Scholarlink

<p>Due to limited feature information and complex background, feature information loss is the main problem in small-object detection. Existing methods often employ multi-scale feature fusion to enhance performance, but this introduces edge distortion and semantic misalignment, limiting further improvements in detection accuracy. To address these issues, this paper introduces the FC-DETR model based on hierarchical feature fusion and calibration. Specifically, to reduce edge distortion, we adopt Inverse Wavelet Pooling (IWP) instead of traditional upsampling methods, which separates high-frequency edge information for precise reconstruction. To address semantic misalignment and feature redundancy after fusion, the Hierarchical Complementary Calibration Module (HCCM) is proposed to enable interaction and coordinated optimization of features across different levels. In addition, we design a Feature Complement Module (FCM) to integrate small-object information from low-level features into high-level features, enabling multi-level feature learning. Experiments on the VisDrone2019, UAVDT, and SIMD datasets demonstrate FC-DETR’s superior performance, with AP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation>, AP<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(_{75}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>75</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> and AP metrics improvements of 4.8%, 4.2%, and 3.8% over the baseline on VisDrone2019. The code is released at: <a href="https://github.com/zzrhero/FC-DETR">https://github.com/zzrhero/FC-DETR</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

Enhancing small-object detection through hierarchical feature fusion and calibration

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Zhuangrui Zhu,
  • Shaoqing Wang,
  • Weiyan Tang,
  • Zihao Jing,
  • Fuzhen Sun

摘要

Due to limited feature information and complex background, feature information loss is the main problem in small-object detection. Existing methods often employ multi-scale feature fusion to enhance performance, but this introduces edge distortion and semantic misalignment, limiting further improvements in detection accuracy. To address these issues, this paper introduces the FC-DETR model based on hierarchical feature fusion and calibration. Specifically, to reduce edge distortion, we adopt Inverse Wavelet Pooling (IWP) instead of traditional upsampling methods, which separates high-frequency edge information for precise reconstruction. To address semantic misalignment and feature redundancy after fusion, the Hierarchical Complementary Calibration Module (HCCM) is proposed to enable interaction and coordinated optimization of features across different levels. In addition, we design a Feature Complement Module (FCM) to integrate small-object information from low-level features into high-level features, enabling multi-level feature learning. Experiments on the VisDrone2019, UAVDT, and SIMD datasets demonstrate FC-DETR’s superior performance, with AP \(_{50}\) 50 , AP \(_{75}\) 75 and AP metrics improvements of 4.8%, 4.2%, and 3.8% over the baseline on VisDrone2019. The code is released at: https://github.com/zzrhero/FC-DETR.