Snapshot learning for color and spectral recovery toward edge deployment

Abstract

We present per-image learning frameworks for color and spectral recovery that convert conventional RGB images into physically meaningful color and spectral representations through the co-design of hardware and learning models. The central concept is to learn an RGB-to-spectrum or RGB-to-color mapping directly from a small amount of paired reference information collected within the same scene and then apply the trained model across the entire image. For spectral recovery, paired RGB and spectral data from a small but representative subregion are used to train a neural network or statistical learning model. The resulting model is then applied to the full image to reconstruct a dense spatiospectral hypercube without requiring complete spatial or spectral scanning. Unlike many data-driven color-correction approaches that require large training datasets, the color-recovery framework trains a lightweight network independently for each image. As representative examples, we apply these one-shot learning frameworks to several clinical applications in resource-constrained settings. We demonstrate that one-shot color and spectral recovery can bring computational intelligence to conventional smartphones through edge computing.