Deep learning enabled reflective coded aperture snapshot spectral imaging

Abstract

Coded aperture snapshot spectral imaging (CASSI) can acquire rich spatial and spectral information at ultra-high speed, which shows extensive application prospects. CASSI innovatively employed the idea of compressive sensing to capture the spatial-spectral data cube using a monochromatic detector and used reconstruction algorithms to recover the desired spatial-spectral information. Based on the optical design, CASSI currently has two different implementations: single-disperser (SD) CASSI and dual-disperser (DD) CASSI. However, SD-CASSI has poor spatial resolution naturally while DD-CASSI increases size and cost because of the extra prism. In this work, we propose a deep learning-enabled reflective coded aperture snapshot spectral imaging (R-CASSI) system, which uses a mask and a beam splitter to receive the reflected light by utilizing the reflection of the mask. The optical path design of R-CASSI makes the optical system compact, using only one prism as two dispersers. Furthermore, an encoder-decoder structure with 3D convolution kernels is built for the reconstruction, dubbed U-net-3D. The designed U-net-3D network achieves both spatial and spectral consistency, leading to state-of-the-art reconstruction results. The real data is released and can serve as a benchmark dataset to test new reconstruction algorithms.