search

yu4u / noise2noise

An unofficial and partial Keras implementation of "Noise2Noise: Learning Image Restoration without Clean Data"
MIT License
1.08k stars 234 forks source link

Target Images #21

Open zchen002 opened 5 years ago

zchen002 commented 5 years ago

Hi, I'm a bit confused, what is the purpose of target images? are target images same as source images?

yu4u commented 5 years ago

In general denoising problems, noisy images and clean images are used for inputs and targets respectively in training. In noise2noise model, both inputs and targets are noisy images, which are the same image with different noises.

zchen002 commented 5 years ago

I got it now, thanks for answering.

tangsipeng commented 5 years ago

sorry,i still don't understand. how can i have one image that have different noises. unless i have a clean image that i can produce different noises

yu4u commented 5 years ago

You might take pictures of a static scene (e.g. very dark scene). Simply by doing so, you can get different noise images sharing unknown clean image. The other assumed examples can be found in the original paper.

tangsipeng commented 5 years ago

thank you, i got it

AllenJac commented 5 years ago

If we have one real image with different level random noise, we don't need to add Gaussian noise with zero-mean. How could we make sure the expectation unchanged, or the expectation is the clean image?

yu4u commented 5 years ago

In real use cases, we cannot know exact noise generation process, and thus we cannot make sure the expectation becomes a clean image. Simply we can infer what kind of statistics can be used and try it.