Research On Tongue Multi-feature Recognition Method Based On Deep Learning

Traditional Chinese Medicine(TCM)image recognition is an interdisciplinary research field that involves the integration of TCM imaging,mathematical modeling,artificial intelligence,and other technologies.TCM tongue diagnosis is based on doctors observing the shape and features of the tongue with the naked eye for diagnosis.However,this method is easily affected by environmental factors and subjective experience of the doctors,making it difficult to achieve standardized and objective diagnosis.With the development of computer technology,people have attempted to use it to solve this problem.Among them,deep learning methods represented by convolutional neural networks have the advantage of directly implicit automatic feature learning from medical images,so they have gradually been applied in traditional Chinese tongue analysis tasks and have achieved certain results.At present,tongue image diagnosis based on Deep Learning is mainly divided into two parts:tongue image recognition and tongue image classification.This article addresses three issues related to how to use deep learning to quickly and effectively distinguish tongue colors,how to improve the effectiveness of deep learning to accurately distinguish tongue shapes,and how to expand the functionality of deep learning to directly detect abnormal features in tongue images.To address these issues,a multi-feature model based on deep learning is proposed,which effectively completes multiple types of tasks.The main research work of this paper is as follows:(1)To address the objectivity issue in the recognition of "color" and "shape" in tongue images,this paper proposes the use of the TDNet network.In this paper,the difficulties of two feature classifications are first introduced.Then,a method of enhancing contrast is proposed to preprocess tongue images for better tongue shape classification,and a color space conversion method is used to preprocess tongue images for better tongue color classification.Finally,the TDNet network is proposed to simultaneously complete both classification tasks.Experimental results such as confusion matrices,ROC curves,and AUC indices show that the fusion model proposed in this paper can better complete both classification tasks than a single model.(2)This article proposes the YOLOX-s model based on attention mechanism to complete the task of detecting three pathological tongue features in tongue images,including Teeth-printed tongue,Wrinkled tongue,and Greasy-fur tongue.Firstly,data augmentation was performed on the original dataset.Then,the fusion method of SENet attention module and YOLOX-s model was introduced,followed by an introduction to the improved loss function.Experimental results show that the proposed fusion model significantly improves the overall recognition effect of the three features.The overall mAP value is 91.95%,with an AP value of 0.85 for tooth-marked tongue,0.96 for cracked tongue,and 0.95 for greasy tongue.