Super-resolution Hyperspectral Image Reconstruction Via Non-negative Structural Sparse Representation

High resolution(HR) hyperspectral images have important applications in many areas, such as anomaly detection, target recognition and image classification. Compared to RGB images, hyperspectral images provide rich spectral information. Every pixel in hyperspectral images contains a continuous spectrum which can be used for accurate and detailed description of materials in the scene. However, due to the limitation of sensors, it is challenging to obtain HR hyperspectral images directly. Current hyperspectral imaging lack spatial resolution seriously. It is difficult and limited for hardware devices to improve the spatial resolution of hyperspectral images. Thus the methods based on software are preferred.The compressive sensing theory provides a new signal acquisition method, which insures perfectly reconstruction of the signal from the few measurements. And it can save resources and reduce cost and demand for hardware. Recently, under the framework of the compressive sensing theory, the methods attending to reconstruct HR hyperspectral images from the pair of low resolution(LR) hyperspectral images and HR RGB image of the same scene have shown promising results. In these methods, sparse non-negative matrix factorization(SNNMF) technique was proposed to exploit the spectral correlations among the RGB and spectral images. However, only the spectral correlations were exploited in these methods, ignoring the abundant spatial structural correlations of the hyperspectral images. This thesis is mainly concerned with the algorithm of super-resolution(SR) hyperspectral image reconstruction. The thesis major contributions are outlined as follows:1.Through the study of non-local similarity of spectral images, this thesis proposes an SR hyperspectral image reconstruction algorithm combining the structural sparsity as priori knowledge and non-negative matrix factorization technique. Compared to the traditional 1-norm constraint, the structural sparsity is more stable and accurate. Structural sparse representation utilizes both the spectral and spatial redundancies of hyperspectral images and non-local similarity of spectral lines, leading to better reconstruction performance. Experiment results show that the approach performs better than other state-of-the-art methods, especially some details are better on the visual effect.2. Through the study of spectral images smoothness, this thesis proposes an algorithm combining the sparsity, smoothness of spectral images and non-negative matrix factorization technique. The sparsity regularization in the algorithm takes advantage of the sparsity of the hyperspectral images. And anisotropic total variation regularization with hyperspectral images smoothness as priori information further raises the algorithm performance and visual effect. Experiment results show that the algorithm has validity and reliability and performs better than other state-of-the-art methods on the visual effect and in the quantitative assessment.