Scholarlink

In response to the escalating security threats posed by unauthorized drone activities, this paper proposes ERUAVNet, a lightweight neural network specifically designed for real-time drone detection through the application of structural reparameterization techniques. The proposed network features an innovative Multi-Group Parallel Convolution Block that effectively captures both local and global channel features via parallel grouped convolutions. Additionally, we introduce the Mixed Local Channel Attention mechanism that integrates both spatial and channel information to strengthen feature representation. Furthermore, we present an improved Efficient RepNeck module that optimizes feature fusion while maintaining rapid inference capabilities. Despite its remarkably compact size of only 3.57 MB, ERUAVNet demonstrates exceptional performance across three challenging datasets: Det-Fly, TIBNet, and DUT-Anti-UAV. Comprehensive experiments indicate that our model outperforms existing mainstream object detection algorithms, exhibiting superior efficiency and precision in drone detection tasks.

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

ERUAVNet: An Efficient Reparameterized Network for Unmanned Aerial Vehicle Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Peizheng He,
  • Bo Yu,
  • Dayu Gao,
  • Shun Li,
  • Lin Xu,
  • Gongbo Chen

摘要

In response to the escalating security threats posed by unauthorized drone activities, this paper proposes ERUAVNet, a lightweight neural network specifically designed for real-time drone detection through the application of structural reparameterization techniques. The proposed network features an innovative Multi-Group Parallel Convolution Block that effectively captures both local and global channel features via parallel grouped convolutions. Additionally, we introduce the Mixed Local Channel Attention mechanism that integrates both spatial and channel information to strengthen feature representation. Furthermore, we present an improved Efficient RepNeck module that optimizes feature fusion while maintaining rapid inference capabilities. Despite its remarkably compact size of only 3.57 MB, ERUAVNet demonstrates exceptional performance across three challenging datasets: Det-Fly, TIBNet, and DUT-Anti-UAV. Comprehensive experiments indicate that our model outperforms existing mainstream object detection algorithms, exhibiting superior efficiency and precision in drone detection tasks.