Single-Image Super-Resolution Analysis in DCT Spectral Domain

Authors : Onur AYDIN, Ramazan Gökberk CİNBİŞ

Pages : 209-217

Doi:10.17694/bajece.714293

View : 20 | Download : 10

Publication Date : 2020-07-30

Article Type : Research Paper

Abstract :Advances in deep learning techniques have lead to drastic changes in contemporary methods used for a variety of computer vision problems. Single-image super-resolution is one of these problems that has been significantly and positively influenced by these trends. The mainstream state-of-the-art methods for super-resolution learn a non-linear mapping from low-resolution images to high-resolution images in the spatial domain, parameterized through convolution and transposed-convolution layers. In this paper, we explore the use of spectral representations for deep learning based super-resolution. More specifically, we propose an approach that operates in the space of discrete cosine transform based spectral representations. Additionally, to reduce the artifacts resulting from spectral processing, we propose to use a noise reduction network as a post-processing step. Notably, our approach allows using a universal super-resolution model for a range of scaling factors. We evaluate our approach in detail through quantitative and qualitative results.
Keywords : deep learning, Super resolution, image process