Research On Tongue Feature Classification Method Based On Deep Transfer Learning

Traditional Chinese Medicine(TCM)pays attention to "Inspection,Listening and smelling examination,Inquiry,Palpation".Tongue diagnosis,as an important part of TCM consultation,plays a key role in clinical diagnosis of TCM.Tongue diagnosis refers to a method for doctors to judge the diagnosis of diseases by observing the changes of the tongue quality and tongue fur morphology of patients.Tongue image refers to the appearance of the tongue.TCM believes that the tongue image is closely related to the health of the human body.With the continuous deepening of modernization of TCM,people have higher requirements for objective and standardized diagnosis of traditional Chinese medicine.However,TCM diagnosis is often restricted by the experience and subjectivity of doctors,and it is difficult to achieve quantitative and objective diagnosis.Therefore,people try to solve the problem of objectivity and standardization of tongue diagnosis of TCM through computer technology.From the computer's point of view,the tongue image feature is the characteristic information of the tongue body image.Generally,the tongue image can be observed in tongue color,tongue coating,cracks,pricks,petechiae and other types.Most traditional tongue image classification methods require a lot of cost to extract manual features,and cannot achieve high accuracy.In recent years,as the concept of deep learning has been proposed,it has been a research hotspot for researchers.It also has been successfully applied in various fields.The combination of deep learning and medical image classification also provides an important solution for the modernization of tongue diagnosis in TCM.Deep learning performs feature learning on images through deep networks,automatically learning features from samples that can accurately identify sample categories,eliminating the need for traditional methods to manually extract features,and greatly improving classification accuracy.However,the training of deep learning often requires a large number of samples,and the particularity of sample acquisition for medical images such as tongue images often has the problem of sample scarcity,which has many obstacles to the training of deep learning models.Therefore,this paper proposes to solve the problem of insufficient sample size through deep transfer learning.Transfer learning refers to training the network parameters on the pre-trained network model with new samples to obtain higher accuracy on new tasks.This can make up for the problem that the model cannot effectively converge due to insufficient sample size,and save the computational cost of retraining on a new neural network through transfer learning,which greatly improves the speed of training.This paper takes three different tongue image features as examples and conducts research through deep transfer learning on neural networks of different depths and analyzes the influence of network depth on classification results.The main work of this article is as follows:1.Through cooperation with Shanghai University of Traditional Chinese Medicine,a total of 2245 cases of tongue image data were collected and collated,and Chinese medicine experts calibrated the data to construct three sets of tongue image datasets with different feature labels.2.Preprocessing of tongue picture.The cascade classifier is used to localize and segment the original tongue image data,and the processed image data is used as the input of the deep network.3.Based on three different neural network structures(Res Net,Goog Le Net,and VGGNet),this paper performs classification training on three different tongue image features to build a tongue image feature classification model.The accuracy of the model is used to compare the influence of different network structures and different network depths on the classification of tongue features.