Lightweight Image Super-Resolution with Scale-wise Network
Xiaole Zhao (School of Computing and Artificial Intelligence, Southwest Jiaotong University), Xinkun Wu (School of Computing and Artificial Intelligence, Southwest Jiaotong University)*
The 34th British Machine Vision Conference
Abstract
Recent advancements in Single Image Super Resolution (SISR) have been achieved by utilizing deep neural networks with a high number of layers. Incorporating multi-scale information is crucial in designing advanced super-resolution networks. In this paper, we present a novel feature upscaling method for SISR tasks which enhances the multi-scale information of images through multi-mode interaction. The multi-scale block design introduces more abundant image information, and the designed feature extraction module works in tandem with the multi-scale module to restore details in low-resolution images, thus enhancing image clarity. The static features and dynamic information in different scales of the image are fused and interacted through the local channel attention module to adjust the importance of different modes in recognizing different actions, thereby improving the final performance of the model. Numerous experiments confirm that our model achieves higher accuracy on Set5, Set14, B100, and Urban100 datasets compared to other state-of-the-art methods, while requiring relatively low computational and memory resources.Video
Citation
@inproceedings{Zhao_2023_BMVC,
author = {Xiaole Zhao and Xinkun Wu},
title = {Lightweight Image Super-Resolution with Scale-wise Network},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0286.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by