Scholarlink

<p>Detecting small objects in UAV imagery is challenging due to limited features and size variations caused by diverse angles and high altitudes. To address this, we propose a novel Unidirectional Guidance Network (UGNet) that enhances detection accuracy for small-scale targets while preventing poor feature learning. UGNet comprises three key components: the Unidirectional Feature Pyramid Network (UnFPN), the Space Transformer Attention (STA) module, and the Auxiliary Guidance Space-to-Depth (AGSD) module. UnFPN provides repeated high-level semantic guidance to low-level detailed features, balancing high resolution and strong semantics. The STA module adaptively enhances small object features, while the AGSD module prevents information loss during downsampling. Experiments on challenging datasets, VisDrone2019-DET and AI-TOD, demonstrate that UGNet improves AP50 from 44.1% to 53.3% on VisDrone2019-DET and from 45.5% to 53.0% on AI-TOD, while simultaneously reducing the parameter count by 18.5M compared to the baseline. These results indicate that UGNet effectively balances performance enhancement and model efficiency, making it suitable for UAV-based object detection tasks. The source code is available at <a href="https://github.com/medeID113/UGNet">https://github.com/medeID113/UGNet</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

Unidirectional guidance network for enhanced small object detection in UAV imagery

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Xin Deng,
  • Zheng Xu,
  • Wenzhu Yang

摘要

Detecting small objects in UAV imagery is challenging due to limited features and size variations caused by diverse angles and high altitudes. To address this, we propose a novel Unidirectional Guidance Network (UGNet) that enhances detection accuracy for small-scale targets while preventing poor feature learning. UGNet comprises three key components: the Unidirectional Feature Pyramid Network (UnFPN), the Space Transformer Attention (STA) module, and the Auxiliary Guidance Space-to-Depth (AGSD) module. UnFPN provides repeated high-level semantic guidance to low-level detailed features, balancing high resolution and strong semantics. The STA module adaptively enhances small object features, while the AGSD module prevents information loss during downsampling. Experiments on challenging datasets, VisDrone2019-DET and AI-TOD, demonstrate that UGNet improves AP50 from 44.1% to 53.3% on VisDrone2019-DET and from 45.5% to 53.0% on AI-TOD, while simultaneously reducing the parameter count by 18.5M compared to the baseline. These results indicate that UGNet effectively balances performance enhancement and model efficiency, making it suitable for UAV-based object detection tasks. The source code is available at https://github.com/medeID113/UGNet.