Research On EIT Image Reconstruction Method Based On Unsupervised Learning

Electrical Impedance Tomography(EIT)is a novel nondestructive functional imaging technology.Because of its advantages of visualization,non-radiation,non-invasiveness,fast response,simple structure and low cost,it has been the hot spot at home and abroad.EIT image reconstruction is an inverse problem.Due to the serious ill-condition and non-linearity of the EIT inverse problem,the traditional reconstruction methods solve the problem by mapping EIT inverse problem to a specific space,which results in reduced image reconstruction accuracy.In the existing machine learning methods,the neural network is easy to fall into the local minimums,and it is difficult to obtain enough labeled samples for practical application,which leads to the poor generalization ability of the model.Unsupervised learning algorithm does not need a separate offline training process and marked training data sets,and can make reasonable decisions by extracting the corresponding inner rules and principles through independent learning data sets.This paper study the EIT image reconstruction methods based on unsupervised learning,the main work is as follows.First,establish a mapping from EIT measurement values to reconstructed image,and transform the mapping relation into a classification problem,so the image pixel values are reconstructed according to the pixel classification result.Second,in order to classify the pixels better and improve the accuracy of the EIT reconstructed images,an EIT boundary measurement voltage preprocessing method based on equipotential theory is proposed.Establish a point-to-point mapping form object field to pixels of the reconstructed images,and take the voltage vector of the object field segmented sub-blocks which obtain after preprocessing the measured voltages as the feature vector to classify the object field segmented sub-blocks,and then get the pixel classification results.Third,in this paper,an unsupervised learning method based on Fuzzy C-Means Algorithm(FCM)clustering is proposed,which is based on the fact that different materials have different electrical impedances and different material regions have different characteristic vectors.By classifying the feature vectors of the object field segmented sub-blocks in EIT system to reconstruct the conductivity distribution in object field to reconstruct the EIT images.Fourth,in order to verify the effectiveness of the algorithm,the image qualities of EIT reconstructed images are evaluated objectively and quantitatively.Some EIT image quality evaluation indexes with physical meanings are put forward.The experimental results of algorithm proposed by this paper,Conjugate Gradient method(CG)and BP neural network algorithm are evaluated respectively by computing the duty cycle,center of figure,center of mass and shape similarity.The effectiveness of the algorithm in this paper is verified by simulation and system experiments.The experimental results show that the unsupervised learning algorithm based on clustering mechanism can reconstruct the material distribution in EIT object field better compared with the traditional CG reconstruction method and the BP neural network algorithm based on supervised learning.