Snapshot spectral imager using orthogonal coding and an untrained spectrally informed decoder network

Abstract

Coded Aperture Snapshot Spectral Imaging (CASSI) has emerged as a powerful technology for acquiring hyperspectral images through a single compressed measurement. However, achieving high spatial and spectral resolution simultaneously remains a significant challenge. In this work, we propose a novel combination of orthogonal coding and a dedicated self-supervised reconstruction algorithm to significantly enhance hyperspectral imaging performance. The key insight behind our approach is that orthogonal coding fundamentally transforms the reconstruction problem from a compressed sensing (CS) problem to an inpainting problem, where each spectral channel is sampled by a random distinct subset of pixels rather than being multiplexed. This shift in problem formulation motivates our integration of a self-supervised reconstruction framework, which is particularly well-suited for inpainting-based restoration. Specifically, we design an orthogonal coded aperture (mask) that ensures spectral bands are modulated by mutually orthogonal patterns, effectively minimizing spectral cross-talk. To complement this encoding strategy, we develop a self-supervised reconstruction algorithm that employs an untrained convolutional decoder network with tailored structure and weight initialization, capturing both spatial and spectral priors. The synergy between our encoding and decoding strategies leads to superior imaging performance, significantly outperforming existing model-driven methods in terms of PSNR, SSIM, and SAM metrics, as demonstrated through simulations and real-world experiments.