Tunable Optimally-Coded Snapshot Hyperspectral Imaging for Scene Adaptation

Snapshot hyperspectral imaging (SHI) is increasing demand for various applications in dynamic scenes. Current mainstream solutions rely on machine learning with open-source datasets to acquire fixed compression encoder and reconstruction decoder, which limits their generalizability across diverse real-world scenarios. Herein, these challenges are addressed by a tunable optimally-coded SHI (TOSHI) system that allows dynamic optimization of optical encoding elements and software decoding strategies based on actual scene data. To improve scene adaptability, a domain-aware adaptive mechanism is introduced that extracts spatial and spectral features from ground truth data to calibrate the system through transfer learning and parameter-conserving fine-tuning. Leveraging spatial division multiplexing technology, TOSHI is equipped with an auxiliary imaging structure to acquire ground truth, enabling more efficient scene adaptation. As a demonstration, a proof-of-concept prototype is developed with an image resolution of up to 5120 x 5120 pixels, an angular resolution of 0.05 degrees, a spectral resolution of 10 nm within the visible wavelength, and a spatial-temporal resolution of up to 2048 x 2048 pixels @14.7fps, achieving a PSNR improvement of approximate to 3.54 dB over conventional SHI systems. Additionally, TOSHI has been verified for online industrial measurements, including active and passive lighting devices, through extensive experiments.