A Novel Deep Structure Auto Encoder for Image Reconstruction in Electrical Resistance Tomography

As a non-invasive and non-radiative process visualization technique, electrical resistance tomography (ERT) has attracted more and more attention in industrial and biomedical fields. Image reconstruction play a vital role in ERT. The research proposes a deep-learning approach to ERT image reconstruction using an auto-encoder network. The stacked layered structure of the auto-encoder neural network can extract latent features for well-posed spatial information that provides improved imaging accuracy. For more thorough retrieval of latent information and improved image accuracy, the stacking design of the auto-encoder neural network is crucial. A novel deep structure auto encoder (SAE) is proposed here by stacking several layers in a neural network and an upgraded adaptive moment estimation (ADAM) optimizer is used to enhance image accuracy beyond the standard reconstruction approach. A well-designed proposed SAE model can enable smooth transitions between various distributions, allowing good generalizations to new instances that best fit the training data’s learned distribution. Simulation and experimental tests are carried out and the results verify the effectiveness of the proposed method.