Scholarlink

<p>Due to the huge differences in object scale, frequent occlusions, and few pixels occupied by targets, small object detection in drone aerial imagery remains challenging. To address these challenges, we propose feature optimization YOLO (FO-YOLO), a method for improving the detection accuracy of small objects in drone imagery. FO-YOLO consists of three main modules: feature augment module (FAM), dynamic detail-aware feature pyramid network (DDAFPN), and optimal transport assignment-EIoU (OTA-E). Firstly, FAM is a feature-rich module that can balance the richness and refinement of feature extraction. Secondly, we designed DDAFPN to tackle the issue of insufficient utilization of detailed information in traditional feature pyramid networks. Meanwhile, it can also reduce the redundant information when fusing features. Thirdly, OTA-E is a label assignment method, designed to optimize the global label allocation strategy and assign labels more accurately. Extensive experiments on the VisDrone and Tinyperson datasets demonstrate the effectiveness and advancement of our method.</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

FO-YOLO for small object detection in drone aerial imagery

  • 期刊论文

  • 发表日期：

    2025年8月3日

查看全文
  • Huaping Zhou,
  • Wei Yin,
  • Kelei Sun,
  • Tao Wu,
  • Bin Deng

摘要

Due to the huge differences in object scale, frequent occlusions, and few pixels occupied by targets, small object detection in drone aerial imagery remains challenging. To address these challenges, we propose feature optimization YOLO (FO-YOLO), a method for improving the detection accuracy of small objects in drone imagery. FO-YOLO consists of three main modules: feature augment module (FAM), dynamic detail-aware feature pyramid network (DDAFPN), and optimal transport assignment-EIoU (OTA-E). Firstly, FAM is a feature-rich module that can balance the richness and refinement of feature extraction. Secondly, we designed DDAFPN to tackle the issue of insufficient utilization of detailed information in traditional feature pyramid networks. Meanwhile, it can also reduce the redundant information when fusing features. Thirdly, OTA-E is a label assignment method, designed to optimize the global label allocation strategy and assign labels more accurately. Extensive experiments on the VisDrone and Tinyperson datasets demonstrate the effectiveness and advancement of our method.