Diagnosis Of Tongue Image Based On Deep Learning In Traditional Chinese Medicine

Tongue diagnosis of TCM is an emerging direction of medical image recognition and has a wide range of applications in the field of computer vision.The early tongue diagnosis algorithm is to extract the features of the pre-processed tongue images,then to classify the features and obtain the diagnostic information of the original image.The disadvantage of this kind of algorithm is that the classification effect depends entirely on the design of the feature.Because the human-designed feature can not cover all the examples,it can not achieve good results in the actual test.In recent years,the deep learning algorithm can well improve the limitations of the traditional algorithms.By using a large number of annotated pictures collected in real environment,the algorithm can learn the feature extraction method autonomously,and the trained model is more practical.Therefore,based on the dataset provided by the Chinese Academy of Medical Sciences,this paper presents a tongue image diagnosis algorithm based on convolutional neural network.The specific work is as follows:(1)Tongue image diagnosis algorithm based on image preprocessing.According to the characteristics of four classification tasks of tongue color,tongue shape,fur character and fur color,different pretreatment algorithms are designed from the aspects of color feature,texture feature and local feature,due to the limitation of dataset.The experimental results show that the classification accuracy of the proposed algorithm improves by about 1%.(2)Tongue diagnosis algorithm based on network structure and loss function optimization.In order to meet the requirement of real-time system identification,the image preprocessing is removed,and the accuracy is improved through the optimization of network structure and loss function.This paper aims at different classification tasks,using different network pruning methods to modify the network structure,making the resulting model size at Alex Net 1/10,1/4 of Goog Le Net.In addition,in view of the fact that the difference intraclass is greater than between classes,the network loss function is improved and Center Loss is added as the auxiliary loss function.After the final optimization of the structure and loss function,the accuracy of the four classification tasks reached more than 90%.(3)In this paper,the network structure is compressed and the tongue diagnostic algorithm is encapsulated into an end-to-end real-time diagnosis system,which realizes multi-user online concurrent diagnosis.