Lightweight Super-Resolution via Grouping Fusion of Feature Frequencies

The larger scale (deeper or wider) of the deep convolutional neural network, the better the performance, and the higher computing and storage capacity are required, which limits the application on resource-constrained devices. Lightweight (small number of parameters) super-resolution networks are highly needed. We propose a novel lightweight image super-resolution network based on grouping fusion of feature frequencies. Firstly, residual concatenation blocks are used to better transmit and fuse local features. Secondly, a hybrid attention block is used to combine the features of different cues and improve the expressiveness of features. Finally, a frequency feature grouping fusion block is applied to fuse the feature information of high frequency and low frequency, and improve the quality of super-resolution image restoration. The proposed network model has been trained on DIV2K dataset using the Pytorch environment and tested on standard Set5, Set14, B100, Urban100, and Manga109 test datasets. The experimental results show that the proposed network model is significantly superior to the other representative network models in terms of subjective visual quality and objective measurement of PSNR, SSIM, and LPIPS. © 2023 Institute of Computing Technology. All rights reserved.