Research On Physique Classification Of Infrared Images Of Traditional Chinese Medicine Based On Deep Learning

As the big data era develops,all walks of life present the trend of digitization and informatization.The development of medical big data has attracted more and more attention due to the high correlation between medical care and people’s livelihood.However,medical data often restricts the development of informatization because of its many difficult-to-solve characteristics,including incompleteness,privacy,and polymorphism.And because medical data is obtained in actual clinical work,patients often go to the doctor after getting sick,so the proportion of data obtained by hospitals for various diseases must be related to the incidence of corresponding diseases.This will lead to unbalanced situation in medical data,and will greatly affect the diagnosis and medical analysis in big medical data.Even for traditional Chinese medicine(TCM),this influence still exists.Although TCM only paid attention to seeing,smelling and asking questions a long time ago,now TCM has long introduced advanced medical equipment as an extra means to diagnose diseases.However,how to match these medical data with the unique symptoms of Chinese medicine and obtain a fast and accurate diagnosis method has become a major problem that restricts the development of Chinese medicine in the new era.In order to judge the physique of TCM more accurately through medical data,this thesis conducted a research on the physical classification of infrared images of TCM.First,this thesis processed the TCM infrared data set at multiple levels based on a complex network data processing method,then designed an improved deployment method of the attention module,after that studied the effectiveness of many other improved methods,added additional features of the sample to the model to participate in training at last.The main research work of this thesis is as follows:(1)Through the infrared image of TCM and the image processing method based on the complex network,this thesis calculated the degree matrix under different thresholds,studied the difference between the greyscale images using different thresholds,and obtained two new image data set to describe the texture of the original image.We designed a model structure suitable for the new data set,and got the data set with better prediction effect.Compared with the original data set,this data set achieved better results on a variety of neural networks.(2)In this thesis,different types of attention mechanisms were added to the neural network and an improved deployment method of attention modules is designed.This method was suitable for a variety of sub-models and the experimental results of most sub-models have been improved;(3)This thesis studied the validity of the gender division in data set,added the patient’s age and gender characteristics to the feature vector processed by the model.And this thesis studied the better value of the proportion of these two features in the feature vector obtained at the end of the model.Finally the accuracy of the model was improved using the better proportion.