Scholarlink

<p>Deep learning has significantly advanced object detection in general scenarios; however, aerial image object detection faces unique challenges, such as scale variations and uneven spatial distributions, that limit the effectiveness of generic object detection models. Addressing these challenges, this paper presents a multi-path aggregation-based detection network (MPA-Det), specifically designed for aerial visual computing. MPA-Det enhances feature learning through three levels of aggregation: multi-path feature aggregation (MPFA), multi-scale path aggregation (MSPA), and multi-detection head cascade (MDHC) aggregation. The MPFA module employs parallel convolutional blocks with identical topologies, improving feature extraction efficiency and mitigating issues associated with uneven object distribution. To address scale variability, MSPA introduces side connections from lower to upper layers, enhancing the feature pyramid network by adding high-resolution bottom-layer features to high-level features, which improves object localization across pyramid layers. Additionally, MDHC progressively raises the Intersection-over-Union (IoU) threshold across multiple cascaded detection paths, enhancing detection accuracy. Empirical evaluations on the VisDrone and Airbus Ship datasets demonstrate the effectiveness of MPA-Det, showing significant improvements in detection performance over existing methods. Our code is available: <a href="https://github.com/Hiyuur/MPA-Det">https://github.com/Hiyuur/MPA-Det</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

MPA-Det: multi-path aggregation-based object detection framework for aerial visual computing

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Yaolin Lei,
  • Kai Jin,
  • Zifeng Qiu,
  • Yiming Sun,
  • Huihui Bai,
  • Wenzhi He

摘要

Deep learning has significantly advanced object detection in general scenarios; however, aerial image object detection faces unique challenges, such as scale variations and uneven spatial distributions, that limit the effectiveness of generic object detection models. Addressing these challenges, this paper presents a multi-path aggregation-based detection network (MPA-Det), specifically designed for aerial visual computing. MPA-Det enhances feature learning through three levels of aggregation: multi-path feature aggregation (MPFA), multi-scale path aggregation (MSPA), and multi-detection head cascade (MDHC) aggregation. The MPFA module employs parallel convolutional blocks with identical topologies, improving feature extraction efficiency and mitigating issues associated with uneven object distribution. To address scale variability, MSPA introduces side connections from lower to upper layers, enhancing the feature pyramid network by adding high-resolution bottom-layer features to high-level features, which improves object localization across pyramid layers. Additionally, MDHC progressively raises the Intersection-over-Union (IoU) threshold across multiple cascaded detection paths, enhancing detection accuracy. Empirical evaluations on the VisDrone and Airbus Ship datasets demonstrate the effectiveness of MPA-Det, showing significant improvements in detection performance over existing methods. Our code is available: https://github.com/Hiyuur/MPA-Det.