Facial Acupoints Location using Transfer Learning on Deep Residual Network

In the field of traditional Chinese medicine, whether it’s acupuncture, moxibustion or massage health care, are inseparable from acupoints location. In this paper, we combine traditional Chinese medicine with artificial intelligence, proposing a facial acupoints location method on deep residual network. We also create a facial acupoints dataset containing 1040 face images and 9 acupoints. Based on transfer learning, we first train a facial landmarks location network using a facial landmarks dataset. Then, we transfer the learned low-level features to facial acupoints location network, and use an adaptive segmented Wingloss function to reducing the location error. Our experiments indicate that transfer learning and Wingloss function are effective in improving the accuracy of facial acupoints location. The methods used by other researchers often have errors caused by intermediate variables, and most of their studies have no scientific location accuracy index. We have improved these problems, and the normalized average error of the facial acupoints location is 2.5% in our research.