Submillimeter-wave holograms and imaging neural network: Experiments at 220-330 GHz

We present the experimental results of a submillimeter-wave standoff imaging system based on a frequency-diverse hologram and image reconstruction via machine learning at 220-330 GHz. The imaging system operates in a single-pixel, monostatic configuration consisting of a transceiver together with a frequency-diverse phase hologram to interrogate the region of interest with quasirandom field patterns. The spatial reflectivity distribution in the region of interest is embedded in the wide-band frequency spectrum of the back-reflected signal and the images are acquired without mechanical or electrical scanning. Images from a visible-light camera are used as the ground truth of the target elements. The targets are scanned in the region of interest, while the wide-band reflection spectrum for the target is measured. The collected imagesignal pair data are used to train a deconvolutional neural network for image reconstruction with the submillimeter-wave reflection spectra as input. In experiments, a corner-cube reflector and a complex test target made of copper foam were imaged in a 28-degree field of view at a distance of 600 mm from the imaging system. The effect of bandwidth on image quality is evaluated using 10-40 GHz bandwidths centered at 275 GHz to image the copper foam target. The resolution in the image predictions was estimated from fitted point-spread functions to be from 12 mm to 30 mm, with the highest resolution at the broadest bandwidth. We have correlated the measured field patterns at the region of interest with the mean squared error (MSE) of the predicted corner-cube images to analyze the effect of field characteristics on imaging accuracy. The results demonstrate increased accuracy in locations with high electric field amplitude and variation over the imaging bandwidth.