基于流模型的低照度超分辨率重建

Low-Illumination Super-Resolution Reconstruction Based on Flow Model

将低照度低分辨率图像增强到正常曝光高分辨率图像具有高度的不确定性,即它们之间的映射关系是一对多的.以往基于像素重建图像和确定图像的过程未能捕捉正常曝光图像的复杂条件分布,导致不适当的亮度、残余噪声和伪影.在本文中,通过一个已经提出的标准化流模型来构建这种一对多的关系.一种以低照度图像的特征为条件,学习将正常曝光图像的分布映射为高斯分布的可逆网络模型,这样可以很好地模拟正常曝光图像的条件分布.可逆网络模型的另一个优点是在训练过程中被一个描述正常图像流形结构的损失函数约束.在LOL(Low-light)数据集上的实验结果表明,该方法相比较其他方法,获得了更高峰值信噪比(PSNR),更高结构相似度(SSIM),更少学习感知图像块相似度(LPIPS).

The enhancement of low-illumination low-resolution images to normal-exposure high-resolution images has a high degree of uncertainty,the mapping relationship between them is one-to-many.Previous pixel-based image reconstruction and image determination processes failed to capture the complex conditional distribution of normally exposed images,resulting in inappropriate brightness,residual noise,and artifacts.In this paper,this one-to-many relationship is modeled by a proposed normalized flow model.A reversible network model that learns to map the distribution of normal exposure images to Gaussian distribution conditioned on the characteristic of low-illumination images,which can well simulate the conditional distribution of normal exposure images.Another advantage of the invertible network model is that it is constrained during training by a loss function that describes the normal image manifold structure.The experimental results on the LOL(Low-light)dataset show that this method achieves higher peak signal-to-noise ratio(PSNR),higher structural similarity(SSIM),and less learned perceptual image patch similarity(LPIPS)than other methods.