Single‐Exposure High‐Dynamic‐Range 3D Measurement via Multi‐Polarization Imaging and Physics‐Informed Zero‐Shot Learning
Fuqian Li, Qican Zhang, Yajun Wang
Abstract
Abstract In optical 3D sensing, high‐efficiency and high‐precision reconstruction of high‐dynamic‐range (HDR) scenes is highly crucial. Current HDR methods require either fusing images under multiple exposures or training network on massive datasets. To this end, a single‐exposure full‐resolution multi‐polarization fringe enhancement method, establishing an effective hardware‐algorithm collaboration mechanism for HDR 3D measurement for the first time, is proposed. Specifically, in the hardware part, a full‐resolution multi‐polarization (FM) imaging method is proposed. Unlike conventional multi‐polarization modulation methods, the method avoids the 3/4‐pixel loss and the instantaneous field‐of‐view error. Furthermore, by eliminating the specular reflection, more reliable scene information is captured, especially in overexposed areas, which significantly reduces the burden of fringe repair algorithm. In the algorithm part, a hybrid repair algorithm is proposed based on physics‐informed zero‐shot learning. The algorithm incorporates the physics priors of phase retrieval, FM fringe noise, and deep neural network into a dual‐stream network architecture. This enables it to address underexposure and overexposure issues in the FM fringe without network pretraining on any dataset. Thus, the advantages of hardware modulation and algorithm optimization are fully integrated. Experiments on static and dynamic scenes with complex reflectivity demonstrate the superiority of the method in efficiency, resolution, fidelity, and generalization ability.