Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记

文章:Underexposed Photo Enhancement using Deep Illumination Estimation (DeepUPE) (CVPR2019)

github: https://github.com/wangruixing/DeepUPE

Loss: Reconstruction Loss + Smoothness Loss + Color Loss

摘要:该文章的任务是对曝光不足的图像进行增强。与之前的图像增强方法(直接学习图像到图像的映射)不同,DeepUPE学习(曝光不足图像)Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记与(illumination map)Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记的映射,这样可以更好地学习Grund Truth的复杂光照。DeepUPE根据光照先验建立Loss函数。

图像增强:图像增强任务可以看作寻找Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记与增强图像Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记的映射关系Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记,增强图像可以表示为Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记=Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记(Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记),在Retinex图像增强方法中,Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记的逆过程被建模为(illumination map)Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记。存在关系Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记=Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记,其中符号Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记表示基于像素的乘法,当已知Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记时,增强图像便可以求得Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记(Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记)=Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记。DeepUPE的任务就是通过网络学习Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记的映射关系。

网络结构:

Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记

首先,将Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记下采样后送入Encoder(采用VGG16的pre-trained模型),得到feature map;然后提取feature map的局部和全局features,将得到的features通过卷积层组合成低分辨率illumination;然后上采样得到Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记;最后结合Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记得到最终的增强图像Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记。训练网络时使用Reconstruction Loss (Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记)、Smoothness Loss (Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记)、 Color Loss (Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记)。

       Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记:重构loss定义为重构的增强图像Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记(Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记)与Ground Truth Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记 的L2误差,参考

       Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记:根据平滑先验,自然图像的光照通常是局部平滑的。在DeepUPE中引入平滑先验有两方面的优点:首先,可以防止过拟合,增加网络的泛化能力;其次,可以增强图像的对比度。

       Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记:将重构的增强图像Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记(Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记) 的每个RGB像素看作三维向量Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记,同样Ground Truth Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记对应的向量为Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记,他们之间的角度为Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记。其中Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记表示一个像素,Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记表示计算两个向量夹角的一种方法。颜色loss可以通过在颜色空间中计算像素的L2距离,之所以不采用这种简单的方法,是因为:一、重构loss已经使用L2计算了像素误差;二、L2进行数值度量,不能保证颜色向量的方向一致性,可能导致颜色的不匹配。

通过Figure.4可以对比不同loss的效果(作用):

Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记

实验:作者只提供了测试代码,如果想用自己的数据训练需要自己写训练代码。下图为使用自己的数据训练与测试的结果:

Underexposed Photo Enhancement using Deep Illumination Estimation(DeepUPE)笔记

其中,Input表示输入图像,DeepUPE Pre-train Model表示源代码模型结果,Ours为我们的训练测试结果。