Scholarlink

<p>A critical step toward edge intelligence in the Internet of Underwater Things (IoUT) is enabling advanced vision capabilities, such as real-time target detection, directly on endpoint devices. However, this task faces significant challenges due to complex underwater environmental factors, such as severe light attenuation, turbidity, and frequent biological occlusions, which often lead to blurred target features and a high rate of missed detections. To address these issues, this paper proposes ARFNet, an adaptively-routed fusion network built upon the YOLO11 framework. The core of ARFNet is a Shared Channel-Split Dual-Backbone module (SCS-DB) that establishes bidirectional feature flow, thereby enhancing information reuse for improved accuracy and efficiency. Furthermore, guided by ablation studies, we introduce a strategic replacement of Space-to-Depth convolution (SPD-Conv) to preserve fine-grained features of small targets, substantially boosting the recall rate. These designs collectively yield two variants: ARFNet-n and ARFNet-s. Finally, a DualPath-iA module incorporates a window-based attention mechanism to augment feature representations, thereby strengthening generalization in complex environments. Extensive experiments on the URPC2020 and RUOD datasets demonstrate that our method achieves mAP scores of 87.2% and 86.3%, respectively, with both variants attaining recall rates above 80% and a reduction in miss detection of over 10%, while maintaining a high inference speed of 214–219 FPS. These results effectively validate the efficacy of ARFNet in balancing detection accuracy, recall rate, and processing speed, offering a solution for IoUT-based applications.</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

ARFNet: a dual-backbone network for object detection in degraded underwater environments

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Lei Ma,
  • Xingxi Hu,
  • Dangguo Shao

摘要

A critical step toward edge intelligence in the Internet of Underwater Things (IoUT) is enabling advanced vision capabilities, such as real-time target detection, directly on endpoint devices. However, this task faces significant challenges due to complex underwater environmental factors, such as severe light attenuation, turbidity, and frequent biological occlusions, which often lead to blurred target features and a high rate of missed detections. To address these issues, this paper proposes ARFNet, an adaptively-routed fusion network built upon the YOLO11 framework. The core of ARFNet is a Shared Channel-Split Dual-Backbone module (SCS-DB) that establishes bidirectional feature flow, thereby enhancing information reuse for improved accuracy and efficiency. Furthermore, guided by ablation studies, we introduce a strategic replacement of Space-to-Depth convolution (SPD-Conv) to preserve fine-grained features of small targets, substantially boosting the recall rate. These designs collectively yield two variants: ARFNet-n and ARFNet-s. Finally, a DualPath-iA module incorporates a window-based attention mechanism to augment feature representations, thereby strengthening generalization in complex environments. Extensive experiments on the URPC2020 and RUOD datasets demonstrate that our method achieves mAP scores of 87.2% and 86.3%, respectively, with both variants attaining recall rates above 80% and a reduction in miss detection of over 10%, while maintaining a high inference speed of 214–219 FPS. These results effectively validate the efficacy of ARFNet in balancing detection accuracy, recall rate, and processing speed, offering a solution for IoUT-based applications.