https://github.com/juglab/n2v

**Noise2Void - Learning Denoising from Single Noisy Images, iccv19,**

1. no proof in the paper.
2. results are not amazing.
3. But it is still important for not using other info and priors, such as noise type.

A conventional network versus our proposed blind-spot network. (a) In the conventional network the prediction for an individual pixel depends an a square patch of input pixels, known as a pixel’s receptive field (pixels under blue cone). If we train such a network using the same noisy image as input and as target, the network will degenerate and simply learn the identity. (b) In a blind-spot network, as we propose it, the receptive field of each pixel excludes the pixel itself, preventing it from learning the identity. We show that blind-spot networks can learn to remove pixel wise independent noise when they are trained on the same noisy images as input and target.

Actually conventional supervised network use GT value instead the input as target.

(b) is implemented by masking!

UPS: a masking procedure for neural network

1. input: values x of a r × r patch / neighborhood / region centered at j;
   1. how to mask xj? by Uniform Pixel Selection (UPS)
      1. replacing xj with xk, k is a pixel chosen uniformly at random from the patch;
2. f: a UNet or DnCNN;
3. output: a patch of values, and the loss is only for pixel j, i.e. only a noise-free value for xj is concerned;
4. pseudo GT / target: xj

Notions

The smaller the patch used for selecting the pixel, the larger the contribution of the noise will be

Noise2self use the average of x excluding xj.

Experiments

Performance of N2V on the BSD68 dataset compared to various baselines. 干不过N2N和supervised deep methods.