Ghost imaging at submillimeter waves: correlation and machine learning methods

We present experimental results on computational submillimeter-wave ghost imaging schemes. The schemes include a dispersive element introducing quasi-incoherent field patterns to the field of view and bucket detection of the back-reflected field across a significantly broad bandwidth. A single bucket detection without discrimination of the field of view into image pixels is used. The imaging experiments at 220-330 GHz with dispersive hologram show successful computational ghost imaging of a corner-cube reflector target at 600-mm distance. Two separate image-forming methods are compared: correlation and machine-learning. In the correlation method, the image is formed by integrating the predetermined quasi-incoherent field patterns weighted with the bucket detections. In the machine-learning method, high image quality can be achieved after non-trivial training campaigns. The great benefit of the correlation method is that, while the quasi-incoherent patterns need to be known, no a priori iterative training to the images is required. The experiments with the correlation method demonstrate resolving of the target at 600-mm distance.