Scholarlink

<p>Although deep learning-based object detection has achieved remarkable progress in natural scenes, unmanned aerial vehicle (UAV) imagery remains challenging due to limited contextual cues, a dominance of small-scale objects, and significant scale variation. To address these issues, we propose the Multiscale Dual Residual Network (MDRNet), which comprises three key modules. The Large Kernel Bottleneck Module (LKBM) expands the receptive field through multi-scale convolutions and cross-scale skip connections, capturing fine-grained contextual features. The Dual Feature Fusion Upsampling Module (DFFUM) enhances small object representation by merging shallow and deep features via dual LKBMs and upsampling. The Dual Residual Feature Pyramid Network (DR_FPN) further strengthens multi-scale feature integration by stacking residual-enhanced layers. MDRNet achieves notable performance gains on VisDrone, and SIMD datasets, outperforming a strong baseline by 3.5%, and 0.8% in mAP@50. And further validated the effectiveness and universality of MDRNet in the autonomous driving road scene dataset (RSUD20K), achieving map@50 significant achievements of 74.0%.</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

MDRNet: Multiscale Dual Residual Network for Small Object Detection on UAV Aerial Images

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Hui Zong,
  • Tianlong Ni,
  • Hao Wang,
  • Quanyin Zhu

摘要

Although deep learning-based object detection has achieved remarkable progress in natural scenes, unmanned aerial vehicle (UAV) imagery remains challenging due to limited contextual cues, a dominance of small-scale objects, and significant scale variation. To address these issues, we propose the Multiscale Dual Residual Network (MDRNet), which comprises three key modules. The Large Kernel Bottleneck Module (LKBM) expands the receptive field through multi-scale convolutions and cross-scale skip connections, capturing fine-grained contextual features. The Dual Feature Fusion Upsampling Module (DFFUM) enhances small object representation by merging shallow and deep features via dual LKBMs and upsampling. The Dual Residual Feature Pyramid Network (DR_FPN) further strengthens multi-scale feature integration by stacking residual-enhanced layers. MDRNet achieves notable performance gains on VisDrone, and SIMD datasets, outperforming a strong baseline by 3.5%, and 0.8% in mAP@50. And further validated the effectiveness and universality of MDRNet in the autonomous driving road scene dataset (RSUD20K), achieving map@50 significant achievements of 74.0%.