A Pan-sharpening Method For Remote Sensing Image Combining Convolution Neural Network And Variational Model

In recent years,with the development of space infrastructure construction and remote sensing technology,remote sensing image technology has been widely used in various fields of national production and life.However,due to the limitations of signal transmission bandwidth and sensor hardware conditions,most satellites can not directly obtain high spatial resolution multispectral images that meet the needs of daily use.Since the information contained in panchromatic images and multispectral images has certain complementary characteristics,the fision of panchromatic images and multispectral images to obtain high-resolution multispectral images has become a research hotspot in recent years.The method researched and proposed in this thesis is aimed at a branch in the field of remote sensing image fusion,Pan-sharpening fusion.This thesis mainly combines the convolutional neural network and the variational model method for image fusion.Compared with the traditional fusion method,the remote sensing image fusion method based on the variational model can better balance the spectral information and spatial structure information of the fused image.However,the fusion method based on the variational model still has some defects,such as the method has many parameters,it also has high time complexity,and the modeling is difficult.In recent years,methods based on convolutional neural networks have been vigorously developed in the field of image fusion,and their fusion results have great advantages compared to traditional methods.This thesis proposes a Pan-sharpening method that combines convolutional neural networks and variational models.The main research contents are as follows:1.A panchromatic sharpening method for remote sensing images that combines convolutional neural networks and variational models.Since interpolation and upsampling will lose image information,this method first uses a convolutional neural network to perform super-resolution reconstruction of the original multispectral image to reduce information loss,and then uses the output multispectral image to construct a spectral fidelity term to extract the panchromatic image.The gradient information constructs the structural fidelity term.The optimization result of the variational model is closely related to the design of the fidelity term.The use of convolutional neural networks can avoid multiple up-sampling operations on the image,reduce the information loss of the image,and improve the quality of the fused image.The experimental results show that the method proposed in this thesis can achieve good results in visual observation and index evaluation,and the fusion image effectively retains the spectral information and spatial structure information.2.A prototype system for remote sensing image fusion.In this thesis,the fusion of convolutional neural networks improves the information of the spectral image as the fidelity term in the variational model,and the quality of the fused image is therefore greatly improved.Because remote sensing image fusion takes into account both research value and practical engineering significance,based on the research in this article,a prototype system for remote sensing image fusion is designed.