Abstract
The conventional method to locate acupuncture points (acupoints) on the human body requires practitioners to have rich experience and skillful performance, leading to high learning costs. Visual positioning technology based on image registration can lower this technical difficulty, allowing more people to benefit from massage therapy. However, existing algorithms often suffer from unstable matching, inaccurate registration, and poor results under local deformation or occlusion. This paper studies a novel visual positioning algorithm for acupoints based on image registration. First, an Image Acupoints Positioning (IAP) algorithm is proposed based on a Convolutional Neural Network (CNN). This algorithm combines prior information with 3D CNN features to maintain high accuracy under unfavorable conditions like noise or illumination changes. Second, a multi-scale parallel Fully Convolutional Network (FCN) is constructed to improve non-rigid registration performance through spatial pyramids, adaptive attention, and direction perception. Experimental results verify the model's validity and precision.
Introduction
Locating massage acupoints is traditionally a "high-barrier" skill requiring years of experience. While automated massage robots and AR-guided therapy are emerging, they struggle with a fundamental biological reality: human bodies are non-rigid. Skin and muscle tissue deform, stretch, and shift depending on posture or pressure, making static anatomical maps unreliable.
Existing visual positioning algorithms often fail because they cannot account for these "squishy" deformations. Furthermore, environmental factors like poor lighting, image noise, or partial occlusion (e.g., a hand or clothing blocking the view) make standard image registration—the process of overlaying a digital map of acupoints onto a real-time photo—unreliable. This research aims to bridge the gap between ideal anatomical maps and the messy reality of diverse human bodies.
Methodology
CNN-based Image Acupoints Positioning (IAP)
The first stage of the proposed system utilizes a CNN-based IAP algorithm. Instead of analyzing the entire body at once, the model employs a patch-based prediction strategy. It crops small local areas around suspected acupoints to reduce background noise and focus the network's attention. To handle varying body types, the researchers implemented a multi-density sampling strategy, training the network on patches at different distances from the target to improve its ability to generalize across different patients.
Multi-scale Parallel FCN for Non-rigid Registration
To handle non-rigid deformations like skin stretching, the authors developed a specialized Fully Convolutional Network (FCN). This network is designed to predict a dense deformation field that maps anatomical landmarks to the patient's current physical state.
Key innovations in this architecture include:
- Spatial Pyramid of Dilated Convolutions: This allows the network to maintain a massive "receptive field," seeing both fine details and broad anatomical context without losing image resolution.
- Adaptive Channel Attention: A trainable "soft threshold" helps the model ignore irrelevant features and focus on the most important anatomical landmarks, specifically weighting minor deformation features.
- Direction Perception Layer: Standard fully-connected layers are replaced with 1x1x1 convolutions to predict pixel offsets in 3D space (X, Y, Z), resulting in a dense deformation field.
Results and Discussion
The proposed method represents a significant advancement over traditional "hand-crafted" feature algorithms. In comparative testing, the system achieved an average positioning error of just 1.27 mm. This significantly outperformed traditional methods such as SIFT (8.60 mm) and ORB (4.14 mm), as well as standard deep learning models like a basic FCN (2.86 mm).
The model also demonstrated high reliability, maintaining a detection rate over 90% even under unfavorable conditions like noise and illumination changes. Optimization experiments revealed that a "trade-off weight" of 0.65 for the loss functions provided the best balance between global positioning accuracy and local precision.
Conclusion
By treating acupoint localization as a non-rigid registration task rather than a simple object detection problem, this algorithm successfully accounts for the biological elasticity of human tissue. While some anatomical regions remain more challenging—for instance, "Acupoint No. 8" showed higher accuracy than "Acupoint No. 4"—the overall system provides a robust foundation for automated massage and guided therapy. Future work will focus on expanding the dataset to include more diverse body types and skin tones to further improve the model's universal applicability.
Summary updated