Mismatch Correction for End-to-End Designed Phase-Encoded-Based Spectral Imaging System

Hyperspectral imaging provides important information about the composition of materials as a function of wavelength, which is useful for various computer vision tasks. Currently, to acquire hyperspectral images, several optical computational architectures have been developed in search of the best results in tasks such as reconstruction. However, the quality of the reconstruction of the acquired hyperspectral information is directly affected by perturbations and misalignments of the optical sensing system that are usually not considered in the system modeling. For the reconstruction of hyperspectral images, correcting the misalignment between the simulated and implemented optical systems is necessary since the acquired data, because of this phenomenon, greatly lowers their quality. Therefore, this work presents a model that predicts the behavior of the system in real scenarios and a regularized projector to match the expected measurements required by the reconstruction step. The performance of the proposed method is tested using a phase-encoded spectral imaging system on simulations and testbed implementation, with overall better results compared to the traditional method of reconstructing real hyperspectral data from the laboratory, obtaining an improvement of 4.79 [dB] compared to state-of-the-art methods.