Research On Remote Sensing Image Fusion Algorithm Based On Low-rank Decomposition And Convolutional Neural Network

With the booming of computer vision theory,the utilization of image processing technology has become increasingly widespread.Image fusion as a key processing technology to improve the visual quality of images has been widely researched by scholars across the world.Among them,remote sensing image fusion occupies a major position.The types of remote sensing images mainly include Synthetic Aperture Radar(SAR)images,multispectral images and panchromatic images.Different images have different imaging characteristics and manifestations.Using technology to fuse remote sensing images can greatly improve image resolution,and high-resolution remote sensing images have a significant role in land planning,national defense security,weather prediction,map navigation and land and resources exploration,so image fusion has great research value.Therefore,this article will take remote sensing image as the main research object,and discuss how to improve the fusion quality of SAR image and multispectral image and the fusion quality of multispectral image and panchromatic image.The main research contents are as follows:Firstly,the process and level of image fusion are introduced,and the significance of remote sensing image fusion in production and life is emphasized.Then,the existing fusion algorithms in remote sensing field are summarized,and the fusion principle and fusion process of component substitution method,multiscale decomposition method and network model method are introduced in detail.The representative algorithms such as IHS(Intensity Hue Saturation)transform,wavelet transform and deep learning model are introduced in detail.At the same time,the problems of the existing methods are also pointed out.Secondly,aiming at the problems of spectral distortion and details ambiguity that often occur during the fusion of SAR and multispectral images,this paper starts from the matrix low-rank theory and proposes a fusion algorithm for SAR and multispectral images that combines spatial transformation and low-rank decomposition.Firstly,IHS algorithm is used to transform the multispectral image,and then the latent low-rank representation(Lat LRR)algorithm is used to extract and fuse the features of SAR image and the intensity component after spatial transformation.The fused component will be used as a new intensity component to perform IHS inverse transform to get the final image.Among them,SAR images need to be preprocessed by multi-looking,filtering and geocoding because of its imaging particularity.Finally,in order to verify the fusion effect of the proposed algorithm,fusion experiments with six mainstream fusion algorithms are carried out.The experiments prove that the algorithm based on IHS-Lat LRR can effectively improve the problem of spectral information loss in the process of SAR and multispectral image fusion,and is superior to the existing fusion algorithms in subjective and objective analysis.Finally,in order to further improve the ability of feature extraction from the model,a Pan-sharpening fusion algorithm based on convolutional neural network is proposed.In this algorithm,CNN is used as the backbone network,and dense connection mode is introduced to enhance the learning ability of network.In data selection,panchromatic image and multispectral image are used as fusion data,and preprocessed according to Wald protocol.The trained model is used for feature extraction and fusion.In order to verify the fusion effect of the proposed algorithm,simulation data experiments and real data experiments are carried out between the proposed algorithm and seven mainstream fusion algorithms.The fusion results show that the fusion images of this algorithm are not only rich in color and good in visual performance,but also have better fusion performance in texture details.At the same time,the optimal values of various evaluation indexes are obtained,which proves that the algorithm can effectively improve the quality of Pan-sharpening.