作者：Yi Huang Xiaoguang Tu Gui Fu Tingting Liu Bokai Liu Ming Yang Ziliang Feng
在弱光条件下拍摄的图像往往能见度低，这会降低图像质量，甚至降低下游任务的性能。基于CNN的方法很难学习能够从各种未知弱光条件下的图像中恢复正常图像的通用特征。在本文中，我们建议将对比学习纳入照明校正网络，以学习抽象表示，从而区分表示空间中的各种弱光条件，目的是提高网络的可推广性。考虑到光照条件会改变图像的频率分量，在空间域和频率域中学习和比较表示，以充分利用对比学习。在LOL和LOL-V2数据集上对所提出的方法进行了评估，结果表明，与其他现有技术相比，所提出的算法取得了更好的定性和定量结果。
Images taken under low-light conditions tend to suffer from poor visibility,which can decrease image quality and even reduce the performance of thedownstream tasks. It is hard for a CNN-based method to learn generalizedfeatures that can recover normal images from the ones under various unknowlow-light conditions. In this paper, we propose to incorporate the contrastivelearning into an illumination correction network to learn abstractrepresentations to distinguish various low-light conditions in therepresentation space, with the purpose of enhancing the generalizability of thenetwork. Considering that light conditions can change the frequency componentsof the images, the representations are learned and compared in both spatial andfrequency domains to make full advantage of the contrastive learning. Theproposed method is evaluated on LOL and LOL-V2 datasets, the results show thatthe proposed method achieves better qualitative and quantitative resultscompared with other state-of-the-arts.
论文链接：http://arxiv.org/pdf/2303.13412v1
更多计算机论文：http://cspaper.cn/