Diagnosis Of Tongue Images Based On Convolutional Neural Networks

The author studies the traditional Chinese medicine tongue diagnosis algorithm based on convolutional neural network.The author uses the combination of object detection,semantic segmentation and image recognition convolution neural network to realize the structure and objectification of TCM tongue image diagnosis.Firstly,the author starts to study the theory of TCM tongue diagnosis and treatment,learn the relevant knowledge of tongue diagnosis,master the tongue diagnosis methods,and lay the foundation for the standardization of tongue diagnosis.According to the theory of traditional Chinese medicine,the tongue image is divided into two directions: tongue color and tongue quality.From the perspective of tongue color,the tongue color is classified into cyan tongue,crimson tongue,light red tongue and light white tongue.From the perspective of tongue quality,the tongue quality is divided into five symptoms: cracked fur,awn thorn fur,greasy fur,dry fur and thin fur.Corresponding to tongue color and tongue quality,various pathologies that may occur in the human body can be inferred according to clear evaluation criteria.With this standard,we can use computer to recognize and classify the tongue color and tongue quality of tongue image through modern image processing technology,and draw diagnostic opinions according to the disease nature and diagnostic relationship of tongue diagnosis in traditional Chinese medicine,so as to complete tongue image diagnosis.Secondly,the author completes the specific tongue objective diagnosis scheme design in four steps: tongue positioning,tongue semantic segmentation,tongue feature recognition and finally objective diagnosis.According to the requirements of each step,the relevant feasibility study is carried out,and the detailed software design and improvement scheme are given for each process of tongue image extraction.The tongue positioning scheme based on yolov5 is given.In the part of tongue body positioning,yolov5 is used to accurately locate and label the tongue body in the tongue image sample,segment the tongue body area and remove most of the invalid background;The tongue semantic segmentation scheme based on UNet is determined.The purpose of tongue semantic segmentation is to completely remove all background interference outside the tongue contour.Here,the improved UNet based on RESNET structure and Se LU activation function proposed in the author is used to segment the tongue and extract the tongue accurately;VGGNet is used for tongue recognition and classification by Multi-task learning.The multi task network connected by hard sharing is used to identify the tongue quality,and the multi task network for tongue image feature recognition is established by connecting the tongue quality and tongue color network by soft sharing;In the objective diagnosis stage,based on the human disease diagnosis relationship table summarized from the diagnosis cases of traditional Chinese medicine,the identified tongue color and tongue quality are corresponding to them,the corresponding human symptoms are obtained,and the diagnosis and treatment opinions are given in combination with the symptoms.Finally,the functions of the above steps are connected to build a complete tongue image diagnosis system.Use standard test sets to test the accuracy and effectiveness of each method;Through the comparison of multiple methods,find the improvement scheme and study the improvement of the selected method;The methods are fused to remove the unnecessary data processing process and simplify the overall structure of tongue image diagnosis.After testing,the recall rate of tongue positioning using yolov5 is 99.331%,m AP@0.599.493%;Using the optimized UNet,the accuracy of tongue semantic segmentation is99.298% and m Io U is 95.873%;The recognition accuracy of tongue color in multi task VGGNet tongue feature classification is also significantly higher than that of single task network.