目录
发表在2019 CVPR。
摘要
The field of image denoising is currently dominated by discriminative deep learning methods that are trained on pairs of noisy input and clean target images. Recently it has been shown that such methods can also be trained without clean targets. Instead, independent pairs of noisy images can be used, in an approach known as NOISE2NOISE (N2N). Here, we introduce NOISE2VOID (N2V), a training scheme that takes this idea one step further. It does not require noisy image pairs, nor clean target images. Consequently, N2V allows us to train directly on the body of data to be denoised and can therefore be applied when other methods cannot. Especially interesting is the application to biomedical image data, where the acquisition of training targets, clean or noisy, is frequently not possible. We compare the performance of N2V to approaches that have either clean target images and/or noisy image pairs available. Intuitively, N2V cannot be expected to outperform methods that have more information available during training. Still, we observe that the denoising performance of NOISE2VOID drops in moderation and compares favorably to training-free denoising methods.
结论
We have introduced NOISE2VOID, a novel training scheme that only requires single noisy acquisitions to train denoising CNNs. We have demonstrated the applicability of N2V on a variety of imaging modalities i.e. photography, fluorescence microscopy, and cryo-Transmission Electron Microscopy. As long as our initial assumptions of a predictable signal and pixel-wise independent noise are met, N2V trained networks can compete with traditionally and N2N trained networks. Additionally, we have analyzed the behaviour of N2V training when these assumptions are violated. We believe that the NOISE2VOID training scheme, as we propose it here, will allow us to train powerful denoising networks. We have shown multiple examples how denoising networks can be trained on the same body of data which is to be processed in the first place. Hence, N2V training will open the doors to a plethora of applications, i.e. on biomedical image data.
要点
本文的思想来源于noise2noise工作。n2n不需要干净图像的参与，而这篇工作进一步加强：甚至不需要加噪图像对（noisy image pairs）。类似于块匹配，N2V只需要输入图像本身。
这种方法特别适用于数据稀缺的任务，如医学图像处理。
N2V方法并不能做到SOTA（尤其是和有监督学习方法相比），但是和不需要训练（training-free）的方法相比，性能没有下降太多。
假设噪声与像素位置无关。
博主没来得及看完全文，但根据这幅图以及题注，我的猜测是：
原来N2N需要加噪图像对。即对同一幅干净图像，随机加噪得到一对有噪图像。保证噪声分布相同并且是零均值噪声，即有噪图像的期望与干净图像的期望相同。
现在，针对有噪图像的每一个像素点，我们都取其空心邻域。我们可以大致假设：该空心邻域的期望也和干净图像的期望是相同的。这样，既然N2N也能工作，那么N2V也是有效的。
在实际操作中，该像素点会被一个随机值（从邻域中选择）取代，然后完成卷积。如果直接学习有噪图像到其本身的映射（不抠掉中心），那么网络会趋近于一个恒等映射函数，失去了意义。