Reference Guided Remote Sensing Image Super-Resolution

In the field of remote sensing,monitoring the ecological and geological environment with high-resolution(HR)images for long time series is meaningful.However,it is difficult to capture high-resolution(HR)remote sensing images for long time series.Therefore,most previous works use single image super-resolution(SISR)algorithm to reconstruct HR images.However,the visual results of these methods are smooth due to the limited information in the single low-resolution(LR)image.Considering that there are usually multiple observations captured by different satellites for the same region at different times,the HR images can be used as references for the target LR images.Therefore,in this thesis,we exploit the reference-based SR(Ref SR)strategy for remote sensing images and propose two methods which achieve significant improvements compared with SISR methods.The main contributions of this thesis are as follows:1.We propose to perform RefSR for remote sensing images via introducing constraints on matched positions.First,we propose a position-encoding based transformer to match and transfer the HR reference details to the LR input in the feature space.The position-constraint strategy successfully improves the accuracy of matching process.Hereafter,we propose to adaptively fuse dense features from different scales and depths,which improves the feature representation for SR reconstruction.Experimental results demonstrate that our method outperforms state-of-the-art SR approaches at both 4× and 8× SR tasks on RRSSRD dataset.2.We propose to perform RefSR for remote sensing images via local-global searching.First,we construct a RefSR dataset for remote sensing images,which contains images with higher resolution than those in RRSSRD dataset.We introduce a pixel-wise alignment method to correct the non-rigid deformations between references and target images.Then we propose a loca-global transformer to find the most similar patches for each query patch at both local and global levels.The similar patches from different levels are the complementarities for each other.In addition,we propose a double-head feature extractor to enrich the feature representations.Experiments demonstrate that our method achieves the best SR performance on both RRSSRD dataset and our dataset.