Remote Sensing Image Fusion Based On Constrained Non-negative Matrix Factorization

Accompanied by the constant development and advance of computer vision,image fusion is also promoted to improve and update.As we all know,a single type of remote sensing device is limited by its own physical conditions or external environmental factors,which makes the remote sensing data acquired often fail to accord with the growing study needs of the current technology and life.In practical applications,multispectral images with low spatial resolution contain multiple spectral bands.It could emerge a variety of color which can be used to identify the ground object,land cover type and Marine spectral distribution characteristics.In terms of the observable object,panchromatic images with low spectral resolution could describe the structure information of its detail and outline.They are applied in environment detection,resource management and town planning.So the incomplete presentation for the real image information of the single image source has restricted seriously the application fields of remote sensing image in actual life.For the sake of displaying the target of high quality,the remote sensing image’s appearance open a new page for its development.NMF（Nonnegative Matrix Factorization）is a good method to address image fusion.But in the process of image fusion,fusion only depends on NMF may lead to the problems of the reduction for data sparsity,image information distortion and so on.For this purpose,the paper proposed the HPF（High Pass Filter）and spectral constrained₂₁ NMF algorithms and the spectral reconstruction and gradient constrained NMF to achieve fusion under the constraint NMF.Thus,the overall quality of the fused image has been greatly improved.The main work is as follows:（1）We proposed the HPF and spectral constrained₂₁NMF algorithms.The detail enhancement is considered necessarily for the fusion by reason of the low spatial resolution for multispectral images.In the first process of integration,the MS is decomposed by non-negative matrix factorization based on₂₁ norm.Subsequently,the spectral feature matrix containing the color information and the abundance matrix describing the spatial structure information are obtained.Then,the high-frequency components of panchromatic images via HPF and the abundance matrix of the multispectral images were superposed together by weighted averaging,and then reconstructed with the abundance matrix.Meanwhile,in order to retain the characteristic information of the original spectrum,spectral Angle mapping was selected as spectral constraint and introduced into the decomposition of₂₁ NMF to acquire the result.Through experiments,the fusion image obtained by this method has complete information performance,clear and intuitive image,and shows good performance in various fusion indexes.（2）The algorithm for a remote sensing image fusion based on spectral reconstruction and gradient constrained NMF is proposed.For improving the quality of spectral and the reducibility of details furtherly,It is essential to effectively unmix the multispectral and panchromatic image respectively so as to extract truthful and accurate trait.Multi-spectral images are also decomposed,and a regular term with minimal spectral reconstruction error is introduced into the decomposition to encourage the decomposed spectral feature matrix to contain the most real spectral features.Then,the NMF decomposition of the whole color image is carried out,and the gradient distance minimization constraint using₁ norm is introduced in the decomposition to encourage the abundance matrix of decomposition to extract more complete and natural details,edges and other nonlinear structures.Finally,the fusion results were obtained by NMF reconstruction.The results indicate that the algorithm could avoid spectral distortion effectively and present details clearly.It has a better performance in both visual effects and objective evaluation than traditional algorithms.