Single Image Super-Resolution Network with Enhanced Octave Convolution for Separating Image Frequencies

In recent years, deep learning-based algorithms for image super-resolution (SR) have been increasingly thriving. However, in most existing algorithms, the model takes the whole low-resolution (LR) image as input and treats the low-frequency (LF) and high-frequency (HF) contents in the image equally. Considering the nature of Bicubic down-sampling, which is the most used degradation in SR models, high-resolution (HR) and LR images share the most part of the LF in the image while differing in HF. To this end, this paper proposes a model that separates frequencies in the image with an enhanced octave convolution (EOctConv) layer. EOctConv applies channel attention to highlight HFs in the image while considering the relation between the weighted HFs with LFs, respectively. Separating image frequencies not only eases the learning of the relation between LR and HR images, it can significantly enable the model to focus most of the computation on recovering the lost details. The proposed model incorporates significantly fewer numbers of parameters which is indeed a remarkable factor to considerably reduce the computation time and complexity yet compares favorably with recent results in literature qualitatively and quantitatively on benchmark datasets.