Abstract

Computed Tomography Imaging Spectrometers (CTIS) capture dense spectrum of dynamic scenes as compressed 2D sensor measurements. Model-based Hyper-Spectral (HS) image reconstruction algorithms devised for such systems are typically very slow, sensitive to the selected data and noise models, and can only restore HS images with poor spatial resolution. On the other hand, deep learning-based approaches, once trained, are capable of performing the reconstruction in real-time and are more suitable for high frame-rate applications but generally suffer from limited generalization capabilities. In this paper for the first time, we jointly address the issues of reconstruction speed and spatial resolution of CTIS through a simple and interpretable deep learning architecture partially inspired by the Filtered Back-Projection (FBP) algorithm used in conventional CT scans. Our model is able to exploit aliased pixel information in CTIS images to recover spatially super-resolved HS cubes. Experimental results on simulated and real data demonstrate the effectiveness of our approach not only in reconstruction quality, but also in computation time and generalization ability.

The full paper can be downloaded from here
The Supplementary Material can be downloaded from here