Scholarlink

<p>With the development of deep learning technology, object detection of Unmanned Aerial Vehicle (UAV) images has demonstrated great potential across various applications. However, it is challenging to accurately identify small targets quickly and effectively due to multi-scale target variations, dense distribution of tiny targets, and interference from complex background textures. In particular, the loss of high-frequency information during downsampling severely hinders high-precision detection. To address these issues, this paper proposes a Full-scale Frequency-aware and Anisotropic Attention Network (FFAA-Net). First, a Frequency-Aware Feature Extraction Backbone is constructed. In this stage, the C2fSep module is used to refine deeper layers to reduce computational redundancy, and a Spatial-Frequency Dual Stream Aggregation Network (SF-DAN) is innovatively proposed, which leverages Discrete Wavelet Transform (DWT) to losslessly reconstruct high-frequency edge details lost during downsampling in the frequency domain, while synergistically capturing global context through cascaded large-kernel convolutions. Second, a Deeply Interleaved Feature Fusion Neck is designed to facilitate interaction between shallow and deep features by introducing a High-Resolution Priority Feedback mechanism. Simultaneously, an Anisotropic Frequency Attention (AFA) module based on Curvelet Transform is integrated, utilizing anisotropic filters to adaptively suppress texture interference in complex backgrounds. Finally, a Full-Scale Prediction System is established, extending the detection range to the P2 layer to ensure effective recall and precise localization of extremely tiny targets. Extensive experiments demonstrate that FFAA-Net achieves strong performance on two public datasets. Specifically, on VisDrone2019, the model achieves 45.2% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> </InlineEquation> and 27.4% <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(mAP_{50-95}\)</EquationSource> </InlineEquation> with only 4.15M parameters. On the AI-TOD dataset, which contains even smaller targets, the model also demonstrates outstanding generalization capability, significantly outperforming current mainstream detection algorithms.</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

FFAA-Net: a full-scale frequency-aware and anisotropic attention network for UAV object detection

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Zhenghua Zhou,
  • Tianning Zhu,
  • Junchuan Xu,
  • Jianwei Zhao

摘要

With the development of deep learning technology, object detection of Unmanned Aerial Vehicle (UAV) images has demonstrated great potential across various applications. However, it is challenging to accurately identify small targets quickly and effectively due to multi-scale target variations, dense distribution of tiny targets, and interference from complex background textures. In particular, the loss of high-frequency information during downsampling severely hinders high-precision detection. To address these issues, this paper proposes a Full-scale Frequency-aware and Anisotropic Attention Network (FFAA-Net). First, a Frequency-Aware Feature Extraction Backbone is constructed. In this stage, the C2fSep module is used to refine deeper layers to reduce computational redundancy, and a Spatial-Frequency Dual Stream Aggregation Network (SF-DAN) is innovatively proposed, which leverages Discrete Wavelet Transform (DWT) to losslessly reconstruct high-frequency edge details lost during downsampling in the frequency domain, while synergistically capturing global context through cascaded large-kernel convolutions. Second, a Deeply Interleaved Feature Fusion Neck is designed to facilitate interaction between shallow and deep features by introducing a High-Resolution Priority Feedback mechanism. Simultaneously, an Anisotropic Frequency Attention (AFA) module based on Curvelet Transform is integrated, utilizing anisotropic filters to adaptively suppress texture interference in complex backgrounds. Finally, a Full-Scale Prediction System is established, extending the detection range to the P2 layer to ensure effective recall and precise localization of extremely tiny targets. Extensive experiments demonstrate that FFAA-Net achieves strong performance on two public datasets. Specifically, on VisDrone2019, the model achieves 45.2% \(mAP_{50}\) and 27.4% \(mAP_{50-95}\) with only 4.15M parameters. On the AI-TOD dataset, which contains even smaller targets, the model also demonstrates outstanding generalization capability, significantly outperforming current mainstream detection algorithms.