Deep Balanced Prior for Refined Self-supervised Reconstruction of Snapshot Spectral Compressive Imaging

Hyperspectral imaging is an advanced instrumentation technology that can obtain the spectral features through measurement. Typically, hyperspectral images (HSIs) contain more spectral bands than standard RGB images, enabling them to store more information and reveal more detailed characteristics of the scene. Relying on this property, HSIs have been widely applied to many computer vision related tasks, agriculture, remote sensing, medicine, etc. Traditional imaging systems equipped with spectrometers scan scenes either spatially or spectrally to acquire HSIs, a process that typically demands considerable time. Consequently, these conventional systems are not ideal for capturing and measuring dynamic scenes. In recent developments, researchers have turned to snapshot compressive imaging (SCI) systems for obtaining HSIs. These SCI systems compress information of snapshots along the spectral dimension into one single 2D measurement. Nowadays, the development of SCI systems has become increasingly diverse . Among these, coded aperture snapshot spectral imaging (CASSI), which uses a physical mask and a prism to implement the multiplexing modulation, has been widely applied as a mature forward compression system.