Scholarlink

<p>Object detection in optical remote sensing, encompassing both UAV-based imagery and satellite observations, plays a vital role in the low-altitude economy as well as in broader applications such as urban security, traffic monitoring, and natural resource management. Unlike natural images, remote sensing imagery is characterized by densely distributed small objects, complex backgrounds, and large-scale variations, which pose significant challenges for accurate detection. To address these issues, this paper introduces WCINet, a lightweight yet powerful framework specifically designed for small object detection under complex backgrounds. Specifically, a hybrid backbone, CTMixer, is proposed to capture fine-grained details across both spatial and frequency domains, thereby enhancing sensitivity to weak object signals. To further improve feature fusion, we design a RepCSP module based on re-parameterization, which strengthens representation during training without increasing inference cost. In addition, we develop a cross-frequency fusion structure, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {CF}^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>CF</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>PAN, which leverages frequency-aware filtering to improve intra-class consistency and reduce boundary ambiguity for dense small objects. Extensive experiments on the VisDrone and DIOR datasets further demonstrate the capability of WCINet. Remarkably, WCINet achieves 36.9% mAP on the VisDrone benchmark with only 2.4M parameters and 7.6 GFLOPs, while running at real-time FPS-level speed.</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

Wavelet-attention cross-frequency interaction network for dense small object detection

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Junying Hu,
  • Jia Su

摘要

Object detection in optical remote sensing, encompassing both UAV-based imagery and satellite observations, plays a vital role in the low-altitude economy as well as in broader applications such as urban security, traffic monitoring, and natural resource management. Unlike natural images, remote sensing imagery is characterized by densely distributed small objects, complex backgrounds, and large-scale variations, which pose significant challenges for accurate detection. To address these issues, this paper introduces WCINet, a lightweight yet powerful framework specifically designed for small object detection under complex backgrounds. Specifically, a hybrid backbone, CTMixer, is proposed to capture fine-grained details across both spatial and frequency domains, thereby enhancing sensitivity to weak object signals. To further improve feature fusion, we design a RepCSP module based on re-parameterization, which strengthens representation during training without increasing inference cost. In addition, we develop a cross-frequency fusion structure, \(\hbox {CF}^{2}\) CF 2 PAN, which leverages frequency-aware filtering to improve intra-class consistency and reduce boundary ambiguity for dense small objects. Extensive experiments on the VisDrone and DIOR datasets further demonstrate the capability of WCINet. Remarkably, WCINet achieves 36.9% mAP on the VisDrone benchmark with only 2.4M parameters and 7.6 GFLOPs, while running at real-time FPS-level speed.