Scholarlink

<p>Aerial images suffer from problems such as extremely small target sizes, drastic multi-scale variations, and excessive background interference, which cause current object detectors to inadequately extract small object features, thereby leading to insufficient detection accuracy. To address these issues, we proposed the Cross-Stage Edge Information Fusion Network (CEIFNet). This network utilizes multiple Cross-stage Transformer (CST) Blocks for feature extraction, which are then fed into the Global Edge Information Interaction (GEII) structure for multiple transmissions of shallow edge information. GEII first generates processed shallow edge information through a Multi-scale Edge Information Generator (MEIG), then, to better perform cross-channel feature fusion, Cross-channel Feature Fusion (CFF) is proposed to fuse edge information with features extracted by conventional convolutions. Finally, after feature interaction through an encoder, detection results are output via a dynamic head (DyHead). Experimental results demonstrate that CEIFNet achieves mAP@50 of 43.1% and 95.0% on the VisDrone2019 and HIT-UAV datasets, respectively. Compared to the baseline model, the mAP@50 improves by 10.5% and 2.8% on the two datasets. Visualization results further validate the effectiveness of CEIFNet under conditions of excessive scale variations and severe background interference.</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

Cross-stage edge information fusion network for small object detection in aerial images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月7日

查看全文
  • Jin Xiao,
  • Congjian Li,
  • Haifeng Chen,
  • Hanyun Li,
  • Linsong Xiao

摘要

Aerial images suffer from problems such as extremely small target sizes, drastic multi-scale variations, and excessive background interference, which cause current object detectors to inadequately extract small object features, thereby leading to insufficient detection accuracy. To address these issues, we proposed the Cross-Stage Edge Information Fusion Network (CEIFNet). This network utilizes multiple Cross-stage Transformer (CST) Blocks for feature extraction, which are then fed into the Global Edge Information Interaction (GEII) structure for multiple transmissions of shallow edge information. GEII first generates processed shallow edge information through a Multi-scale Edge Information Generator (MEIG), then, to better perform cross-channel feature fusion, Cross-channel Feature Fusion (CFF) is proposed to fuse edge information with features extracted by conventional convolutions. Finally, after feature interaction through an encoder, detection results are output via a dynamic head (DyHead). Experimental results demonstrate that CEIFNet achieves mAP@50 of 43.1% and 95.0% on the VisDrone2019 and HIT-UAV datasets, respectively. Compared to the baseline model, the mAP@50 improves by 10.5% and 2.8% on the two datasets. Visualization results further validate the effectiveness of CEIFNet under conditions of excessive scale variations and severe background interference.