A neural network-based local decomposition approach for image reconstruction in Electrical Impedance Tomography

Electrical Impedance Tomography (EIT) is a method of imaging the impedance distribution inside a non-homogeneous medium based on current or voltage measurements on its surface. Being a non-invasive and non-ionizing image modality, its application can be extended to a multitude of areas, including robotics and specifically, tactile sensing. The use of EIT, however, is limited by the complexity of the inverse image reconstruction problem, which is non-linear and ill-posed. In this contribution, we propose a data-driven approach to image reconstruction, using Neural Networks. Specifically, the image containing the target object is divided into partially overlapping sub-images, where each sub-image is modelled with a bi-variate polynomial. The forward problem is solved using the EIDORS toolbox in MATLAB, thus resulting to a set of voltage measurements. A set of feedforward neural networks, one for each sub-image, are then trained using the voltage inputs and the target polynomial coefficients to perform image reconstruction. The simulation experiments demonstrate promising performance for the case of a 2D square object in a noisy background.