Scholarlink

<p>Object detection in unmanned aerial vehicle (UAV) imagery remains challenging due to small object sizes, dense distributions, and frequent occlusions. These challenges are further aggravated by limited onboard computing resources, which impose strict constraints on model complexity and efficiency. In this paper, we propose LEA-DETR, a lightweight and efficient transformer-based framework for UAV object detection. Specifically, we introduce an adaptive sparse–dense attention fusion mechanism to enhance discriminative feature modeling, together with an Efficient Multi-Scale Convolutional Feature Pyramid Network (EMS-FPN) to better integrate fine-grained details and high-level semantics, especially for small objects. To further improve local representation under lightweight constraints, we design a compact backbone based on gated inverted bottleneck convolutions. Experiments on the VisDrone2019 dataset show that LEA-DETR reduces model parameters by up to 65.7% while achieving a 1.3% mAP improvement over the baseline. These results demonstrate that LEA-DETR provides an effective and practical solution for real-world UAV object detection.</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

LEA-DETR: a lightweight and efficient attention-enhanced model for UAV object detection

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Haohao Ma,
  • Mingliang Zuo,
  • Qiqi Ge

摘要

Object detection in unmanned aerial vehicle (UAV) imagery remains challenging due to small object sizes, dense distributions, and frequent occlusions. These challenges are further aggravated by limited onboard computing resources, which impose strict constraints on model complexity and efficiency. In this paper, we propose LEA-DETR, a lightweight and efficient transformer-based framework for UAV object detection. Specifically, we introduce an adaptive sparse–dense attention fusion mechanism to enhance discriminative feature modeling, together with an Efficient Multi-Scale Convolutional Feature Pyramid Network (EMS-FPN) to better integrate fine-grained details and high-level semantics, especially for small objects. To further improve local representation under lightweight constraints, we design a compact backbone based on gated inverted bottleneck convolutions. Experiments on the VisDrone2019 dataset show that LEA-DETR reduces model parameters by up to 65.7% while achieving a 1.3% mAP improvement over the baseline. These results demonstrate that LEA-DETR provides an effective and practical solution for real-world UAV object detection.