Scholarlink

<p>To address the challenges of insufficient detection accuracy in UAV aerial imagery, this paper proposes EDC-Net, a small object detection method based on an Edge-Aware Dynamic Context Network. First, an Edge-Aware Enhancement Module (EAEM) is designed to compensate for contour degradation in deep features. By embedding gradient cues into intermediate feature learning and combining them with spatial-context information, EAEM enhances the network’s sensitivity to tiny-object boundaries. Second, we propose a Multi-scale Adaptive Context Module (MACM), which coordinates dynamic non-linear activation, statistical self-attention, and multi-receptive-field refinement to suppress background interference, aggregate global context with linear complexity, and strengthen multi-scale feature representation. Finally, a Dynamic Structural Alignment Head (DSAH) is constructed by introducing Dynamic Snake Convolution into the regression branch, enabling adaptive geometric alignment for irregular targets and improving localization accuracy. Across the VisDrone2019 dataset, experimental results indicate that the proposed approach outperforms the baseline YOLO11s, with gains of 4.5% and 3.0% in mAP@50 and mAP@50:95, respectively, and a 29.8% reduction in the total number of parameters. Additionally, generalization tests on the TinyPerson dataset confirm the capability of the presented method in small object detection tasks.</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

EDC-Net for small object detection in UAV imagery using edge-aware dynamic context modeling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Yangyang Gao,
  • Shuxian Liu

摘要

To address the challenges of insufficient detection accuracy in UAV aerial imagery, this paper proposes EDC-Net, a small object detection method based on an Edge-Aware Dynamic Context Network. First, an Edge-Aware Enhancement Module (EAEM) is designed to compensate for contour degradation in deep features. By embedding gradient cues into intermediate feature learning and combining them with spatial-context information, EAEM enhances the network’s sensitivity to tiny-object boundaries. Second, we propose a Multi-scale Adaptive Context Module (MACM), which coordinates dynamic non-linear activation, statistical self-attention, and multi-receptive-field refinement to suppress background interference, aggregate global context with linear complexity, and strengthen multi-scale feature representation. Finally, a Dynamic Structural Alignment Head (DSAH) is constructed by introducing Dynamic Snake Convolution into the regression branch, enabling adaptive geometric alignment for irregular targets and improving localization accuracy. Across the VisDrone2019 dataset, experimental results indicate that the proposed approach outperforms the baseline YOLO11s, with gains of 4.5% and 3.0% in mAP@50 and mAP@50:95, respectively, and a 29.8% reduction in the total number of parameters. Additionally, generalization tests on the TinyPerson dataset confirm the capability of the presented method in small object detection tasks.