Experimental Research On Segmentation And Recognition Of Tongue Image In Traditional Chinese Medicine Based On Deep Learning

Objectives:Using deep learning technology to study the segmentation of traditional Chinese medicine tongue image to remove interference background.A tongue image recognition algorithm is proposed to determine the types of tongue color and coating color,in order to develop a tongue image diagnosis and treatment system to assist Chinese medicine doctors in better diagnosis of diseases.The use of statistical techniques to analyze the quantitative indicators of different tongue colors and coating colors provides an objective basis for the identification of the tongue in traditional Chinese medicine and provides data support for the objectification of tongue diagnosis.Methods:1.Randomly select 232 subjects as research objects,use CASIO-3000 EX digital camera to collect tongue images under standard light source D65,use Photo Shop V 13.0 for preliminary manual segmentation,and divide the processed data set into training sets,Validation set and test set,using data enhancement technology to increase the training set,preliminary exploration of the TCM tongue image segmentation model through U-Net,and propose a new tongue image segmentation network to improve the shortcomings of U-Net.Finally,the accuracy rate,Dice coefficient,m Io U coefficient and misclassification error are used to evaluate the effect of the model.2.Screen the tongue images with normal lip color from the tongue image data set,and use the "Thirteenth Five-Year" version of the "Thirteenth Five-Year" version of the "Diagnosis of Traditional Chinese Medicine" textbook to determine the types and identification standards of tongue color and coating color,using data enhancement technology Increase the training set,use the new tongue image segmentation network to segment the tongue image,construct the traditional Chinese medicine tongue image recognition model through the separation attention network(Res Ne St),and finally use the accuracy,recall,and precision to evaluate the effect of the model.3.Use statistical methods to analyze the tongue image data set.Analyze the color characteristics of the tongue,and refer to the difference between the lips and the tongue to determine the color of the tongue.Learn the color characteristics of the tongue coating to determine the color of the tongue coating,and quantify the classification standard of the tongue coating color.Results:1.For U-Net and the new tongue image segmentation network model,use the TCM tongue image of the test set to evaluate the performance of the model.For each tongue image,we use four evaluation indicators to evaluate the prediction results,and take the average of the test set as the final result of the model test.For the tested data set,the accuracy of U-Net is 0.9659,the Dice coefficient is 0.9707,the m Io U coefficient is 0.9642,and the misclassification error is0.16.The accuracy of the tongue image segmentation network is 0.9760,the Dice coefficient is 0.9809,the m Io U coefficient is 0.9760,and the misclassification error is 0.08.From the analysis of the experimental results,the U-Net segmentation network can more completely segment the tongue body and background of the TCM tongue image.Compared with U-Net,the new tongue image segmentation network can achieve better results.2.For the recognition model,use the data-enhanced test set to evaluate the effect of the model.We use three evaluation indicators to evaluate the predicted results.For the test set,the accuracy rate of moss color recognition was 0.9262,the accuracy rate was 0.9060,and the recall rate was0.9263;the accuracy rate of tongue color recognition was 0.8862,the accuracy rate was 0.8845,and the recall rate was 0.8961.From the analysis of the experimental results,the network model based on deep learning can better identify the color of tongue coating and tongue texture.3.For the quantification of TCM tongue examination standards,pale tongue was the highest in tongue GB value,followed by pale red tongue,purple tongue,and red tongue,and crimson tongue was the lowest(P<0.01).In the tongue quality I value,pale tongue was higher,followed by pale red tongue,red tongue,purple tongue,and crimson tongue was the lowest(P<0.01).The ranking order of the difference in GB value of tongue color and lip color is similar to that of quantized GB value of tongue color,but the degree of difference is greater.In the G value of the moss color,white moss is the highest,followed by thin white moss,thin yellow moss,and yellow moss the lowest(P<0.01).The values of thin white fur are lower than those of white fur(P<0.01).Conclusions:1.The segmentation performance of the new tongue image segmentation network proposed in this study is more prominent than that of U-Net,which can provide technical support for the subsequent objective research of tongue diagnosis.2.The tongue image recognition algorithm framework proposed in this study can better identify the tongue color and the fur color,and is not affected by the camera angle and the shooting distance.It has strong anti-interference ability and adaptability,suggesting that the framework can assist TCM doctors The role of diagnosing disease.3.Deep learning technology can well segment the tongue from the tongue image,and is not affected by the external environment,and has a strong generalization ability.For the quantitative indicators of tongue color and fur color,a single R,G,and B value cannot fully reflect the tongue color and fur color category,and it is necessary to comprehensively analyze all the values in the comprehensive color space.