A Heatmap Regression-Based HR Net Method for Hand Acupoint Recognition
DOI:
https://doi.org/10.47363/r6e67k62Keywords:
Deep Learning, Key Point Detection, Acupoint RecognitionAbstract
Objective: This study investigates hand acupoint recognition using deep learning, leveraging image processing and deep learning technologies to automatically identify acupoint categories and locations on the hand. This approach aims to enhance the accuracy of acupoint recognition, providing a reference for young traditional Chinese medicine practitioners and advancing the modernization of acupuncture diagnosis and treatment.
Methods: This study developed an HRNet-based hand acupoint recognition method using heatmap regression, applied to the recognition task of six common dorsal hand acupoints, including Shaozhe (SI1) and Zhongzhu (SJ3). Standardization, normalization, and data augmentation techniques were employed during the image preprocessing stage. Model training employed either heatmap regression or coordinate regression strategies. The proposed method was compared with four deep learning models—Simple Baseline, ResNet50, and VGG16—on a self-created dataset of 1,102 hand acupoint images, and model performance was evaluated using multiple metrics including PCK, IoU, OKS, and precision.
Results: Experimental results show that the HRNet model based on heatmap regression performs best across all metrics (PCK: 95.37%, IoU: 80.04%, precision: 94.89%, OKS: 90.07%), outperforming Simple Baseline (PCK: 94.12%, etc.); among coordinate regression models, VGG16 outperforms ResNet50. The overall recognition accuracy of the heatmap regression model was significantly higher than that of the coordinate regression model.
Conclusion: This study validated the feasibility and effectiveness of deep learning models in hand acupoint recognition, particularly the HRNet model based on heatmap regression, which demonstrated significant advantages in terms of accuracy and stability. This method not only provides precise acupoint localization assistance for traditional Chinese medicine practitioners but also offers critical data support for precise diagnosis and treatment in intelligent medical devices such as acupuncture robots.