Scholarlink

<p>To address challenges posed by complex background interference, multi-scale target variation, and image degradation in underwater garbage detection, this study proposes an enhanced underwater garbage detection model based on the YOLOv11 architecture. The model incorporates a novel multi-branch dilated attention mechanism (MBDA) to capture multi-scale features using dilated convolutions with varying dilation rates. It also integrates a channel attention mechanism to support collaborative modeling of local details and global context. To mitigate detail loss associated with dilated convolutions, the model employs depthwise separable convolution branching (DWB), aiming to enhance fine-grained feature extraction while maintaining relatively low computational complexity. Additionally, the model adopts the NWDLoss function, based on normalized Wasserstein distance, to improve bounding box regression training and detection performance for small and irregular objects. Experimental results show a 1.1% and 5.4% improvement in the mAP50-95 metric on the J-EDI validation and test sets, respectively. The model achieves an inference speed of 208.3 FPS with a lightweight parameter size of 2.75M. Compared with other comparison models, the proposed approach demonstrates competitive detection accuracy and efficiency, suggesting its potential applicability in complex underwater environments.</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 garbage detection method for environmental protection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Yuhong Chen,
  • Huicong Ning,
  • Peinan Hao,
  • Jiajia Chen,
  • Ying Zhang

摘要

To address challenges posed by complex background interference, multi-scale target variation, and image degradation in underwater garbage detection, this study proposes an enhanced underwater garbage detection model based on the YOLOv11 architecture. The model incorporates a novel multi-branch dilated attention mechanism (MBDA) to capture multi-scale features using dilated convolutions with varying dilation rates. It also integrates a channel attention mechanism to support collaborative modeling of local details and global context. To mitigate detail loss associated with dilated convolutions, the model employs depthwise separable convolution branching (DWB), aiming to enhance fine-grained feature extraction while maintaining relatively low computational complexity. Additionally, the model adopts the NWDLoss function, based on normalized Wasserstein distance, to improve bounding box regression training and detection performance for small and irregular objects. Experimental results show a 1.1% and 5.4% improvement in the mAP50-95 metric on the J-EDI validation and test sets, respectively. The model achieves an inference speed of 208.3 FPS with a lightweight parameter size of 2.75M. Compared with other comparison models, the proposed approach demonstrates competitive detection accuracy and efficiency, suggesting its potential applicability in complex underwater environments.