Research On Machine Learning Model Of Tongue Characteristics In Traditional Chinese Medicine

According to the four diagnostic methods in TCM,TCM doctors can determine the physiological and pathological status of qi,blood,yin and yang of human body.The four diagnosis of TCM contains looking,smelling,asking and cutting,meanwhile,tongue diagnosis is an important part of looking diagnosis and the important basis of diagnosing diseases.As an important content of tongue diagnosis,tongue body can reflect the deficiency and reality of yin and yang of human body,tongue body is diagnosed by TCM doctors through visual observation and TCM doctors give the diagnosis based on their treatment experience.The level of doctors' diagnosis and treatment skills and external environmental conditions directly affect the diagnosis results,thus,the judgement of tongue body exists the problems of lacking of objectification and low repeatability.From the perspective of judging tongue body in TCM,in this thesis,we introduce image processing technology to realize the qualitative and quantitative research of tongue body image for increasing objectification.Three core contents of crack tongue recognition,tooth-marked tongue recognition and tongue color recognition are studied.C# is used to develop the server and WPF is used to realize the TCM tongue diagnosis and treatment recommendation system platform.The contents of this thesis are as follows:1.Faster R-CNN method is used to identify the cracked tongue.In view of the shortage of tongue image data samples,the data set is expanded by the method of data enhancement.We use the Labelimg tool to process the crack tongue data set,construct the network feature extractor to extract the target crack features,built the region proposal network to generate the crack region proposal box,and identify the crack tongue by the classification and regression of region proposal box.Experiments are conducted by ZF,VGG16 and ResNet101 network extractors and regional proposal networks,their accuracy is 69.61%,77.40% and 74.58%.2.We use convolution neural network and multi-instance learning method to carry out tooth-marked tongue recognition.Specifically,we use two-dimensional gamma function to correct the brightness of tongue image data sample,and adopt color reduction and convex hull detection to generate the suspected tooth-marked area,then the suspected tooth-marked area as the input of convolution neural network and extract tooth-marked feature,finally adopt multi-instance classifier to carry out tooth-marked tongue recognition.Experiments are conducted by ZFNet,Mi-SVM+ZFNet,LMMK+ZFNet,LSK+ZFNet,VGG16,Mi-SVM+VGG16,LMMK+VGG16,LSK+VGG16,their accuracy is 62.52%?65.46%?64.58%?68.70%?69.81%?73.53%?70.21% and 76.30%.3.The improved homomorphic filter is used to preprocess the light in the recognition of tongue constitution color.The fuzzy c-means clustering is used to separate tongue constitution color and tongue coating and extract tongue color features.In RGB,HSV,Lab color space models,KNN,SVM,CNN,Bayes are used for qualitative color recognition.The average accuracy is 57.84%,81.36%,67.37% and 65.26%,respectively.4.This thesis uses visual studio development tools and MySQL database to design and implement the recommendation platform of TCM tongue body diagnosis and treatment,in which we complete the functions of tongue information input,tongue feature mark,and recommendation of TCM medicine treatment plans.It can be realized to recommend TCM diagnosis and treatment plan for patients through the characteristics of tongue.