Non-line-of-sight imaging has important applications in the medical imaging, navigation, robotics and defense industries. In a new study, scientists have taken a step toward making such a variety of applications possible.
Non-line-of-sight imaging reconstructs the image of hidden objects through light scattered from these objects on other surfaces. In this method, a wall can be used as a mirror. Although most imaging methods use travel time information, research has shown that spatial correlations in scattered light have enough information to recovering hidden objects. These methods recover the latent image of the hidden object by solving the variations of a phase retrieval (PR) problem.
NLoS correlography allows NLoS imaging at sub-millimeter resolutions. However, the limitations of existing PR methods, in particular their sensitivity to noise, prevent this. In fact, the significant decrease in intensity with increasing distance is an important barrier to NLoS imaging. Now researchers are one step closer to solving this problem.
They first analyzed a NLoS correlography noise model so they obtained enough data to learn the NLoS correlation problems. They then, using deep learning, trained convolutional neural networks (CNNs) to solve the noisy PR problem. The results were efficient and robust to several forms of noise, a feature that the existing algorithms lack, and were also consistent with the laboratory data. The researchers were able to successfully reconstruct the shape of small hidden objects from a distance of about one meter using very short exposure times. This research is an important step towards high-resolution real-time NLoS imaging.