Study On Local Feature Extraction And Auxiliary Diagnosis Of Tongue Image

Generally speaking,traditional Chinese medicine(TCM)analyzes and determines the cause of disease by combining the four methods of diagnostic(observation,auscultation and olfaction,interrogation,pulse feeling and palpation).The tongue is closely related to the five internal organs(heart,liver,spleen,lungs and kidneys)and the bottom of one’s heart.Some physiological changes in the human body can be mapped out by the tongue image.Therefore,tongue diagnosis as an important branch of traditional Chinese medicine inspection has been attached great importance to medicine.With the rapid development of digital image technology,the research on the objectification of tongue diagnosis in TCM has also made big progress.Current studies on the objectification of tongue diagnosis are mostly focused on the whole tongue image,which can’t reflect the relationship between different regions of tongue images.On the other hand,the tongue data set obtained through often small,and limited categories of the diseases can be judged.The purpose of this paper is to study the region of interest in tongue image,find out the characteristics with positive guiding significance and give pathological analysis results.In this paper,we extract and optimize various features based on the region of interest in tongue image,and then build a classification tree which can give a variety of disease diagnosis.In addition to manual feature extraction,we also use the improved PCANet network for feature extraction and then complete the multi-classification operation.The database of this subject was collected in cooperation with Guangdong Hospital of Traditional Chinese Medicine,and the sample types are the judgment results given by the doctor.Because the number of different diseases is quite different,in order to ensure the reliability of the results,we only study the categories with relatively large sample size.Some important parts of the tongue image,such as the tip of the tongue,is segmented from the complete tongue image to explore the areas of interest to us.In this project,we quantify and optimize the features of the tongue blocks in the region of interest.In view of the characteristics of tongue image blocks,we extract three characteristics including color,texture and red thorn.In order to remove redundant features,we choose some useful features through some feature selection methods.The experimental results show that our feature extraction based on region of interest has achieved good results in accuracy.The tongue image block of region of interest with the positive guiding significance for disease diagnosis.In this study,the correlation between health and disease was analyzed by using the previous feature selection results,including the binary classification of health and each disease,and the experiment of constructing classification tree from top to bottom and from bottom to top.In addition,in order to explore the most suitable classifier for this data set,we use a variety of classifiers to carry out comparative experiments,and ultimately choose the best classifier to get the classification model.Finally,we build a disease classification tree based on the region of interest,some diseases such as diabetes have achieved good results.Although the manual feature has some positive significance,the tongue image block information extraction is not comprehensive enough and may contain some unknown information.Therefore,we use the PCANet feature cascade network to convolute the tongue image block of the region of interest,and then use SVM classifier to carry out multi-classification experiments.In order to prove the effectiveness of the PCANet feature cascade method,this paper compares the PCANet feature cascade extraction with the manual extraction of features(color,texture,red thorn)and some depth convolution learning methods.The experimental results show that the sensitivity of the proposed method is very high,and good results can be obtained in some diseases.