Scholarlink

<p>Small object detection has always been a key challenge in the field of object detection due to its low resolution and limited information. Traditional object detection algorithms often struggle with small objects, as they fail to fully leverage low-level features and overlook the semantic gaps between feature maps. To address these issues, we introduce a novel small object detector called SO-YOLO. Our approach incorporates a cross-scale feature alignment (CSFA) module that employs high-level semantics to guide the spatial redistribution of low-level features, thereby reducing semantic gaps during feature fusion. Additionally, a small object feature enhancement (SOFE) module is designed to extract relevant spatial features from low-level feature maps and filter out background noise. Subsequently, a high-resolution feature pyramid network (HRFPN) is utilized to upsample all feature maps to the same high resolution, thereby increasing the likelihood of detecting small objects. Extensive experiments conducted on the challenging VisDrone2019, TT-100&#xa0;K, and SeaDronesSee datasets demonstrate that SO-YOLO significantly improves the detection performance of small objects. Code is available at <a href="https://github.com/Anewman211/SO-YOLO">https://github.com/Anewman211/SO-YOLO</a>.</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

Cross-scale feature alignment and feature enhancement for small object detection

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Binglin Wu,
  • Yi Wang,
  • Xiaxu Huang,
  • Guoliang Xie

摘要

Small object detection has always been a key challenge in the field of object detection due to its low resolution and limited information. Traditional object detection algorithms often struggle with small objects, as they fail to fully leverage low-level features and overlook the semantic gaps between feature maps. To address these issues, we introduce a novel small object detector called SO-YOLO. Our approach incorporates a cross-scale feature alignment (CSFA) module that employs high-level semantics to guide the spatial redistribution of low-level features, thereby reducing semantic gaps during feature fusion. Additionally, a small object feature enhancement (SOFE) module is designed to extract relevant spatial features from low-level feature maps and filter out background noise. Subsequently, a high-resolution feature pyramid network (HRFPN) is utilized to upsample all feature maps to the same high resolution, thereby increasing the likelihood of detecting small objects. Extensive experiments conducted on the challenging VisDrone2019, TT-100 K, and SeaDronesSee datasets demonstrate that SO-YOLO significantly improves the detection performance of small objects. Code is available at https://github.com/Anewman211/SO-YOLO.