Abstract
With the development of efficient Super-Resolution (SR), many CNN-based methods adopt re-parameterization techniques to accelerate inference while training a wider network. However, the wide feature maps often lead to difficulty in reaching convergence due to information redundancy. To expand network width with a positive effect on restoration quality, we propose a novel Wide Feature Projector (WFP) module to get more benefits from wider feature. Besides, previous attention structures were computationally complex and occupied much memory. To address this issue, we investigate the peak memory consumption of attention structures in order to design a Fast and Memory-Economic Attention (FMEA) module, which factorizes the element-wise attention map to speed up inference and minimize memory consumption. Consequently, we propose a novel efficient SR network, termed as Wide Feature Projection Network (WFPN), which achieves a compelling super-resolution performance, and consistently beats its competitors in terms of computation complexity and memory cost.Video
Citation
@inproceedings{Fu_2022_BMVC,
author = {Minghao Fu and Dongyang Zhang and Min Lei and Kun He and Changyu Li and Jie Shao},
title = {Wide Feature Projection with Fast and Memory-Economic Attention for Efficient Image Super-Resolution},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0615.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by