Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging

Abstract

Due to the quartic attenuation of intensity with distance, low signal-to-noise ratios are arguably the fundamental barrier to real-time, high-resolution, non-line-of-sight (NLoS) imaging at long stand-offs. We use results from spectral estimation theory to derive a noise model for NLoS correlography; a speckle correlation-based technique for recovering occluded objects from indirect reflections. Then, using only synthetic data sampled from the proposed noise model, and without knowledge of the experimental scenes nor their geometry, we train a deep convolutional neural network (CNN) to solve the noisy phase retrieval (PR) problem associated with correlography. We validate that the resulting CNN is exceptionally robust to noise, far exceeding the capabilities of existing algorithms in low-flux regimes. We use the proposed technique to demonstrate NLoS imaging with 300um resolution at a 1m standoff, using just two 1/8-th second exposure-length images from a standard CMOS detector. This lifts two critical restrictions of existing NLoS imaging methods: (1) Typical NLoS reconstruction methods that utilize transients produce resolutions on the order of 3-10 mm; our results are an order of magnitude improvement. (2) Typical NLoS capture times are of the order of tens of minutes to hours; our results represent a 2 orders of magnitude improvement.