Hyperspectral imaging through scattering media via physics-informed learning

Overcoming light scattering in natural light conditions is a challenging but rewarding problem. For this problem, both physics-based and data-driven methods encounter bottlenecks to recover large amounts of hyperspectral data encoded in the scattered light spot. In this paper, we present a physics-informed learning method to solve the light scattering problem with the neural network and the incoherent light transmission matrix method. With the proposed method, we achieve single-shot hyperspectral imaging through scattering media exceeding the memory effect range. Hyperspectral images with a maximum resolution of 3.2-megapixel are recovered in 64 wavebands, which represents an order of magnitude improvement over the memory effect range. No additional imaging or dispersion elements are required in the system except for the scatter medium, thus ensuring a compact and inexpensive hyperspectral imaging system. Our recovery method outperforms other physics-based and data-driven methods in terms of recovery quality and time cost.