Adaptive local sparse representation for compressive hyperspectral imaging

Coded aperture snapshot spectral imaging (CASSI) is an effective way for hyperspectral imaging. In CASSI, the key issue is to accurately and efficiently reconstruct the 3D hyperspectral image from its corresponding coded 2D image. Due to the ill-posed nature, reconstruction errors are inevitable, a feasible solution is to add an RGB camera for complementary sampling to reduce the reconstruction error. In this paper, we investigate the structural changes of local image patches in different bands and their correlation with RGB observation, propose a reconstruction method for dual-camera CASSI system. Specifically, we learn an adaptive dictionary with RGB observation, then use RGB observation to guide the selection of the adaptive dictionary for each local image patch of the reconstruction target, and finally reconstruct the original hyperspectral image through an iterative numerical algorithm. This method fuses the spatial and spectral information obtained from RGB observations into the reconstruction process, experimental results show that the proposed method can greatly improve the reconstruction quality, especially the reconstruction of the details, and reduce more time compared with past dictionary-based reconstruction methods.

Automated summary

This paper proposes a reconstruction method for dual-camera CASSI system, which learns an adaptive dictionary with RGB observation and uses RGB observation to guide the selection of the adaptive dictionary for each local image patch. Experimental results show that the proposed method can greatly improve the reconstruction quality, especially the reconstruction of the details, and reduce the reconstruction time compared with past dictionary-based reconstruction methods.