2019/11/09 by Jae Woong Soh, Gu Yong Park, Soh, Jae Woong +5 · 1 citation
Computer Science · Engineering · #Advanced Image Processing Techniques #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image Processing Techniques and Applications #Image and Video Processing (eess.IV) #Integrated Circuits and Semiconductor Failure Analysis #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1911.03624
openalex publication_date 2019/11/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Recently, many convolutional neural networks for single image\nsuper-resolution (SISR) have been proposed, which focus on reconstructing the\nhigh-resolution images in terms of objective distortion measures. However, the\nnetworks trained with objective loss functions generally fail to reconstruct\nthe realistic fine textures and details that are essential for better\nperceptual quality. Recovering the realistic details remains a challenging\nproblem, and only a few works have been proposed which aim at increasing the\nperceptual quality by generating enhanced textures. However, the generated fake\ndetails often make undesirable artifacts and the overall image looks somewhat\nunnatural. Therefore, in this paper, we present a new approach to\nreconstructing realistic super-resolved images with high perceptual quality,\nwhile maintaining the naturalness of the result. In particular, we focus on the\ndomain prior properties of SISR problem. Specifically, we define the\nnaturalness prior in the low-level domain and constrain the output image in the\nnatural manifold, which eventually generates more natural and realistic images.\nOur results show better naturalness compared to the recent super-resolution\nalgorithms including perception-oriented ones.\n