With the prevalence of Traditional Chinese Medicine (TCM), automation techniques are highly required to support the therapy and save human resources. As the fundamental of the TCM treatment, acupoint detection is attracting research attention in both academic and industrial domains, while current approaches suffer from poor accuracy even with sparse acupoints or require extra equipment. In this study, considering the decision-making knowledge of human experts, an image-based deep learning approach is proposed to detect facial acupoints by localising the centre of acupoints. In the proposed approach, high-resolution networks are selected as the backbone to learn informative facial features with different resolution paths. To fuse the learnt features from the high-resolution network, a resolution, channel, and spatial attention-based fusion module is innovatively proposed to imitate human decision, that is, focusing on the facial features to detect required acupoints. Finally, the heatmap is designed to integrally achieve the acupoint classification and position localisation in a single step. A small-scale real-world dataset is constructed and annotated to evaluate the proposed approach based on the authorised face dataset. The experimental results demonstrate the proposed approach outperforms other baseline models, achieving a 2.4228% normalised mean error. Most importantly, the effectiveness and efficiency of the proposed technical improvements are also confirmed by extensive experiments. The authors believe that the proposed approach can achieve acupoint detection with considerable high performance, and further support TCM automation.

中文翻译：

---

基于图像的高分辨率网络和注意力融合的面部穴位检测方法

随着中医 (TCM) 的普及，自动化技术对支持治疗和节省人力资源提出了更高的要求。作为中医治疗的基础，穴位检测正引起学术界和工业界的研究关注，而目前的方法即使在穴位稀疏或需要额外设备的情况下也存在准确性差的问题。在这项研究中，考虑到人类专家的决策知识，提出了一种基于图像的深度学习方法，通过定位穴位中心来检测面部穴位。在所提出的方法中，选择高分辨率网络作为骨干，以学习具有不同分辨率路径的信息丰富的面部特征。为了融合从高分辨率网络、分辨率、通道、创新性地提出了基于空间注意力的融合模块来模仿人类的决策，即关注面部特征来检测所需的穴位。最后，设计热图，一步完成穴位分类和位置定位。构建并注释了一个小规模的真实世界数据集，以基于授权的人脸数据集评估所提出的方法。实验结果表明，所提出的方法优于其他基线模型，实现了 2.4228% 的归一化平均误差。最重要的是，所提出的技术改进的有效性和效率也得到了大量实验的证实。作者认为，所提出的方法可以实现相当高性能的穴位检测，并进一步支持中医自动化。