Research On Tongue Image Classification In Traditional Chinese Medicine Based On Deep Learning

Traditional Chinese Medicine(TCM)comprehensively judges disease syndromes through the analysis of the internal connections of the four diagnostics,and the tongue examination of TCM is an important part of the inspection.In the process of objectification of tongue diagnosis in TCM,there are problems in that the accuracy of tongue image segmentation is not high,the color correction effect of tongue image and the tongue feature recognition are not good.The Django framework is used for server-side development,and Vue,Boot Strap and other frameworks are used for browser-side development to realize the auxiliary platform of TCM tongue image.The work content of the thesis is as follows:1.In view of the low accuracy of traditional model segmentation and the small receptive field of the segmentation method based on deep learning,a pyramid pooling tongue image segmentation algorithm(PPTIS-Net,Pyramid Pooling Tongue Image Segmentation Network)is proposed based on the pyramid scene analysis network.First,the Res Net101 network is used for backbone feature extraction,and then multi-scale feature extraction is performed on tongue image features through pyramid pooling.Finally,the segmented tongue image is restored to the original scale through full connection and deconvolution operations to generate the segmented result.Compare with algorithms such as Mask-RCNN in segmentation pixel accuracy,F1 value and average intersection ratio,and achieve good results in pixel accuracy and average intersection ratio.2.Aiming at the problem of poor tongue image color correction effect and improving the effect of tongue image color constancy,the "subjective and objective two-stage tongue image color correction algorithm" is adopted and the model is improved,and the original five-layer convolutional layer in the objective stage is adopted.The network is modified to the Alex Net,which creatively combines the classic network with the existing algorithm.The improved algorithm is called the two-stage tongue image color correction algorithm(TSTICC-Net,Two-Stage Tongue Image Color Correction Network).In view of the lack of color correction data set,in addition to the use of data enhancement methods to expand the data set,the four dimensions of image contrast,chroma,brightness and sharpness have been expanded.The algorithm first corrects the original image through the objective stage of color correction,using a trained Alex Net to correct the original image for the first time,and then perform subjective stage color correction based on the7) color space to obtain the final the corrected image.In addition,the thesis puts forward the ablation experiments based on the three-channel pixel difference and ratio average evaluation index.The ablation experiments with different configurations of the original algorithm and the improved algorithm show that the improved algorithm is 6.37%better than the original algorithm.3.In view of the problem that the recognition accuracy of tongue marks and tongue crack features in TCM tongue diagnosis is not high,because this recognition is a finegrained recognition scheme,the extrusion expansion tongue feature recognition algorithm based on the network of extruded expansion models(SETFR-Net,Squeeze Expansion Feature Tongue Recognition Network)is proposed.The data set is divided into two parts by using the five-person marker,which is adopted by more than half,and the extremely characteristic tongue and tongue cracks are collected over the network for pre-training.By combining the extruded expansion structure with the backbone network,the algorithm re-calibrates the features of each channel in the process of extracting the features in each calculation,enhances the effective feature weight,suppresses the invalid feature weight,and finally improves the feature recognition results.The algorithm creatively combines the squeeze expansion network algorithm with the tongue feature recognition task,and compares the four indicators of average accuracy,macro accuracy,F1 value and recall rate with network models such as VGG16,and evaluates the characteristics of tongue cracks.As a result,the four indicators achieved excellent results;in the evaluation results of tongue and tooth marks,except for the macro-accuracy indicators,other indicators achieved excellent results.4.Design and implement the TCM tongue auxiliary platform,using B/S architecture,based on the Django framework,using My SQL database and Redis database,the platform has doctor and patient role functions.The doctor side implements diagnosis,tongue processing,medical record processing and auxiliary display,while the patient side implements tongue processing,medical record processing and auxiliary display.