Hyperspectral images have important applications in many areas of civil and military filed such as geological prospecting,agriculture and military applications.In hyperspectral imaging,the acquisition of each spectral image corresponds to a very narrow spectral window,so the imaging system must use a longer exposure time to collect enough photons to obtain a spectral image with good signal to noise ratio,resulting in acquisition the hyperspectral image is seriously lacking in spatial resolution,which greatly limits the application of hyperspectral images.Recently,high-resolution spectral images have been extensively studied by using high-resolution RGB images and low-resolution spectral images combined solutions in the same scene to obtain high-resolution spectral images.Such as the classical non-negative sparse matrix decomposition(SNNMF)algorithm,which only considers the correlation between the spectral and does not take advantage of the spatial correlation of hyperspectral images.In this thesis,the hyperspectral image reconstruction algorithm includes two parts: spectral dictionary learning and sparse coding.First,we learn the spectral dictionary representing the spectral reflectance of the scene from the low resolution spectral image,and then learn from the low resolution spectral image and the high resolution RGB image the sparse coefficients needed to reconstruct hyperspectral images,and the main work and innovation is:1.The non-negative dictionary learning of hyperspectral images has specific physical meaning,that is,the spectral reflectance of the scene is non-negative.The traditional non-negative dictionary learning algorithm has the problem of high computational complexity and slow convergence.This thesis proposes a more efficient non-negative spectral dictionary learning method.This method only updates one of the base vectors in the nonnegative spectral dictionary in each iteration,and uses the block coordinate descent algorithm to obtain the closed solution in order to improve the efficiency and robustness of the dictionary learning.2.In order to improve the visual effect of reconstruction results,a new method of combining the total variation sparse model to describe the smoothness of hyperspectral images is presented.Compared with the traditional method,the gradient field of image is proposed.In this thesis,the non-negative sparse coefficients of hyperspectral images are constrained by total variation.In order to reduce the computational complexity of the objective optimization function,the objective function is solved by using the piecewise optimization method and the ADMM method.It is proved by experiments that the total variation regular model can have better reconstruction results in the sparse domain.3.In order to improve the accuracy of nonnegative sparse coding and use non-local self-similar prior knowledge,this thesis first uses a nonlocal sparse regular model to characterize the high degree of non-local self-similarity in hyperspectral images and then reconstruct the high resolution spectral image by solving the target optimization function.Finally,this thesis makes a large number of experiments on the hyperspectral image reconstruction results under simulated and real conditions.The experimental results show that the proposed method substantially outperforms existing state-of-the-art methods in terms of both objective quality metrics and visual effect,indicating that non-local self-similar prior knowledge for the reconstruction effect of the upgrade has a very important role. |