Research On Application Of TCM Tongue Diagnosis Based On Convolutional Expert Neural Network

Tongue diagnosis is viewed as the key part in TCM observation,as much important information can be acquired concerning human health by the analysis of the human tongue.With the improvement of living standard,people pay increasing attention to their health.As the essence of domestic medicine,tongue diagnosis theory of TCM plays an irreplaceable role in human health diagnosis.However,quantitative mechanism is insufficient in the traditional tongue diagnosis,which is vulnerable to be influenced by external environment and the level of doctors,so the results are always demonstrated to be more subjective.To solve this problem,this paper innovatively proposes a method of automatic tongue image classification using computer vision technology.The main research work and innovation points of this paper are as follows:(1)Design the tongue image data acquisition and calibration system.Firstly,the imaging principle of Intel Real Sense D435 depth camera is studied,then the principle of RGB-D image alignment is analyzed,and the data acquisition software is made combined with the camera.Based on the collected data,the principle of data annotation is studied,and a practical data annotation software is made.Finally,a 32-point 3D tongue image localization model is proposed innovatively.(2)The algorithm of tongue image preprocessing is studied.The first step is to analyze the noise types and denoising indicators.The second step is to study a variety of denoising algorithms and do the corresponding comparative experiments.The performance of various denoising algorithms is evaluated according to the noise evaluation indicators.The third step studies a variety of color correction algorithms,and completes the corresponding comparative experiments.Through the comparative analysis of the experimental results,it is concluded that the grayscale world algorithm based on image entropy constraint can best retain the true information of the tongue image.The median filter denoising algorithm combined with the grayscale world algorithm with entropy constraint is used to preprocess the tongue image,which lays a foundation for the subsequent tongue image segmentation algorithms.(3)Research on localization and segmentation of tongue body.To begin with,the principle of 3D Procrustes analysis algorithm and the principle of tongue 3D coordinate reconstruction algorithm are studied.Then,a three-dimensional reconstruction of tongue is proposed by using the point-distribution model,and the algorithm principle of the point-distribution model is analyzed,and the corresponding experiments are done.Moreover,the algorithm principle of convolutional expert neural network is analyzed,and the convolutional expert neural network is trained by using the samples collected and labeled above,and the training results are analyzed experimentally.Finally,the three-dimensional point distribution model combined with the convolutional expert network constraint landmark model formed by the convolutional expert neural network is used to locate and segment the tongue image landmarks,and the corresponding experiments and optimization analysis are done.(4)Study on tongue image classification and recognition.Firstly,the tongue images were labeled with multiple labels to extract the directional gradient histogram(HOG),and then the multi-classification support vector machine(SVM)was designed to carry out the multi-classification experiment.Finally,a control group was added in the experiment for algorithm performance analysis.K-Nearest Neighbors(KNN)algorithm combined with HOG method and HOG combined with SVM classification algorithm were used for comparative experimental analysis.In comparison,HOG combined with SVM tongue image classification algorithm has better performance.The effectiveness of the proposed method is verified by a systematic tongue image multi-classification and recognition experiment,and the innovative method has a certain application value in engineering.