Research On Feature Extraction Of Face And Tongue Image Based On Face Landmark Detection In Traditional Chinese Medicine

As one of the four national quintessences,traditional Chinese medicine(TCM)has developed a set of systematic diagnostic methods in the long history.Inspection is important content and key technology in TCM diagnostics.Physicians extrapolate and predict the formation and development of disease in a patient through observing the patient’s body,face,tongue,and expression,and further analytical assessment and treatment.The changes in internal organs are shown in the human face and tongue in TCM theory.Therefore,face and tongue examinations have always been the main basis for inspections and played an extremely important role in TCM.The TCM inspection process relies heavily on the expertise and clinical experience accumulation of the diagnosing physicians,and it is extremely susceptible to interference from the diagnosis environment such as light.The inspection results usually vary from person to person,with poor reproducibility,which greatly limits the development of TCM inspection techniques.With the development of computer vision technology and medical image processing technology,modernization research of TCM diagnosis based on objectivity,quantification and standardizztion has become the trend of TCM development.Based on the feature extraction task of face and tongue image in TCM diagnosis,this paper uses deep learning technology to process and analyze the face and tongue image,and research the image feature extraction task in the automated system of TCM inspection and diagnosis.The main contents of this paper are as follows:(1)The parallel HRNet network is proposed to detect facial landmark points.The proposed method performs independent detection for different parts of the face area,especially the mouth,to improve the detection accuracy of the model.The complexity of the model is also reduced by decreasing the number of layers of different resolutions in the HRNet network.(2)The method based on 3DMM pre-trained model and graph convolutional network is built for single-image 3D face reconstruction and texture optimization.Combining the detection results of face landmark points in this paper,the feature extraction of the face region is realized based on Mask.In this paper,the texture of the face model is optimized based on 3DMM pre-training network and network convolution using the extracted texture features of a human face.Experiments show that compared with the original 3DMM model,the texture of the 3D face model generated by the procesed method has a certain enhancement effect on the visual senses,and it also has a certain improvement in the objective evaluation index.(3)An improved HRNet network is used to segment the tongue image.And the separation of the moss of the segmented tongue image is realized by K-means clustering method.The tongue region is extracted based on the detection results of facial features,and the semantic segmentation of tongue image is proposed based on the improved HRNet network.By comparing several common loss functions to train the improved HRNet network separately,the edges of the tongue region image segmented by the proposed method are sharper.Based on the above-mentioned tongue image segmentation results,the K-means algorithm is used to separate the tongue fur texture,and a good separation result is also obtained.