Revisiting Single Image Super-Resolution Under Internet Environment: Blur Kernels and Reconstruction Algorithms

摘要

Due to limited network bandwidth, the blurred and down-sampled high-resolution images in the spatial domain are inevitably used for transmission over the internet, and so single image super-resolution (SISR) algorithms would play a vital role in reconstructing the lost spatial information of the low-resolution images. Recently, it has been recognized that the blur kernel is crucial to the SISR performances. As most of the existing SISR methods typically assume the blur kernel is known, and in fact the blur kernel is either fixed with the scaling factor or unknown, it thus would be of high value to investigate the relationship between blur kernels and reconstruction algorithms. In this paper, we first propose a fast and effective SISR method based on mixture of experts and then give an empirical study on the sensitivity of different SISR algorithms to the blur kernels. Specially, we find that different algorithms have different sensitivity to the blur kernels and the most suitable blur kernels for different algorithms are different. Our findings highlight the importance of the blur models for SISR algorithms and may benefit current spatial information coding methods in multimedia processing.