Scholarlink

<p>Hair follicle detection in complex scalp environments remains challenging due to small target sizes, morphological similarities, and background interference. To address these issues, we propose HFD-NET, a novel real-time detection framework that significantly enhances weak feature representation while maintaining computational efficiency. Unlike existing methods, HFD-NET introduces three key innovations: (1) the CSRFConv module, which dynamically fuses spatial-channel features to suppress background noise and amplify discriminative follicle characteristics; (2) the C3k2_IBC module, optimized for multi-scale small-object detection through inverted bottleneck convolutions; and (3) the ADown module, which minimizes information loss during downsampling by preserving critical edge and texture details. Extensive experiments on the FDU_HairFollicleDataset demonstrate HFD-NET’s superiority, achieving a 67.1% mAP@0.5—outperforming YOLO11n by 5.5%—with only a 0.1M parameter increase and negligible computational overhead. HFD-NET also generalizes effectively to public datasets (e.g., 71.6% mAP@0.5 on HFDC), surpassing Faster R-CNN, SSD, and recent YOLO variants in accuracy-efficiency trade-offs. This work bridges the gap between high-precision detection and real-time applicability, offering a practical solution for clinical hair transplantation and scalp health monitoring.</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

HFD-NET: a real-time deep learning-based visual algorithm with weak feature enhancement for hair follicle detection

  • 期刊论文

  • 发表日期：

    2025年3月12日

查看全文
  • Liyang Zhang,
  • Juntao Tong,
  • Yue Chen,
  • Nian Wang,
  • Weiang Li,
  • Mengxin Chen,
  • Jinghong Tian

摘要

Hair follicle detection in complex scalp environments remains challenging due to small target sizes, morphological similarities, and background interference. To address these issues, we propose HFD-NET, a novel real-time detection framework that significantly enhances weak feature representation while maintaining computational efficiency. Unlike existing methods, HFD-NET introduces three key innovations: (1) the CSRFConv module, which dynamically fuses spatial-channel features to suppress background noise and amplify discriminative follicle characteristics; (2) the C3k2_IBC module, optimized for multi-scale small-object detection through inverted bottleneck convolutions; and (3) the ADown module, which minimizes information loss during downsampling by preserving critical edge and texture details. Extensive experiments on the FDU_HairFollicleDataset demonstrate HFD-NET’s superiority, achieving a 67.1% mAP@0.5—outperforming YOLO11n by 5.5%—with only a 0.1M parameter increase and negligible computational overhead. HFD-NET also generalizes effectively to public datasets (e.g., 71.6% mAP@0.5 on HFDC), surpassing Faster R-CNN, SSD, and recent YOLO variants in accuracy-efficiency trade-offs. This work bridges the gap between high-precision detection and real-time applicability, offering a practical solution for clinical hair transplantation and scalp health monitoring.