Scholarlink

<p>Remote sensing image detection using deep learning has emerged as a pivotal research area. Recent studies have focused on enhancing feature extraction and multi-scale fusion, often at the cost of increased inference latency. In UAV remote sensing, achieving a balance between detection accuracy and inference speed is crucial. This paper introduces GSR-Net, a lightweight Transformer-based model tailored for real-time UAV remote sensing object detection. GSR-Net integrates HGNetv2 with Ghost convolution and ARConv, specifically designed for small object detection, to create an efficient and lightweight architecture. This design significantly reduces model parameters and computational complexity without sacrificing detection accuracy. Furthermore, GSR-Net adopts a streamlined feature fusion strategy (SFF) that fuses only low-noise upper-level features, optimizing performance. An improved particle swarm optimization algorithm is introduced for hyperparameter tuning during training, shifting computational burden from inference to training. Experiments on the VisDrone2019 dataset demonstrate GSR-Net’s superiority, achieving 93.9 FPS on an NVIDIA RTX4090D GPU, outpacing RT-DETR by 35.8 FPS, with improved AP and AP50 by 1.6 and 2.2%, respectively, using 19.6M fewer parameters and only 50.7 GFLOPs. Real-world UAV experiments further validate GSR-Net’s robustness against frame loss and missed detections, offering a practical solution for drone inspection applications. The code is open source at <a href="https://github.com/mpwmpw/GSR-Net">https://github.com/mpwmpw/GSR-Net</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

GSR-Net: enhancing real-time UAV remote sensing object detection via a lightweight transformer model

  • 期刊论文

  • 发表日期：

    2026年5月3日

查看全文
  • Pengwei Ma,
  • Hongmei Fei,
  • Yao Zhang,
  • Nan Lian,
  • Fuyong Liu,
  • Jie Zhou

摘要

Remote sensing image detection using deep learning has emerged as a pivotal research area. Recent studies have focused on enhancing feature extraction and multi-scale fusion, often at the cost of increased inference latency. In UAV remote sensing, achieving a balance between detection accuracy and inference speed is crucial. This paper introduces GSR-Net, a lightweight Transformer-based model tailored for real-time UAV remote sensing object detection. GSR-Net integrates HGNetv2 with Ghost convolution and ARConv, specifically designed for small object detection, to create an efficient and lightweight architecture. This design significantly reduces model parameters and computational complexity without sacrificing detection accuracy. Furthermore, GSR-Net adopts a streamlined feature fusion strategy (SFF) that fuses only low-noise upper-level features, optimizing performance. An improved particle swarm optimization algorithm is introduced for hyperparameter tuning during training, shifting computational burden from inference to training. Experiments on the VisDrone2019 dataset demonstrate GSR-Net’s superiority, achieving 93.9 FPS on an NVIDIA RTX4090D GPU, outpacing RT-DETR by 35.8 FPS, with improved AP and AP50 by 1.6 and 2.2%, respectively, using 19.6M fewer parameters and only 50.7 GFLOPs. Real-world UAV experiments further validate GSR-Net’s robustness against frame loss and missed detections, offering a practical solution for drone inspection applications. The code is open source at https://github.com/mpwmpw/GSR-Net.