Remote Sensing Image Fusion Based On Image Detail Extraction And Convolution Sparse Representation

With the rapid development of science and technology,the development of remote sensing satellite technology is also changing with each passing day.Researchers can obtain a large number of remote sensing images every day,but it is limited to the technical limitations of remote sensing satellites,under the condition of keeping the signal-to-noise ratio constant,if we want to obtain high-resolution spatial information,we have to sacrifice the spectral resolution.On the contrary,if we want to obtain rich spectral information,we have to sacrifice the spatial information,therefore,the direct acquisition of remote sensing image is not conducive to the subsequent feature extraction and image classification.Remote sensing image fusion can solve the problem of contradiction between detail information and spectral information,and obtain remote sensing image with high spatial resolution and rich spectral information.In recent years,with the application of machine learning in various fields,convolutional sparse representation also plays an important role in the field of remote sensing image fusion,it can sparsely represent the whole image,fully consider the correlation between pixels,reduce the loss of spectrum and detail of fusion results,so convolution sparse representation is widely praised in the field of image fusion.Based on this,this paper proposes a remote sensing image fusion method based on convolution sparse representation and morphological filter,but in this method,some details will be missing when training dictionary filter,and the training samples are few,which affects the accuracy of prediction results.Moreover,the traditional fusion methods are aimed at improving the fusion rules,and ignore the image pre-processing work,in fact,the interpolation and amplification of multispectral images before fusion will make some details missing.Therefore,a remote sensing image fusion method based on convolution sparse representation and Non-Subsampled Contourlet transform(NSCT)is proposed.The specific work is as follows:1.The theory of morphological filter is studied.Its operation mode can make things interact with structural elements,and it has more advantages than other filters in detail extraction.Therefore,the filter composed of semi gradient operator is selected to obtain the details of the image,so that the fusion result can have higher spatial resolution.2.The principle of convolutional sparse representation is studied.Convolution sparse representation is a direct sparse representation of a whole image,which has translation invariance and fully considers the semantic relationship within the image.Therefore,a remote sensing image fusion method based on convolution sparse representation and morphological filter is proposed,first,morphological filter is used to obtain the detail information of the image,and then high-resolution and low-resolution dictionary filter is trained.Feature mapping can be obtained through low-resolution dictionary filter and detail image,and then feature mapping and high-resolution dictionary filter are used to reconstruct the detail information,so as to obtain the final fusion result,The spatial resolution is also improved while maintaining the spectral information.3.Remote sensing image fusion based on convolution sparse representation and NSCT is proposed.Firstly,convolution sparse representation is used to build the model to complete the super-resolution of the image.In the stage of image fusion,NSCT is used to decompose the super-resolution image and panchromatic image to obtain their own multi-scale and multi-directional high-frequency subband and low-frequency subband,in order to make the fusion result have more spectral information and spatial information,the fusion rule used in the high frequency subband is weighted average,while the fusion rule used in the low frequency subband is that the sparse coefficient is larger.Finally,the new subband obtained by fusion is transformed by NSCT to get the fusion image.Experimental results show that the fusion results obtained by this method perform well in the aspects of feature representation and spectral retention.