Millimeter- and Submillimeter-Wave Imaging Through Dispersive Hologram and Deep Neural Networks

We present imaging results with dual-band millimeter- and submillimeter-wave hologram and deep neural networks (NNs). The imaging method uses a single transceiver, which interrogates the region of interest (RoI) through a dispersive transmission-type hologram. The hologram was designed to cover two bands 50–75 and 220–330 GHz. Two separate single-transceiver imaging experiments were carried out with two test objects translated in the RoI at
$101\times101$
locations. NNs were trained to images of the test objects with wideband reflection spectra from the RoI as the input. The deep NNs were based on deconvolutional (DC) layers that mapped the latent information of the test objects in the spectra to image pixel values. The two
$\sim 10$
-cm test objects were imaged in 200
$\times200$
mm
²
and
$300\times300$
mm
²
field-of-view at 600 mm from the hologram aperture (19°–28° angular field-of-view). The experimental resolution was estimated from point-spread functions extracted from the predicted images. The full width at half maximum resolution was 21 and 16.5 mm, for the 50–75 and 220–330 GHz bands, respectively. These are close to the theoretical limits of 25–19 mm, for the lower band, and 19–16 mm for the higher band as predicted with hologram aperture size and edge taper. Augmented reflections were constructed from corner-cube measurements to evaluate the ability to predict the images of vast collection of objects. The results with augmented data show performance comparable with the experimental ones with limited test object space. The latent representations for both the experimental and augmented data indicate sparsity—a demonstration of feasibility to generalize from reflection spectra to images. The performance of the developed imaging technique is in par with the current, multichannel state of art, and has the advantage of substantially reduced hardware complexity.