关键词: 迁移学习, 舌体分割, 舌象分析, 深度学习, 中医舌诊

Abstract: The traditional Chinese medicine tongue diagnosis,because of its intuition and easy to be observed,as well as its high clinical value,convenience and practicability,has become one of the important research subjects.At present,the combination of medical image processing technology,artificial intelligence technology and clinical experience of Chinese medicine experts to achieve objectification,quantification and automation of TCM tongue diagnosis is the mainstream of modernization research of TCM tongue diagnosis.In this paper,the key techniques of tongue segmentation and tongue image feature recognition based on migration learning and deep learning are studied.A tongue segmentation method based on region-based single pixel loss function is proposed.It can instruct the training and learning of the model by combining the color correlation and the semantic correlation between neighboring pixels,and the semantic information of target pixel labels.The experiments show that it partly improves the segmentation effect of the model,the MIoU index on the test set reached 96.32%.Then,a classification model of the tongue image geometric features,which combines spatial transformation network and VGG16 model,is proposed to identify and extract the geometric features of tongue image,providing a basis for syndromic inference of tongue image.Considering the orderliness of the geometric features of the data on the two-dimensional plane,the spatial transformation network is used to explicitly learn the spatial invariance in the model.And the convolution part of the VGG16 model is reused,so that the knowledge learned from the tongue segmentation task can be used for parameter transfer learning.Through two sets of comparative experiments,the validity of the spatial transformation network is proved to improve the spatial invariance of the model,and the knowledge of transfer learning is proved to make the model converge faster and more smoothly.At the same time,a classification model of the tongue image texture features,based on the deep texture coding network and VGG16 model,is proposed to recognize and extract the texture features of tongue image,providing a basis for syndromic inference of tongue image.According to the disorder of texture features in two-dimensional plane,a deep texture coding network is used to encode the ordered feature map,obtained by convolution layers,into an orderless texture semantic representation,which can express texture information more effectively.And the deep texture encoding network can enable the whole model to input images of any size,which gets rid of the loss of texture information caused by scaling operations of fixed input size.The validity of the orderless encoding of the deep texture encoding network for texture semantic representation is verified by the comparative analysis of experiments.

Key words: Deep learning, Tongue image analysis, Tongue segmentation, Traditional chinese medicine tongue diagnosis, Transfer learning