Multi-Resolution Coded Apertures Based on Side Information for Single Pixel Spectral Reconstruction

Abstract

Compressive spectral imaging (CSI) architectures allow to reconstruct spectral images from a lower number of measures than the traditional scanning-based methods. In these architectures, the coded aperture design is critical to obtain high-quality reconstructions. The structure of coded apertures is traditionally designed without information about the scene, but recently side information-based architectures provide prior information of the scene, which enables adaptive coded aperture designs. This work proposes the development of an adaptive coded aperture design for spectral imaging with the single pixel camera, based on a multi-resolution approach. An RGB side image is used to define blocks of similar pixels, such that they can be used to design the coded aperture patterns. This approach improves the reconstruction quality in up to $23\mathbf{dB}$ compared with traditional single pixel camera, and the computation time in up to 99.5% because it does not require an iterative algorithm.