Scholarlink

<p>This paper introduces LSOD-DETR, a lightweight small object detection model based on the real-time detection transformer, designed to address the challenges of low pixel counts and indistinct features in unmanned scenes. To enhance multi-scale feature perception, we design a multi-scale edge information enhancement (MSEIE) module that combines learnable edge extraction with convolution operations. Additionally, a small object detection feature enhancement network (SDENet) is proposed to strengthen the detection of tiny targets by fusing shallow details. Furthermore, a knowledge distillation method is introduced to optimize the trade-off between performance and complexity. Experimental results on the Tsinghua-Tencent 100K dataset demonstrate its superiority, achieving a precision of 91.6%, recall of 86.7%, and mAP (0.5) of 88.9%. Significantly, the model demonstrates exceptional efficiency suitable for real-time edge deployment, operating with only 14.1 million parameters and 44.3 GFLOPs, while achieving an inference speed of 65 FPS. Generalization tests on the VisDrone dataset further validate its robustness, with mAP (0.5) reaching 40.3%.</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

LSOD-DETR: a lightweight small object detection model based on real-time detection transformer

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Lili Zhang,
  • Wenshuo Han,
  • Kang Yang,
  • Ke Zhang,
  • Long Zhang,
  • Ruiyang Xiao,
  • Jing Li,
  • Wei Wei,
  • Pei Yu,
  • Hongxin Tan

摘要

This paper introduces LSOD-DETR, a lightweight small object detection model based on the real-time detection transformer, designed to address the challenges of low pixel counts and indistinct features in unmanned scenes. To enhance multi-scale feature perception, we design a multi-scale edge information enhancement (MSEIE) module that combines learnable edge extraction with convolution operations. Additionally, a small object detection feature enhancement network (SDENet) is proposed to strengthen the detection of tiny targets by fusing shallow details. Furthermore, a knowledge distillation method is introduced to optimize the trade-off between performance and complexity. Experimental results on the Tsinghua-Tencent 100K dataset demonstrate its superiority, achieving a precision of 91.6%, recall of 86.7%, and mAP (0.5) of 88.9%. Significantly, the model demonstrates exceptional efficiency suitable for real-time edge deployment, operating with only 14.1 million parameters and 44.3 GFLOPs, while achieving an inference speed of 65 FPS. Generalization tests on the VisDrone dataset further validate its robustness, with mAP (0.5) reaching 40.3%.