Abstract : Convolutional neural networks (CNN) are applied
to the time-harmonic electromagnetic diagnostic of a dielectric
micro-structure. The latter consists of a finite number of circular
cylinders (rods) with a fraction of wavelength radius that are set
parallel to and at sub-wavelength distance from one another.
Discrete scattered fields are made available around it in a freespace
multisource-multireceiver configuration. The aim is to
characterize this micro-structure, like positions of rods or their
absence, and in effect to map their dielectric contrasts w.r.t. the
embedding space. A computationally efficient field representation
based on a method of moments (MoM) is available to model the
field. Iterative, sparsity-constrained solutions work well to find
missing rods, but may lack generality and need strong priors.
As for time-reversal and like noniterative solutions, they may
fail to capture the scattering complexity. These limitations can
be alleviated by relying on deep learning concepts, here via
convolutional neural networks. How to construct the inverse
solver and which resulting architecture appears the most efficient
is focused onto. Representative numerical tests illustrate the
performance of the approach in typical situations. Comparisons
with results from a contrast-source inversion (CSI) introduced in
parallel are performed. Emphasis is on potential super-resolution
in harmony with subwavelength features of the micro-structure
