Research On Application Of Tongue Recognition Model Based On Convolutional Neural Network

At present,most modern tongue clinics use tongue diagnostic instruments as a means of collecting data.Their popularity is severely constrained by factors such as cost,venue,time,etc.It is difficult to meet the needs of people’s "self-service" tongue diagnosis at any time and anywhere.Therefore,based on the tongue image data collected by portable devices such as smart phones,the following research work is carried out based on convolutional neural network for tongue image recognition.Firstly,according to the theory of tongue diagnosis of traditional Chinese medicine,under the guidance of traditional Chinese medicine experts,the attributes and categories of tongue image data are defined.For the problem of multi-attribute labeling for tongue image recognition,the tongue labeling tool is designed and developed,and the tongue image data is developed.Label work,build a tongue coating data set and a tongue data set.Secondly,based on the analysis of classical image segmentation methods such as threshold segmentation,region segmentation and graph segmentation,according to the characteristics of tongue image data,an automatic segmentation method based on GrabCut algorithm is proposed,and the morphology expansion algorithm is applied to the tongue.The data is segmented for hole detection and filling,and further cropped to obtain the final segmentation result.Experiments show that the tongue data after automatic segmentation retains the tongue image feature information completely,which can provide data support for tongue image recognition research.Thirdly,based on the performance analysis of classical convolutional neural networks,a tongue recognition model TonNet is proposed.Based on the AlexNet network structure,the model introduces dense connections in DenseNet,including convolutional layer,Sedense block,pooling layer,global average pooling layer,full connectivity layer and Softmax layer.Based on TonNet’s requirements for tongue and tongue and tongue quality multidirectional comprehensive diagnosis,a Tongue identification model Multi-TonNet containing two outputs of tongue coating and tongue was designed.Finally,TonNet is trained on the tongue coating data,and compared with the training effect of the classical network structure,and the TonNet is extended and applied on the tongue data set;the tongue-like tongue coating 2 attribute data set is constructed.MultiTonNet is trained and tested to verify its effectiveness.Finally,under the Ubuntu 16.04 operating system,using the Python 2.7 language in the Anaconda 3.0 environment,respectively based on TonNet in the tongue coating dataset,based on TonNet in the tongue dataset and based on Multi-TonNet in the tongue coating and tongue quality The tongue image recognition experiment on the attribute data set verifies the validity of the tongue recognition model TonNet and the feasibility of the multi-output tongue recognition model Multi-TonNet.