Scholarlink

<p>Addressing the limitations of subjective experience in traditional Chinese medicine back acupoint localization and the insufficient accuracy of automated identification, this paper introduces a lightweight multi-network fusion model. We construct a dual-backbone architecture featuring MobileNetV2-HRNet. MobileNetV2 is employed for efficient feature extraction, followed by HRNet’s parallel multi-branch structure to achieve multi-resolution feature fusion, integrating both local details and global structural information. Furthermore, a channel-spatial dual-dimension attention mechanism dynamically focuses on crucial regions. We also design an anatomical constraint hybrid loss function. The model utilizes heatmap regression to pinpoint acupoint coordinates, ultimately achieving precise localization of 13 back acupoints. Our method achieves an average detection accuracy of 92.3% on our self-built back acupoint dataset, demonstrating a 24.6% improvement over traditional image processing techniques. With a parameter size of only 4.2&#xa0;M and a single-frame inference time of under 15ms, this model holds significant application potential in health monitoring and related fields.</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

A method for detecting key points of back acupoints based on deep learning

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Chengjun Tian,
  • Guangqiang Song,
  • Yang Li

摘要

Addressing the limitations of subjective experience in traditional Chinese medicine back acupoint localization and the insufficient accuracy of automated identification, this paper introduces a lightweight multi-network fusion model. We construct a dual-backbone architecture featuring MobileNetV2-HRNet. MobileNetV2 is employed for efficient feature extraction, followed by HRNet’s parallel multi-branch structure to achieve multi-resolution feature fusion, integrating both local details and global structural information. Furthermore, a channel-spatial dual-dimension attention mechanism dynamically focuses on crucial regions. We also design an anatomical constraint hybrid loss function. The model utilizes heatmap regression to pinpoint acupoint coordinates, ultimately achieving precise localization of 13 back acupoints. Our method achieves an average detection accuracy of 92.3% on our self-built back acupoint dataset, demonstrating a 24.6% improvement over traditional image processing techniques. With a parameter size of only 4.2 M and a single-frame inference time of under 15ms, this model holds significant application potential in health monitoring and related fields.