Research On Satellite Image Inpainting Based On Generative Adversarial Network

Due to mechanical jitter of satellite and radiation in the space environment,satellite images may have certain defects,such as dead pixels and lines,etc.These defects may affect the application of satellite images in radiation transmission calculation and numerical model prediction.Therefore,the inpainting of satellite images becomes an important task.Although many researchers have done a lot of researches on satellite image inpainting at home and abroad,the current inpainting performance are still unsatisfactory.Aiming at the above-mentioned problems,we conduct researches based on satellite images collected by the domestic Fengyun-1 satellite.The current image inpainting models based on single explicit structure inference have achieved good inpainting results,but these models are prone to structural inaccuracy and blurring due to inaccurate edge prediction.In order to solve this problem,this paper proposes a satellite image inpainting model Structure Connect based on a variety of explicit structural reasoning and generative adversarial networks.The Structure Connect model divides the image inpainting into two stages.First,it explicitly reasons the edge and threshold segmentation of the image,and then combines these structural information to repair the image.The newly proposed threshold segmentation image repair network based on the generative adversarial network not only alleviates imbalanced samples and poor continuity of the edge prediction network,but also provide more accurate structural information.Experiments show that at different noise ratios,the repair results of the threshold segmentation image repair network are more accurate than the edge repair network,and the PSNR of the Structure Connect model is improved by about 0.2db compared with the benchmark model Edge Connect.In order to make better use of the multiple channels of satellite images for image inpainting,this paper proposes an exemplar-based generative adversarial network model Ex UGANs.Ex UGANs uses a multi-task learning framework,which simultaneously learns image inpainting task and the identity mapping task of reference image.Specifically,the Ex UGANs model designs a structure of a shared encoder and two independent decoders,so that the encoder can better learn the feature representation of the reference image to improve the generalization ability of the model.In order to alleviate the information loss of the bottleneck and the inconsistency between the background and foreground,this paper also introduces the U-Net structure and gated convolution to help the model retain more useful information.Experiments show that at different noise ratios,the PSNR of the Ex UGANs model is improved by about 1.5db compared to the benchmark model Ex GANs and measures mentioned above all help to improve the result.