Scholarlink

<p>Underwater target detection is crucial for marine monitoring, ecological assessment, and intelligent operations. However, there are significant challenges for existing detection algorithms, such as light attenuation, color distortion, background noise, and target scale variations. To address these issues, an underwater multi-scale adaptive detection network with cross-stage pyramid extraction, adaptive multi-scale fusion, and multi-granularity perception (CDM-YOLOv8n) is proposed. The network consists of three core components: the Efficient Cross-Stage Bottleneck with a Pyramid Structure Feature Extraction (Conv2CSP), the Multi-Scale Adaptive Information Enhancement and Fusion (DownCSRA), and the Multi-Granularity Information Awareness (MGIA). Firstly, Conv2CSP enhances feature extraction to resolve overlapping object boundaries in dense underwater scenes. Secondly, DownCSRA improves feature representation under complex backgrounds by adaptively integrating multi-scale information. Finally, MGIA employs dedicated fine-grained perception branches to better capture target features. Experiments on the URPC2020 and RUOD datasets show that CDM-YOLOv8n improves mAP@0.5 by 2.9% and 2.3% and mAP@0.5:0.95 by 2.1% and 2.3%. Evaluation on Small and Overlapping Objects highlights substantial improvements, with mAP@0.5 improving by 2.6% on URPC2020 and 3.5% on RUOD for overlapping objects and by 2.2% on URPC2020 and 3.0% on RUOD for small objects. Compared with state-of-the-art underwater methods, CDM-YOLOv8n achieves over 30 FPS on GPU, ensuring accurate real-time detection in compute-intensive underwater scenes and demonstrating scalability to HPC platforms.</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

Multi-scale adaptive small and overlapping target detection in underwater images

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Lisha Luo,
  • Qiang Gao,
  • Qi Liu,
  • Xuan Li,
  • Xiao Yu,

摘要

Underwater target detection is crucial for marine monitoring, ecological assessment, and intelligent operations. However, there are significant challenges for existing detection algorithms, such as light attenuation, color distortion, background noise, and target scale variations. To address these issues, an underwater multi-scale adaptive detection network with cross-stage pyramid extraction, adaptive multi-scale fusion, and multi-granularity perception (CDM-YOLOv8n) is proposed. The network consists of three core components: the Efficient Cross-Stage Bottleneck with a Pyramid Structure Feature Extraction (Conv2CSP), the Multi-Scale Adaptive Information Enhancement and Fusion (DownCSRA), and the Multi-Granularity Information Awareness (MGIA). Firstly, Conv2CSP enhances feature extraction to resolve overlapping object boundaries in dense underwater scenes. Secondly, DownCSRA improves feature representation under complex backgrounds by adaptively integrating multi-scale information. Finally, MGIA employs dedicated fine-grained perception branches to better capture target features. Experiments on the URPC2020 and RUOD datasets show that CDM-YOLOv8n improves mAP@0.5 by 2.9% and 2.3% and mAP@0.5:0.95 by 2.1% and 2.3%. Evaluation on Small and Overlapping Objects highlights substantial improvements, with mAP@0.5 improving by 2.6% on URPC2020 and 3.5% on RUOD for overlapping objects and by 2.2% on URPC2020 and 3.0% on RUOD for small objects. Compared with state-of-the-art underwater methods, CDM-YOLOv8n achieves over 30 FPS on GPU, ensuring accurate real-time detection in compute-intensive underwater scenes and demonstrating scalability to HPC platforms.