Research On Tongue Image Segmentation Method And Tongue Image Collection System Based On Generative Adversarial Networks

With the rapid development of artificial intelligence-related fields,the new model of smart medical care has become a hot topic in the field of cross-research.The rich veins of the tongue are connected with the internal organs of the human body,which can timely reflect the internal organs and the deficit of qi and blood in the human body.The tongue image is one of the important visual external representations of human diseases and health conditions.However,in clinic,doctors are greatly influenced by the environment,which may cause inaccuracy in diagnosis.Therefore,using modern deep learning technology to automatically segment the tongue body from the tongue image is a very valuable work,and the tongue image segmentation is the precondition and important task of tongue image intelligent recognition.In recent years,some researchers have studied the method of tongue image segmentation,and achieved certain research results,but the existing methods still have some shortcomings.This paper analyzes and studies the structure and characteristics of deep convolutional neural network and adversarial neural network,combines the complex characteristics of human tongue image feature extraction,and integrates convolutional neural network and adversarial neural network methods.In the process of tongue image segmentation,when convolution and pooling downsampling,part of the spatial relationship between pixels will be lost to a certain extent,and there are problems such as blurred or inaccurate edges of the tongue image segmentation result.This paper proposes a VNet-GAN segmentation method.-Adversarial network structure model,using generative adversarial network in tongue image segmentation.On the basis of the UNet segmentation network,the network structure is improved,the pooling layer is omitted,a 4*4 convolution kernel with a stride of 2 is used,a discriminator is added to automatically judge the segmentation results,and a VNet-GAN segmentation-confrontation is constructed.network model.The experimental results show that the application of generative adversarial network to tongue segmentation can better solve the problem that the edge of tongue segmentation is not accurate enough for ordinary networks,and the segmentation effect is better,and the segmentation accuracy reaches97.99%.Although VNet-GAN segmentation-adversarial network can effectively solve the problems of edge blur and accuracy in segmentation,but in the process of learning and training,there is a certain fluctuation in the convergence of the loss function,the robustness is not high,and there are noise holes in the segmentation of tongue images.The phenomenon.To this end,this paper proposes a new WNet-GAN cascade network model by adding a cascade mode to make the segmentation from coarse segmentation to fine segmentation,and at the same time adopting a feature cross fusion mechanism to protect the overall efficiency of the network.The experimental results show that the learning segmentation training converges quickly,the loss function fluctuates less,the segmentation robustness and visual effect are better,and the segmentation accuracy reaches 98.23%.Finally,according to the training model,a tongue image collection and recognition system is designed to realize tongue image collection,image preprocessing,intelligent tongue image segmentation,and data storage,which provides strong support for the subsequent realization of a smart Chinese medicine system.