基于双重迭代的零样本低照度图像增强

针对低光照条件下拍摄图像质量低下的问题，该文提出一种基于双重迭代的零样本低照度图像增强方法。其外层迭代通过卷积神经网络估计增强参数，再由内层迭代进行图像增强，增强结果进一步用于计算损失函数并反馈更新外层的参数估计网络，最终通过多轮迭代生成高质量的图像。在该框架下，还设计了多尺度增强系数估计模块、基于注意力的像素级大气光估计模块，并提出了基于亮度对比度、大气光、颜色均衡以及图像平滑性先验的无监督损失函数。大量实验结果表明，该方法可有效将低光照图像增强为高质量的清晰图像，其性能优于现有的同类方法。同时该方法基于零样本学习，不需任何训练数据集，具有良好的普适性。

Zero-shot Learning for Low-light Image Enhancement Based on Dual Iteration

In this paper, a novel zero-shot low-light image enhancement framework is proposed based on dual iterations. The outer iteration uses a network to estimate the enhancement parameters, with which the inner iteration improves actually the image, and the results are applied to calculating the loss functions and updating the outer network. After multiple rounds of iterations, high-quality images can be obtained. Within this framework, an adaptive parameter estimation module and an attention-based pixel-wise atmosphere estimation module are designed. In addition, unsupervised loss functions based on light, contrast, color balance and image smoothness priors are proposed. Experiments demonstrate that the proposed method obtains high quality clear images from low-light ones, and outperforms state-of-the-art methods. Furthermore, the proposed method belongs to zero-shot learning, which does not need training dataset and thus can be widely applied.