Research On Tongue Cleft Extraction Technology Based On Deep Learning

Tongue cleft is a reliable basis for tongue diagnosis in Traditional Chinese Medicine(TCM).This paper is oriented towards mobile application background.There is a big difference in the quality of pictures collected by mobile phones,and traditional segmentation methods are difficult to obtain good results.In this paper,deep learning and transfer learning are used to segment the cleft accurately,and then the fissured tongue are classified according to the symptoms.This is of great significance for the realization of intelligent medical treatment and digitalization of Traditional Chinese medicine.The main work and innovation of the paper are as follows:(1)We propose a tongue cleft segmentation method based on transfer learning and improved U-Net structure.Firstly,we use data expansion methods such as horizontal flipping,random clipping and normalization to alleviate the problem of small samples.Secondly,the road crack training model is transferred to tongue cleft by transfer learning method to improve training efficiency and accuracy.Finally,we add SE module to U-Net structure to improve the accuracy of segmentation,and use Focal Loss function as the loss function to make the network pay more attention to the segmentation of small targets.Experiments show that the proposed method is more robust and can deal with image quality problems such as background,illumination and texture.Its MPA can reach 71.06%,MIo U can reach 67.35%,which has better visual performance.(2)We proposed a classification method for fissured tongue based on Inception-V4+FRN structure.The Inception-V4 network combines the high-level depth of the residual network with the faster training speed.FRN layer can standardize the data,so as to accelerate the network convergence and improve the model accuracy.The experimental results show that the accuracy of the classification of the Inception-V4+FRN network can reach 87.5%.(3)In order to realize the deployment of the segmentation and classification model in mobile phones,8-bit quantization is adopted to transform float32 data into uint8 data.In order to avoid the problem of large sampling dynamic range,the Moving Average Min Max method is used for data calibration.During quantization,asymmetric quantization algorithm is used to retain data information more completely.The quantization model greatly improves the computation speed and reduces the storage space and memory consumption while slightly sacrificing the accuracy.The tongue cleft segmentation and classification platform was developed based on Android system.The program has the functions of registration,login,logout and tongue cleft segmentation and classification.Users can complete tongue cleft segmentation and diagnosis with a mobile phone.