Scholarlink

<p>Dense swarms of tiny objects and tight real-time budgets in unmanned aerial vehicle (UAV) imagery place severe pressure on detection systems. Conventional detectors lose critical edge cues during downsampling and struggle to model dense spatial relations, forcing difficult accuracy–speed trade-offs. We present ME-DETR (multi-scale edge-enhanced detector for real-time UAV object detection), a lightweight multi-scale edge-enhanced detector designed for resource-efficient deployment. ME-DETR introduces a multi-scale edge-enhanced cross-stage partial backbone with four-directional pooling to retain boundaries, a spatial global decomposition transformer (AIFI_SGD) that fuses global and local cues for dense-scene reasoning, and a dynamic scale selection slimneck that delivers efficient multi-scale fusion. On VisDrone2019, ME-DETR attains 51.2% mean average precision at 0.5 intersection-over-union, denoted mAP@50, on validation and 41.1% on the test set. Relative to the real-time detection transformer with a residual network (ResNet-18) backbone, it trims parameters by 33.8% and giga floating-point operations by 20.7%. These results indicate that ME-DETR is edge-oriented and maintains strong small-object detection 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

ME-DETR: a lightweight multi-scale edge-enhanced detector for real-time UAV object detection

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Xiaochuan Qian,
  • Haiwen Feng

摘要

Dense swarms of tiny objects and tight real-time budgets in unmanned aerial vehicle (UAV) imagery place severe pressure on detection systems. Conventional detectors lose critical edge cues during downsampling and struggle to model dense spatial relations, forcing difficult accuracy–speed trade-offs. We present ME-DETR (multi-scale edge-enhanced detector for real-time UAV object detection), a lightweight multi-scale edge-enhanced detector designed for resource-efficient deployment. ME-DETR introduces a multi-scale edge-enhanced cross-stage partial backbone with four-directional pooling to retain boundaries, a spatial global decomposition transformer (AIFI_SGD) that fuses global and local cues for dense-scene reasoning, and a dynamic scale selection slimneck that delivers efficient multi-scale fusion. On VisDrone2019, ME-DETR attains 51.2% mean average precision at 0.5 intersection-over-union, denoted mAP@50, on validation and 41.1% on the test set. Relative to the real-time detection transformer with a residual network (ResNet-18) backbone, it trims parameters by 33.8% and giga floating-point operations by 20.7%. These results indicate that ME-DETR is edge-oriented and maintains strong small-object detection performance.