A Multi-Level Reconstruction Algorithm for Electrical Capacitance Tomography Based on Modular Deep Neural Networks

Electrical capacitance tomography (ECT) enables the imaging of multiphase flow systems in industrial processes. Recovering flow profiles from measured capacitance data in ECT is an inverse problem that is traditionally solved using numerical algorithms such as the Landweber and Tikhonov regularization (TV) methods. This paper, however, proposes a machine learning-based approach to the ECT inverse problem through the use of modular deep neural networks (MDNNs) in a multi-level image reconstruction scheme. The basis behind the method put forth is, instead of having a single neural network take in the capacitance measurements and perform the inverse imaging task on its own, the reconstruction is delegated to separate sub-neural networks that each only recovers the image on a particular subsection of the imaging domain. Our proposed approach has demonstrated improvement upon the Landweber and TV algorithms in terms of reconstruction accuracy, suggesting that MDNNs are suitable candidates for tackling inverse imaging problems.