Effect of data noise on LSTM-FC scattered-field processing for microwave imaging

A strategy to mitigate model error in microwave imaging by introducing a neural-network-based preprocessing of the scattered field is considered in this paper. In particular, the approach consists of a long short-term memory (LSTM) cell combined with fully-connected (FC) neural layers. Such a network, which works in the time domain, aims at extracting the scattered field contributions as they were measured by a canonical two-dimensional setup with line-source antennas and ideal probing elements. The extracted data are then given in input to a quantitative tomographic technique formulated in the mathematical context of Lebesgue spaces with variable exponents. Here, the effect of input data noise on the whole imaging procedure is evaluated for the first time. Results obtained on simulated data involving circular dielectric cylinders are presented to assess the processing error and the imaging performance against the signal-to-noise ratio.