Tongue Image Segmentation With Deep Convolution Neural Network And Conditional Random Field

Tongue diagnosis plays an important position in Traditional Chinese Medicine(TCM).It is one of the important means to understand the state of the human body.Doctors can judge the physical condition of patients by observing the color and shape of the tongue texture and coating.The tongue diagnosis assistant system in TCM can objectively and quantitatively evaluate the patient’s physical condition,which is an important content of the modernization of the tongue diagnosis of Traditional Chinese Medicine.Since the collected image contains many interferences,it is very important to accurately segment the tongue from the image.It is a key step in the tongue diagnosis assistant system and an important prerequisite for subsequent judgment of the patient’s health status.In this thesis,a tongue image segmentation algorithm based on deep convolutional neural network and conditional random field is proposed to improve the shortcomings of the original methods and the emerging semantic segmentation network in tongue image segmentation.Experiments were carried out on two tongue image data sets.The research of this paper mainly includes the following three aspects:1.An algorithm for detecting the candidate region of tongue body is proposed.In order to solve the problem that the traditional tongue image segmentation method and convolutional neural network have limitations on the size of the input image,the maximum stable extreme value region algorithm is used to select the candidate region of the tongue body part in the original image.Segmenting the tongue part in the candidate area enables the original high-resolution image to be fully utilized,and avoids the loss of image information caused by the downsampling operation.2.A segmentation algorithm based on dense atrous convolutional neural network is proposed.In order to solve the problems of gradient disappearance and network degradation in existing convolutional neural networks,and the problem of excessive feature maps in the application of existing networks in tongue image segmentation tasks,the neural network was improved to propose the tongue image segmentation network in this paper.The dense connected network structure can alleviate the gradient disappearance and degradation problems of the network through feature reuse and multi-channel transmission of information.By using the atrous convolution algorithm and atrous convolution connection criterion,the network can expand the feature map without increasing the parameters,so as to improve the segmentation accuracy.Through the hole convolution multi-scale feature extraction module,the multi-scale features of the image are extracted and merged under multiple receptive fields.The network can automatically extract the features in the image,effectively avoid the limitations of manually designing the features,and overcome the problem of the lack of robustness in the traditional tongue segmentation method.3.A segmentation algorithm combining dense convolution network and conditional random field is proposed.In order to solve the problem of locality of feature extraction by convolution operation,the fully connected conditional random field is used to optimize the tongue edge.