Scholarlink

<p>Head detection in dense scenes remains challenging due to the small scale of head targets and severe occlusions, which often cause feature loss and reduced detection accuracy. DETR leverages the Transformer architecture to establish robust long-range dependency modeling and achieves end-to-end training. Nevertheless, existing approaches often lack specific feature enhancement tailored for small-scale targets. To address these issues, this paper proposes HFE-DETR. This method improves the representation capability for small objects by implementing enhancement and fusion strategies on high-resolution feature maps. We design a Space-to-Depth Convolution enhanced ResNet-18 backbone to better preserve fine-grained spatial information during downsampling. Moreover, we develop a Multi-channel High-resolution Feature Fusion (MHFF) module to strengthen shallow high-resolution features, and integrate a Dynamic Channel and Position Attention Mechanism (DyCPAM) to adaptively highlight informative semantic and spatial cues. In addition, we employ a learnable weighted multi-scale feature fusion strategy to improve feature interaction efficiency. Experimental results on the SCUT-HEAD small-scale head detection benchmark demonstrate that HFE-DETR outperforms the baseline RT-DETR model, sustaining a high-throughput speed of 164 FPS with only 25.8M parameters and effectively validating its suitability for high-performance computing environments. The source code is available at <a href="https://github.com/TomoriNaoA/HFE-DETR">https://github.com/TomoriNaoA/HFE-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

HFE-DETR: a high-resolution feature enhancement framework for dense head detection

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Jun Li,
  • Haochuan Zhang,
  • Heran Wang,
  • Junnan Jiang

摘要

Head detection in dense scenes remains challenging due to the small scale of head targets and severe occlusions, which often cause feature loss and reduced detection accuracy. DETR leverages the Transformer architecture to establish robust long-range dependency modeling and achieves end-to-end training. Nevertheless, existing approaches often lack specific feature enhancement tailored for small-scale targets. To address these issues, this paper proposes HFE-DETR. This method improves the representation capability for small objects by implementing enhancement and fusion strategies on high-resolution feature maps. We design a Space-to-Depth Convolution enhanced ResNet-18 backbone to better preserve fine-grained spatial information during downsampling. Moreover, we develop a Multi-channel High-resolution Feature Fusion (MHFF) module to strengthen shallow high-resolution features, and integrate a Dynamic Channel and Position Attention Mechanism (DyCPAM) to adaptively highlight informative semantic and spatial cues. In addition, we employ a learnable weighted multi-scale feature fusion strategy to improve feature interaction efficiency. Experimental results on the SCUT-HEAD small-scale head detection benchmark demonstrate that HFE-DETR outperforms the baseline RT-DETR model, sustaining a high-throughput speed of 164 FPS with only 25.8M parameters and effectively validating its suitability for high-performance computing environments. The source code is available at https://github.com/TomoriNaoA/HFE-DETR.