of 131. Field tests and quantitative environmental performance analysis further validate its practicality, highlighting SG-YOLO’s excellent balance between lightweight design and efficient small object detection.</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

A lightweight model for efficient detection of small floating debris in complex aquatic environments

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Jianhua Ye,
  • Hao Liu,
  • Pan Li,
  • Zongsheng Yuan,
  • Fang Liu

摘要

Detection and removal of floating objects are crucial for water pollution control and the development of sustainable aquatic ecosystems. While unmanned platforms offer a promising solution, their limited computational resources, coupled with complex background interference (e.g., illumination, reflection), make accurate and efficient detection of small objects a major challenge. To overcome this challenge, we propose the SG-YOLO model, a novel lightweight model that excels in complex scenes and small object detection. Our approach incorporates three key innovations to address the core challenges: A lightweight cross-channel architecture combining heavily parameterized convolutions and an attention mechanism to enhance feature representation and robustness to complex interferences. The introduction of SPD-Conv and a specially designed SG-C2f module to comprehensively preserve and refine the feature details of small objects and prevent them from being lost in deep networks. A detection head with multi-scale feature fusion to enhance scale invariance and improve the accuracy and efficiency of small object detection. Extensive experiments on the IWHR_AI_Lable_Floater_V1 and FLOW-IMG datasets demonstrate that SG-YOLO achieves mAP@0.5 accuracy of 91.8 and 84.2%, respectively, improving upon YOLOv8 by 2.4 and 5.0%, while maintaining a parameter size of only 3.3 M and achieving a frame rate of 131. Field tests and quantitative environmental performance analysis further validate its practicality, highlighting SG-YOLO’s excellent balance between lightweight design and efficient small object detection.