| Electromagnetic tomography(EMT)is a tomography technology based on the principle of electromagnetic induction.The measured voltage at the boundary of the measured object field is acquired by the electromagnetic sensor,and then the internal conductivity distribution of the measured object is reconstructed through the image reconstruction algorithm.Because EMT technology has the advantages of non-radiation,non-contact,non-invasive,visualization,high safety performance,and rich measurement information,the technology is widely used in the industrial and medical fields.The inverse problem of image reconstruction is the key of EMT.Traditional image reconstruction algorithms linearize the mathematical model,resulting in blurred reconstructed images,lack of detailed features,and poor imaging quality.Inspired by deep learning in solving nonlinear problems,many deep learning networks have been applied to the inverse problem solving of EMT.But these networks have the problem of unsatisfactory recovery of defect details and poor robustness.This subject proposes a deep learning method based on the fusion network of VAE(Variational Auto-encoder,VAE)and GAN(Generative Adversarial Networks,GAN)to improve the detail accuracy,robustness of image reconstruction,and the quality of image reconstruction.The main work of this dissertation is as follows:1.A VAE-GAN fusion network is designed for EMT image reconstruction in this disseratation.The VAE with a multi-layer convolution structure is used as the generator of the GAN network to generate the initial reconstructed image.The discriminator network of GAN is retained,and the generator parameters are continuously optimized for adversarial learning to obtain more accurate reconstructed images.The VAE-GAN network improves the accuracy and robustness of image reconstruction.2.The electrical conductivity distribution of metal parts is studied in this disseratation,and the simulation database is made to train the VAE-GAN network.Defect samples of different shapes are simulated.On the basis of traditional round samples,square and triangle samples are added to improve the edge preservation of the network.Corresponding simulation experiments and noise robustness experiments are carried out to verify the feasibility of the VAE-GAN fusion network proposed in this dissertation for EMT image reconstruction.The results show that the imaging effect of the VAE-GAN fusion network proposed in this dissertation is better than the traditional reconstruction algorithm.The results of the VAE-GAN network algorithm are reconstructed,and the image evaluation parameters are introduced for verification.The image error is smaller and the correlation coefficient is larger,which confirms the feasibility of applying deep learning theory to the traditional EMT field.3.The EMT system experiments of metal parts with surface defects was carried out.The experiment results verified that the network constructed by simulation data can also be directly used for experimental data and proved the practicability of the VAE-GAN fusion network.Compared with other networks,the VAE-GAN fusion network proposed in this dissertation has higher accuracy in reconstructing the position and size of the defect in practical applications. |
