| Spectral imaging can acquire not only two-dimensional spatial information of the object, but also one-dimensional spectrum information of each pixel. It has become an important role for agriculture, medical, marine and military. Benefiting from compressive sensing theory, observation data of hyperspectral images is reduced greatly under the sampling rate below the traditional sampling theorem. In this thesis, a low rank reconstruction algorithm for hyperspectral images in coded aperture imaging system is studied with a framework of compressed sensing and the similarity theory of nonlocal structure.Firstly, the current situation of compressive sensing and its applications in imaging system is introduced. And the principle of compressed sensing theory is analyzed from the sparse representation of signals. The mathematical model of coded sensing imaging is introduced in the case of multi-dimension scene.Secondly, the principle of nonlocal image restoration algorithms is analyzed based on the nonlocal approximation theory. And the nonlocal structural similarity in the spectral data cube is excavated which is inspired by the redundancy of 2D image. The noise is removed utilizing low-rank approximation within similarity-based clustering.Finally, a code pattern guided by the saliency information of color image is proposed for CASSI. A similarity measure combined with color information is developed to cluster nonlocal patch from the estimation of data cube. And the low-rank characteristic of the nonlocal clustering is used as a priori information in the realization of CS reconstruction, bringing forward a Color Guided Nonlocal Low Rank(CGNLR) reconstruction algorithm. The experiment of reconstruction algorithm is based on the measured data of hyperspectral images of natural scenes. Simulation results demonstrate that CGNLR can significantly outperform classical total variation algorithm in the PSNR and image detail and texture of reconstruction results. |
PDF Full Text Request