Scholarlink

<p>Synthetic Aperture Radar (SAR) can work under all-weather and all-day conditions, and it has been widely applied in maritime fields. However, existing algorithms based on SAR still have limitations in ship detection due to the complex scenarios and small-sized targets. In order to solve those problems, a Symmetric Convolution Network based on YOLOv7 (SCNet-YOLO) was proposed for SAR ship target detection. Firstly, we have designed a Symmetric Convolutional structure (SConv),&#xa0;it is highly effective in enhancing the capability of feature extraction for small targets. Secondly, we add the dynamic head to fuse the attention mechanism and better capture the ship targets of various sizes. Finally, the WIoU is used as the loss function to capture the border information while avoiding overfitting. The experiment results show that the SCNet-YOLO proposed in this paper achieves a precision of 93.5% and mAP50:95 of 68.1% on the HRSID, and achieves a precision of 96.1% and mAP50:95 of 71.7% on the SSDD, and it is superior to other state-of-the-art SAR ship detection methods in the overall performance.</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

SCNet-YOLO: a symmetric convolution network for multi-scenario ship detection based on YOLOv7

  • 期刊论文

  • 发表日期：

    2025年3月15日

查看全文
  • Weina Zhou,
  • Yuqi Yang,
  • Ming Zhao,
  • Wenhua Hu

摘要

Synthetic Aperture Radar (SAR) can work under all-weather and all-day conditions, and it has been widely applied in maritime fields. However, existing algorithms based on SAR still have limitations in ship detection due to the complex scenarios and small-sized targets. In order to solve those problems, a Symmetric Convolution Network based on YOLOv7 (SCNet-YOLO) was proposed for SAR ship target detection. Firstly, we have designed a Symmetric Convolutional structure (SConv), it is highly effective in enhancing the capability of feature extraction for small targets. Secondly, we add the dynamic head to fuse the attention mechanism and better capture the ship targets of various sizes. Finally, the WIoU is used as the loss function to capture the border information while avoiding overfitting. The experiment results show that the SCNet-YOLO proposed in this paper achieves a precision of 93.5% and mAP50:95 of 68.1% on the HRSID, and achieves a precision of 96.1% and mAP50:95 of 71.7% on the SSDD, and it is superior to other state-of-the-art SAR ship detection methods in the overall performance.