Scholarlink

<p>To address the challenges of inaccurate and missed detections of small and occluded targets in ship detection tasks under complex scenes, this study proposes a novel small-object ship detection method based on multi-level feature enhancement and cross-layer fusion. First, the backbone network is redesigned with a convolution attention fusion module. By synergistically leveraging the local detail capture capability of convolution and the global dependency modeling of attention, it significantly enhances the ability to extract detailed features of small targets and occluded targets.Subsequently, a cross-level interactive feature reorganization pyramid network is developed to strengthen multi-scale feature fusion through cross-level interaction, enhance the transmission of high-level semantic information, and improve the semantic representation of shallow and medium features. Furthermore, an efficient local attention module is introduced to effectively capture key ship features, suppress interference from complex maritime backgrounds, and reduce missed and false detections. Experimental results demonstrate that the proposed method achieves detection accuracies of 83.30% and 83.52% on the self-built USSD dataset and the HRSC2016 dataset, respectively, demonstrating superior performance compared to existing methods.The code is available at <a href="https://github.com/ZDS56/mlcf-sod">https://github.com/ZDS56/mlcf-sod</a>.</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-level feature enhancement and cross-layer fusion for small object ship detection

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Jing Zhang,
  • Zining Wu,
  • Gaoxiang Zheng,
  • Zhan Liu,
  • Yongheng Tang

摘要

To address the challenges of inaccurate and missed detections of small and occluded targets in ship detection tasks under complex scenes, this study proposes a novel small-object ship detection method based on multi-level feature enhancement and cross-layer fusion. First, the backbone network is redesigned with a convolution attention fusion module. By synergistically leveraging the local detail capture capability of convolution and the global dependency modeling of attention, it significantly enhances the ability to extract detailed features of small targets and occluded targets.Subsequently, a cross-level interactive feature reorganization pyramid network is developed to strengthen multi-scale feature fusion through cross-level interaction, enhance the transmission of high-level semantic information, and improve the semantic representation of shallow and medium features. Furthermore, an efficient local attention module is introduced to effectively capture key ship features, suppress interference from complex maritime backgrounds, and reduce missed and false detections. Experimental results demonstrate that the proposed method achieves detection accuracies of 83.30% and 83.52% on the self-built USSD dataset and the HRSC2016 dataset, respectively, demonstrating superior performance compared to existing methods.The code is available at https://github.com/ZDS56/mlcf-sod.