Scholarlink

<p>As a core technology for intelligent interpretation of low-altitude remote sensing, UAV image object detection faces two inherent and interrelated challenges: the localization sensitivity across scales and the long-tail distribution problem. Existing IoU-based metrics fail to properly reflect bounding box quality across multiple scales, while the long-tail distribution often causes information of tail categories to be overwhelmed during training. To address these issues, we propose YOLO-KLS, a collaborative optimization method that combines a Dual-branch Sampler (DBS) to refocus training on tail classes and a Transformed KL divergence (TKLD) metric to improve scale-robust localization. Furthermore, we introduce a High-resolution Detection Layer (HDL) to provide more samples and feature information about small object, and employ Quality Focal Loss (QFL) to prioritize challenging targets, ensuring stable training under class imbalance. Experiments on VisDrone2019 and UAVDT demonstrate performance gains of 5.1% and 5.6%, respectively, showing the effectiveness of our approach.</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

Transformed KL divergence and dual-branch sampler for UAV image object detection

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Junbao Wu,
  • Hao Meng,
  • Zhe Lu,
  • Chang Liu

摘要

As a core technology for intelligent interpretation of low-altitude remote sensing, UAV image object detection faces two inherent and interrelated challenges: the localization sensitivity across scales and the long-tail distribution problem. Existing IoU-based metrics fail to properly reflect bounding box quality across multiple scales, while the long-tail distribution often causes information of tail categories to be overwhelmed during training. To address these issues, we propose YOLO-KLS, a collaborative optimization method that combines a Dual-branch Sampler (DBS) to refocus training on tail classes and a Transformed KL divergence (TKLD) metric to improve scale-robust localization. Furthermore, we introduce a High-resolution Detection Layer (HDL) to provide more samples and feature information about small object, and employ Quality Focal Loss (QFL) to prioritize challenging targets, ensuring stable training under class imbalance. Experiments on VisDrone2019 and UAVDT demonstrate performance gains of 5.1% and 5.6%, respectively, showing the effectiveness of our approach.