期刊名称：Balkan Journal of Electrical & Computer Engineering
印刷版ISSN：2147-284X
出版年度：2020
卷号：8
期号：3
页码：209-217
DOI：10.17694/bajece.714293
语种：Turkish
出版社：Kirklareli University
摘要：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.
