Electro-optic liquid crystal hyperspectral imaging via U-Net-based deep learning

Hyperspectral imaging (HSI), characterized by its capacity to resolve fine-grained spectral signatures alongside spatial distributions, has emerged as a pivotal frontier in optical sensing. Here, we propose and demonstrate a compact computational HSI framework that leverages the electrically tunable birefringence of a uniformly aligned liquid crystal (LC) module. By modulating the LC molecular orientation with distinct gating voltages, the object’s hyperspectral information is optically encoded as a sequence of compressed grayscale intensity patterns via a polarization interference scheme. Subsequently, a deep learning architecture based on U-Net is employed to reconstruct the high-dimensional hyperspectral data cube from the multi-channel input through end-to-end learning. Our experimental results demonstrate high-fidelity reconstruction across 31 spectral channels within the visible spectrum (400–700 nm), exhibiting agreement with ground-truth data in both spatial morphology and spectral consistency. Our approach paves the way for miniaturized, cost-effective HSI solutions, with potential applications in environmental monitoring, precision agriculture, and biomedical diagnostics.