Multi-spectral Image Fusion Method Based On Multi-scale Geometric Analysis And Sparse Representation

The development of remote sensing technology provides a powerful tool for human observation of earth.Due to technical limitations of the spectral imaging sensor,human can only get spectral images with a lower spatial resolution,which cannot satisfy the requirements of most applications.By combine multispectral images and panchromatic images with high spatial resolution,a multispectral image with high spatial resolution can be obtained,which can effectively solve the problem of low spatial resolution of multispectral images.These methods are also called Pan-sharpening methods of multi-spectral image,or the fusion methods of multi-spectral image and panchromatic image.It is widely concerned by researchers at home and abroad.Aim at solving the problem of Pan-sharpening,in this paper,we studies kinds of classical remote sensing image fusion methods and the evaluation criteria of remote sensing image fusion quality,also we comprehensively compares the performance of kinds of fusion algorithms with different evaluation indexes.Based on the theories of multi-scale geometric analysis,sparse representation and dictionary learning,the current mainstream image fusion technology is summarized,and the existing algorithm is improved.The main work and research results of this paper are as follows:(1)Based on multi-scale geometric analysis,an adaptive pan-sharpening method is proposed.Most of the traditional researches which based on multi-scale fusion are based on the substitution of different frequency information,ignoring the relationship between the bands.In this paper,an adaptive pan-sharpening method based on non-subsampled contourlet transform is proposed:non-subsampled contourlet transform is used to decompose panchromatic image and multispectral image,then using adaptive fusion criterion,keep the low-pass sub band of multispectral images unchanged;directional-sub bands are adaptively injected spatial details using variable regression method,parameters of each directional sub band include injection gain parameter and average luminance weighting parameters,which were estimated jointly;high-pass sub band is replaced with panchromatic image' s high-pass sub band.The simulated experimental result shows that compared with the traditional methods,the proposed method can get a better fusion effect.(2)Due to the huge calculation of the fusion method which using K-SVD algorithm to learn the dictionary,this paper proposes an improved algorithm based on non-subsampled contourlet transform and dictionary learning.The method combining multi-scale geometric analysis and sparse representation,extract the image details with non-subsampled contourlet transform,with the luminance component constitute the high and low resolution Dictionary pair,according to the assumption that scale invariance,combined with the theory of sparse representation based on sparse representation,obtained spatial details need to be injected into the multi spectral image.The simulated experimental result shows that the proposed method effectively reduces the time of the dictionary training and keeps the spectral information of the multi spectral image very well.(3)The experimental results on GeoEye1?WorldView2 and Quick Bird remote sensing images show that the methods proposed in this paper can effectively reduce the spectral distortion while enhancing the spatial resolution of multi-spectral images.The visual effects and objective evaluation indexes,such as RMSE,SAM,ERGAS,UIQI,sCC,Q4 also shows good results.