Joint Low-light Enhancement and Super Resolution with Image Underexposure Level Guidance
Mingjie Xu (Beihang University), Chaoqun Zhuang (Beihang University), Feifan Lv (Beihang University), Feng Lu (Beihang University)*
The 34th British Machine Vision Conference
Abstract
Obtaining high-quality images with high resolution in poor illumination environments using a limited spatial resolution image sensor poses a significant challenge. Low-light Enhancement (LLE) and Super-Resolution (SR) are crucial technologies for overcoming this challenge. However, current approaches usually generate normal-light high-resolution images with non-uniform brightness and loss of details from low-light low-resolution images, and suffer from significant performance degradation in cross-dataset settings. To alleviate these problems, we propose a novel solution for low-light image super-resolution. For non-uniform brightness problem, we propose a Relative Underexposure Level Estimation Module (RUL-EM) that estimates the relative underexposure levels of input images to adjust the image brightness to a uniform level and avoid artifacts. For detail loss and cross-dataset problems, we introduce the Multi-Scale Sampling (MSS) strategy for sampling multi-scale patches. MSS involves randomly cropping low-light and low-resolution patches of different sizes and positions and resizing them to a given patch size. Combining RUL-EM with MSS can improve the model performance in detail restoration and generalization. Additionally, we also incorporate channel attention to enable the Joint LLE & SR Network (JLSN) to adaptively adjust the influence of estimated relative underexposure levels. Our proposed method can be applied to various backbone architectures. Experimental results show that our proposed method achieves state-of-the-art performance on the joint LLE & SR task in both within-dataset and cross-dataset settings. Our proposed solution can convert low-resolution low-light images into high-resolution images with satisfactory brightness, vivid colors, and more details.Citation
@inproceedings{Xu_2023_BMVC,
author = {Mingjie Xu and Chaoqun Zhuang and Feifan Lv and Feng Lu},
title = {Joint Low-light Enhancement and Super Resolution with Image Underexposure Level Guidance},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://bmvc2022.mpi-inf.mpg.de/BMVC2023/0046.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by