Scholarlink

<p>To address the limitations of traditional acupoint localization methods—notably high subjectivity and poor consistency—this research proposes RT-HALN, a high-precision lightweight network for hand acupoint localization, designed to automate and standardize the localization process. This model employs a dual-stream parallel fusion Stem architecture (DPF-Stem) to enhance the robustness of acupoint localization in complex backgrounds. Additionally, it incorporates efficient spatial pyramid pooling (Efficient-SPP) to tackle issues related to scale feature segmentation and computational redundancy in acupoint detail information. Furthermore, a progressive feature enhancement module (CBR-Block) is introduced to improve the accuracy of acupoint localization. The experimental results demonstrated that the RT-HALN model achieved an AP95 of 94.74% on the hand acupoint dataset, while reducing model computation and parameter count by 0.024G and 0.444&#xa0;M, respectively. On the cross-pose and cross-scene Hand Keypoints Dataset, both AP95 and AR95 improved by 1.16% and 1.15%, respectively, confirming the real-time capabilities and generalization performance of the RT-HALN model. Additionally, the RT-HALN network is designed for deployment on edge devices to meet real-time detection requirements, achieving a balance between positioning accuracy and detection speed while maintaining a lightweight structure.</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

RT-HALN: a lightweight network for hand acupoint localization using dual-branch collaborative fusion and progressive feature enhancement

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Yi Wu,
  • JianHua Qin,
  • ZhenLun Chen

摘要

To address the limitations of traditional acupoint localization methods—notably high subjectivity and poor consistency—this research proposes RT-HALN, a high-precision lightweight network for hand acupoint localization, designed to automate and standardize the localization process. This model employs a dual-stream parallel fusion Stem architecture (DPF-Stem) to enhance the robustness of acupoint localization in complex backgrounds. Additionally, it incorporates efficient spatial pyramid pooling (Efficient-SPP) to tackle issues related to scale feature segmentation and computational redundancy in acupoint detail information. Furthermore, a progressive feature enhancement module (CBR-Block) is introduced to improve the accuracy of acupoint localization. The experimental results demonstrated that the RT-HALN model achieved an AP95 of 94.74% on the hand acupoint dataset, while reducing model computation and parameter count by 0.024G and 0.444 M, respectively. On the cross-pose and cross-scene Hand Keypoints Dataset, both AP95 and AR95 improved by 1.16% and 1.15%, respectively, confirming the real-time capabilities and generalization performance of the RT-HALN model. Additionally, the RT-HALN network is designed for deployment on edge devices to meet real-time detection requirements, achieving a balance between positioning accuracy and detection speed while maintaining a lightweight structure.