The Reseach And Implementation Of Super-resolution Algorithm Of Hyperspectral Imagery Based On Redundant Dictionary

Hyperspectral imagery has been widely used in various fields for its rich amountof information. However, the problems brought by its low spatial resolution and largeamount of data make it subject to certain restrictions. So people imagine to improvethe level of the resolution of hyperspectral imagery by means of mathematics andsignal theory, namely the super-resolution restoration method. However, it is foundthat most super-resolution restoration technologies for the hyperspectral imageryusually process the image as multi-frame image which has complementaryinformation between each other. They did not consider the maintain capability of thespectral-dimension information in the super-resolution restoration process. In addition,due to the many bands of the hyperspectral imagery, the mount of the computingincreases. So, in this paper we present the over-complete sparse decompositionalgorithm for super-resolution restoration of the spatial and spectral resolution of thehyperspectral image. This algorithm not only effectively reduces the overallalgorithm’s computational, but also ensures the spectral characteristics consistency inthe super-resolution recovery process. The main work includes:We studied and implemented a kind of super-complete hyperspectral imagesparse decomposition algorithm. The algorithm builds the redundant dictionary bylearning the hyperspectral images. This makes hyperspectral image pixel can berepresented by a linear atom in dictionary. The redundant dictionary obtained by thelearning and training has laid a good foundation for the super-resolution ofhyperspectral imagery.We designed and implemented super-resolution restoration algorithm based onredundant dictionary to improve the dimensional spectral resolution of thehyperspectral imagery. Based on the training method of single redundant dictionary,we obtained a pair of high and low resolution redundant dictionary. In thesuper-resolution recovery process, firstly, we extract the corresponding spectral curveand sparse decomposition based on low-resolution redundant dictionary for the verypixel in the hyperspectral imagery. Then use the obtained sparse coefficients and thehigh-resolution dictionary to super-resolution restoration and getting the highresolution hyperspectral image. The experimental results show that the algorithmcombines the characteristics of redundant dictionary effectively and get a good result in the hyperspectral image restoration. Comparing with the interpolation recovery, thisalgorithm has obvious advantage.We designed and implemented super-resolution restoration algorithm based onredundant dictionary to improve the spatial and spectral resolution of thehyperspectral imagery. Based on the spectral dimension improvement, we add thesuper-resolution restoration to the dimension of space for hyperspectral image.Recovering the inputed low-resolution images using the super-resolution restorationwith original constraints, and using the maximum posterior with the accelerated edgeoptimization to sharpen edge of the rebuild image, we get high resolutionhyperspectral images finally. Experimental data show that the PSNR value of theimage which was recovered by the algorithm proposed by the paper and optimized bythe accelerated marginalization maximum a posteriori probability improved2.6dBcomparing with the interpolation recovery. The algorithm improves the spatial andspectral resolution of the hyperspectral imagery.