Spatio-Spectral Enhancement of Single-Pixel Reconstruction by a Fully Convolutional Neural Network

Spectral images (SI) contain a large amount of information, which increases the cost of their collection, storage, and processing. Compressive spectral imaging (CSI) methods allow the reconstruction of that information from a small set of random projections. However, compressed sensing with high spatial-spectral resolution requires expensive sensors. The single-pixel camera is an architecture that provides a low-cost solution for compressive SI acquisition, but with a significant limitation in image resolution. This work proposes a deep learning approach by means of a fully convolutional neural network (FCNN) with the ability to retrieve the spatio-spectral information from a single-pixel reconstruction in order to produce a spatio-spectral enhancement, suitable for applications requiring higher quality and low-cost acquisition systems. Simulations and experimental results show that the trained FCNN with a limited data set improves the quality of the single-pixel reconstruction in up to 20 dB, without requiring additional measurements. The performance of the proposed approach is compared against the traditional method of the single-pixel reconstruction.