Scholarlink

<p>In response to the burgeoning security concerns associated with the proliferation of unmanned aerial vehicles (UAVs), this study introduces an innovative network architecture dubbed big-LITTLE-Net, tailored for detecting small-sized UAVs within the image. Conventional UAV detection methods need to strike a balance between the scale of the model and the detection accuracy, and face difficulties in detecting small targets. The proposed big-LITTLE-Net encapsulates a dual-branch approach, with the big branch conducting multi-scale feature extraction, while the LITTLE branch focuses on feature collaboration and supplementation, enhancing the overall field of view. Besides, two innovative bridge modules, the feature reconstruction module and the feature fusion bus module, promote functional collaborative processing in features and branches. The feature reconstruction module enables information flow between different network architecture levels, and the feature fusion bus module acts as a hub to combine image features from different branches. In comparison, within three datasets, the detection capability of the big-LITTLE-Net is enhanced by an average of 3.1%, and the detection accuracy for small targets is increased on average by 10.87%. At the same time, the parameter of the model is reduced by 44%. This research also covers ablation studies that contrast single and big-LITTLE architectures, with results underscoring the efficacy and innovation of the new architecture design.</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

Big-LITTLE-Net: a dual-branch network for small UAV detection

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Yinjie Chen,
  • Wenyi Tang,
  • Yunbo Rao,
  • Hui Ding,
  • Shuzhen Zhu,
  • Yuanyuan Wang

摘要

In response to the burgeoning security concerns associated with the proliferation of unmanned aerial vehicles (UAVs), this study introduces an innovative network architecture dubbed big-LITTLE-Net, tailored for detecting small-sized UAVs within the image. Conventional UAV detection methods need to strike a balance between the scale of the model and the detection accuracy, and face difficulties in detecting small targets. The proposed big-LITTLE-Net encapsulates a dual-branch approach, with the big branch conducting multi-scale feature extraction, while the LITTLE branch focuses on feature collaboration and supplementation, enhancing the overall field of view. Besides, two innovative bridge modules, the feature reconstruction module and the feature fusion bus module, promote functional collaborative processing in features and branches. The feature reconstruction module enables information flow between different network architecture levels, and the feature fusion bus module acts as a hub to combine image features from different branches. In comparison, within three datasets, the detection capability of the big-LITTLE-Net is enhanced by an average of 3.1%, and the detection accuracy for small targets is increased on average by 10.87%. At the same time, the parameter of the model is reduced by 44%. This research also covers ablation studies that contrast single and big-LITTLE architectures, with results underscoring the efficacy and innovation of the new architecture design.