Application Of Image Super Resolution Algorithms In Remote Sensing Images

The need for high-resolution imaging becomes particularly important in remote sensing image applications such as ground-object identification.High resolution remote sensing image has high quality and clarity,and thus it is useful in applications.In this paper,image super-resolution algorithms and their applications in remote sensing images are studied.The main work includes the following three aspects:(1)We investigate the development of image super-resolution algorithms,and introduce the popular deep learning algorithms in detail.(2)We used a multi-frame image SR algorithm based on the Bayesian framework to improve the resolution of images acquired by a micro-nano carbon satellite.The algorithm estimates noise,motion,and blur kernel parameters in the outer iteration and reconstructs an HR image in the inner iteration.We worked on improving the motion estimation component of the algorithm for our images and reducing the estimation error of this component.Our experimental results indicate that our algorithm is superior to similar algorithms during both simulation and actual application.Further,the generated HR images exhibit better quality than the original images(3)We used a SISR algorithm based on a GAN to improve the resolution of images acquired by the Sentinel-2 satellite.The network contains a generator and a discriminator.In the generator,we first use an Inception like multi scale structure to extract the image features.Then we use residual blocks to learn the non-linear mapping between low resolution features and high resolution features.Finally we use a subpixel layer to reconstruct high resolution images.The discriminator is mainly used to determine whether the input image is a real picture or a computer generated image.By the adversarial learning between generator and discriminator,we can obtain high resolution images which is similar to real world images.In the adversarial learning,we use a loss function which combines perceptual loss,adversarial loss,regularization loss and mean square error.In the experiment,our proposed obtains better results than similar methods,as our experimental results indicate.F,due to the influence of the atmospheric,some input satellite images are hazy.In order to improve the resolution of these hazy images,we designed the Dehaze-SenSR network.Compared with the dark channel priori based dehazing algorithm,the network can improve the resolution of the input image while effectively removing the haze in the images,and the details of the generated image are more abundant.