Scholarlink

<p>Aerial small target detection in complex scenarios has been widely adopted in daily applications, yet it remains challenging due to diverse target scales, cluttered backgrounds, and small target aggregation. This paper proposes a novel network architecture for aerial small target detection in complex scenarios. First, a lightweight LIConv-block module is designed to enhance high-contribution information and suppress redundant data. Subsequently, the IDSS module is introduced to strengthen intra-scale feature aggregation and forward-propagate the most effective information. Additionally, DLDHead is proposed as the detection head to capture fine-grained details of small targets, addressing the issues of small scale and blurred edges inherent in small targets. Finally, linear interval mapping is employed to achieve dynamic sample balancing, resolving the problem of imbalanced sample distribution. Experimental results on the VisDrone2019 dataset demonstrate that the proposed model improves the mAP50 and mAP50:95 metrics by 5% and 3.7%, 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

SF-DETR: aerial small target detection network based on scale fusion and fine-grained enhancement

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Renzheng Xue,
  • Shijie Hua

摘要

Aerial small target detection in complex scenarios has been widely adopted in daily applications, yet it remains challenging due to diverse target scales, cluttered backgrounds, and small target aggregation. This paper proposes a novel network architecture for aerial small target detection in complex scenarios. First, a lightweight LIConv-block module is designed to enhance high-contribution information and suppress redundant data. Subsequently, the IDSS module is introduced to strengthen intra-scale feature aggregation and forward-propagate the most effective information. Additionally, DLDHead is proposed as the detection head to capture fine-grained details of small targets, addressing the issues of small scale and blurred edges inherent in small targets. Finally, linear interval mapping is employed to achieve dynamic sample balancing, resolving the problem of imbalanced sample distribution. Experimental results on the VisDrone2019 dataset demonstrate that the proposed model improves the mAP50 and mAP50:95 metrics by 5% and 3.7%, respectively.