search

juglab / n2v

This is the implementation of Noise2Void training.
Other
387 stars 107 forks source link

Noise including crumple and creases in images #62

Closed harnitbakshi closed 4 years ago

harnitbakshi commented 4 years ago

Hi I had a question regarding the model, is it able to remove noise such as crumpling or creases as well? If not then what kind of noises can it remove

Thanks

tibuch commented 4 years ago

Hi @harnitbakshi

Noise2Void has to be trained on pixel-wise independent noise. Crumpling or creases span multiple pixels, which makes them not pixel-wise independent. Since N2V has only one picture to train, it is impossible for the network to distinguish between real structure and noise patterns. We discuss this limitation in our paper (https://arxiv.org/pdf/1811.10980.pdf) in Section 4.4 Errors and Limitations.

This being said, the network architecture is able to learn the removal of non-pixel-wise independent noise. But you would need a pairs of training images where the input image has crumpling or creases and the target images has no such patterns. Figure 6 in the paper illustrates this well.