Scholarlink

<p>Small object detection remains a significant challenge in computer vision due to the minimal size and vulnerability to overshadowing by larger objects. Current approaches often suffer from limited perspectives and inadequate feature collaboration. In this study, we introduce MicroSight, an efficient transformer-based middleware designed to enhance small object detection. By incorporating global attention interaction and local-range optimization, MicroSight extracts and enriches features from both global and refined perspectives. Additionally, a semantic attention contribution matrix is introduced to facilitate multi-perspective feature collaboration. Our experiments on the COCO dataset demonstrate that MicroSight improves detection accuracy on small objects by up to 21.3% AP and accelerates training speed by 5.0x. These findings underscore the potential of transformer-based approaches in advancing small object detection. Code address: <a href="https://github.com/Dazi11/MicroSight">https://github.com/Dazi11/MicroSight</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

Enhancing small object detection: a transformer-based middleware approach

  • 期刊论文

  • 发表日期：

    2026年6月21日

查看全文
  • Fei Gu,
  • Zijun Wang,
  • Ziheng Zhang,
  • Zeyang Zhang,
  • Jing Wu

摘要

Small object detection remains a significant challenge in computer vision due to the minimal size and vulnerability to overshadowing by larger objects. Current approaches often suffer from limited perspectives and inadequate feature collaboration. In this study, we introduce MicroSight, an efficient transformer-based middleware designed to enhance small object detection. By incorporating global attention interaction and local-range optimization, MicroSight extracts and enriches features from both global and refined perspectives. Additionally, a semantic attention contribution matrix is introduced to facilitate multi-perspective feature collaboration. Our experiments on the COCO dataset demonstrate that MicroSight improves detection accuracy on small objects by up to 21.3% AP and accelerates training speed by 5.0x. These findings underscore the potential of transformer-based approaches in advancing small object detection. Code address: https://github.com/Dazi11/MicroSight.