Generative Data Augmentation for Learning-based Electrical Impedance Tomography via Variational Autoencoder

Electrical Impedance Tomography (EIT) owns lots of potential industrial and biomedical applications due to its high temporal resolution and non-intrusive advantages. To improve the spatial resolution of EIT, a neural network-based image reconstruction method is proposed. Compared with the traditional neural network-based image reconstruction methods, the proposed method is constructed by the variational auto-encoder. To improve the generalization ability of the proposed network, a data generation strategy is proposed. Artificial conductivity images can be automatically generated following the same manifold of the preset image set. Numerical results proved that the proposed generation model can generate a desirable dataset for significantly improving the accuracy and generalization of the neural network.