Scholarlink

<p>The ocean plays a key role in the global climate and natural cycles, while also providing abundant resources for sustainable development. As marine exploration increases, underwater image recognition technology has become increasingly important. However, existing object detection methods face difficulties when processing underwater targets like seagrass, sea snakes, and jellyfish. These targets are often characterized by their small size, weak features, and long, thin shapes. To solve these problems, this paper proposes a dynamic snake convolution YOLO (DSCYOLO) algorithm for underwater image recognition. The proposed model includes three main improvements. First, dynamic snake convolution (DSConv) is used to replace standard convolutions, which improves the feature extraction capability for thin, tube-like targets. Second, a mixed local channel attention (MLCA) mechanism is added to combine local and global features effectively. Finally, an extra detection head is added to improve the detection accuracy for small underwater organisms. The effectiveness of these improvements is verified through comparative ablation experiments. Experimental results show that DSCYOLO performs well, achieving mean Average Precision (mAP) values of 72.6%, 91.0%, and 84.4% on the URPC2020, DUO, and RUOD datasets, respectively.</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

Dscyolo: dynamic snake convolutional YOLO network for underwater image recognition

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Mengqi Xue,
  • Baoju Zhang,
  • Cuiping Zhang,
  • Bo Zhang

摘要

The ocean plays a key role in the global climate and natural cycles, while also providing abundant resources for sustainable development. As marine exploration increases, underwater image recognition technology has become increasingly important. However, existing object detection methods face difficulties when processing underwater targets like seagrass, sea snakes, and jellyfish. These targets are often characterized by their small size, weak features, and long, thin shapes. To solve these problems, this paper proposes a dynamic snake convolution YOLO (DSCYOLO) algorithm for underwater image recognition. The proposed model includes three main improvements. First, dynamic snake convolution (DSConv) is used to replace standard convolutions, which improves the feature extraction capability for thin, tube-like targets. Second, a mixed local channel attention (MLCA) mechanism is added to combine local and global features effectively. Finally, an extra detection head is added to improve the detection accuracy for small underwater organisms. The effectiveness of these improvements is verified through comparative ablation experiments. Experimental results show that DSCYOLO performs well, achieving mean Average Precision (mAP) values of 72.6%, 91.0%, and 84.4% on the URPC2020, DUO, and RUOD datasets, respectively.