Remote Sensing Image Fusion Based On Generative Adversarial Network

Remote sensing image fusion is one of the important branches in the field of image research.The fusion of multi-source remote sensing images with different image characteristics is of great significance for ground object recognition,disaster warning,land change monitoring,environmental quality assessment and other applications.This paper studies the fusion of high-resolution panchromatic images and multispectral images with rich spectral information in remote sensing images,and the fusion process is also known as "panchromatic sharpening".Compared with traditional image fusion methods,deep learning,which is hot in recent years,has a good performance in the research of remote sensing image fusion.Among them,the generated adversarial network has been widely used by many scholars in the research of image fusion due to its strong learning and generalization ability.Based on the analysis of the current situation of remote sensing image fusion at home and abroad,this paper proposes a remote sensing image fusion algorithm based on attention mechanism to generate adversarial network and a remote sensing image fusion algorithm based on unsupervised learning dual discriminator to generate adversarial network.The main work of this paper is as follows:(1)In view of the problem that the spectral and spatial structure information of source image is easily lost in the fusion result of traditional algorithm,the generation adversal network model is adopted to realize the fusion of multi-source remote sensing image.In this model,feature extraction is carried out on two source images respectively by means of two-stream branching.However,considering the uneven distribution of feature information during feature fusion,an attention mechanism module is added to the generator part of generating antagonistic network,which can re-assign weight ratios to extracted features according to the importance of feature information.It achieves the goal of highlighting important target information,removing redundant information such as noise artifact,and improving the perception effect of fusion image.(2)Single discriminator often causes the feature information of generated images to focus on one source image and ignore the image features of another source image.Therefore,based on the above model,a dual discriminator generation adversarial network model with unsupervised learning is proposed.The model no longer carries out supervised learning according to the simulated data,but takes two source images as the input of the generator and the identification criteria of the two discriminators.Unsupervised learning improves the learning generalization performance of the model,and makes full use of the spectral and spatial structure information of the source images.At the same time,in order to increase the stability of the model,the balance coefficient and gradient penalty term based on structural similarity are added to the loss function to ensure the dynamic balance between the generator and discriminator during adversarial learning.In addition,the spectral loss function and spatial structure loss function are also introduced in this paper,which improves the quality of fusion renderings to a certain extent.Through a series of experiments,it is proved that the fusion algorithm proposed in this paper has good performance both subjectively and objectively.