Super-resolution through gradient-based neighbor embedding

Abstract

By UHD TV appeared, research to restore the high-resolution video efficiently from low-resolution video has been studied. Example-based super-resolution is intended to increase the size of the image using the patch from the training database. Performance of the algorithm is depending on the size of the database. Hence, the reconstruction method through the synthesis of patches in the relatively small size of database has been studied; super-resolution through neighbor-embedding.Low-resolution patches are estimated with neighbor patches through weighting parameter, and then the target high-resolution patch is reconstructed by neighbor patches and the weighting parameter.Performance is changed significantly depends on the feature that is used to obtain neighbor. In this paper, super-resolution through gradient-based neighbor-embedding is suggested to improve spatial quality of an image. Hierarchical searching is to find the reasonable neighbor patches. First, searching by k-d tree is to find the patch based on the structure of the overall pixel values .Second, patches are searched by considering the gra-dient of patches. Not only similar gradient but also artifacts are searched. Detected patches are reconstructed via the principal component analysis (PCA). As a result, artifacts are removed. Neighbor-embedding algorithm assumes the geometric relationship between low-resolution patches and high-resolution patches, this relationship must be preserved during the PCA process. And patch is refined by decomposing finely in accordance with the texture of patch to strengthen the texture. Decomposition of the patch uses a method of threshold-based segmentation. Experimental results show the proposed method is effective in improving quality of an image than the other example-based super-resolution algorithm. Performance of the proposed method is superior to other methods in case of changing the size of database.