Non-line-of-sight (NLOS) imaging is a technology that computes images of targets outside the camera’s field of view or behind obstacles with the help of a relay surface. In real-world scenarios, the quality of NLOS images is poor due to the random and uncertain scattering characteristics of the relay surface. The noise and scattering information are coupled, making the reconstruction of clear images difficult. Time-of-flight (TOF) camera is a kind of active 3D imaging camera with the advantages of low cost, real-time capability, and rich data. This paper explores an NLOS 3D reconstruction method based on TOF cameras. Since pure Lambert surfaces do not exist in the natural world, two common materials as the relay surfaces, polypropylene (PP) plastic and polymethyl methacrylate (PMMA) plastic sheets, were randomly employed during the reconstruction of NLOS depth images. A deep neural network model is constructed to learn the scattering features of the relay surfaces, so as to realize the separation of scattering features and target features. The scattering features model base of common materials should be established first in practical application. In the experiment, 12 plaster portraits were measured, and each of them was captured depth image by 360° rotation. Then, based on U-Net network and attention mechanism, a DU-Net deep neural network is proposed. The attention mechanism channel focused on the target information and ignoring the noisy data. After training, the network model has a good reconstruction result for the degenerated images outside the training set.
|