BiUNet: Towards More Effective UNet with Bi-Level Routing Attention
Kun Dong (University of Chinese Academy of Sciences), Jian Xue (University of Chinese Academy of Sciences), Xing Lan (University of Chinese Academy of Sciences), Ke Lu (University of Chinese Academy of Sciences)*
The 34th British Machine Vision Conference
Abstract
The UNet-like structure has indeed emerged as the paradigm for medical image segmentation due to its excellent performance. However, many variants of UNet tend to be parameter-heavy or computationally complex, limiting their utilization of fast image segmentation in practical applications. In this paper, we propose BiUNet, a powerful and efficient model which well incorporates a lightweight attention module, Bi-Level Routing Attention (BRA). Besides, to compensate for the information loss caused by downsampling and further enhance the network’s performance, we introduce two innovative techniques termed pixel merging and pixel expanding, which are seamlessly integrated into BiUNet. Extensive experiments demonstrate that our model can achieve better performance than the latest networks with fewer parameters and lower FLOPs.Video
Citation
@inproceedings{Dong_2023_BMVC,
author = {Kun Dong and Jian Xue and Xing Lan and Ke Lu},
title = {BiUNet: Towards More Effective UNet with Bi-Level Routing Attention},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0482.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by