search

juglab / n2v

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

strange results with out-of-sample prediction #59

Closed benmuhlmann closed 4 years ago

benmuhlmann commented 4 years ago

Hi all, I'm running into some strange results when making predictions on images not used in training. Both images were produced from the same microscope setup, so the assumption is noise should be similar.

I've included python files used to train and predict, based on the 2D RBG examples given in the n2v repo.

image 251 was cropped then used to train the u-net. Then this model was used to make predictions for image 251, and another 'very noisy' image which is mostly out of focus.

The prediction on image 251 seems to have the effect of reducing some noise. The prediction on the 'very noisy image' seems to change the image's entire color map. This is the main isssue

https://drive.google.com/drive/folders/1h2C4qWxS7g0NeaiM6-Guy69zEbue-qhr?usp=sharing

fjug commented 4 years ago

Hi Ben,

after having a look at your data, multiple things need to be pointed out.

Screenshot 2019-11-20 19 02 21

Fig.1: left, image 251; middle: zoom into some area of image 251. Data is not subject to much pixel noises and looks rather nice! right: the second image you refer to. It is very different, actually not showing much at all, certainly not structures similar to the ones seen in image 251. Histograms show how pixel intensities in image 251 (left) and the second image (right) are distributed. Note the massive difference!!!



Anyways: here my main points for you:

  1. None of your images are particularly noisy.
  2. If they would be, I would strongly suggest NOT to train the RGB example (you have microscopy images, to RGB images... you even have single channel images!)
  3. The second image you have is, as you say, out of focus. This is not the same as being in-focus but noisy, you see?
  4. Even if you data would be subject to large amounts of pixel independent noise (this is what Noise2Void can address, as the paper states), the second image is not as the first. I have included some histograms for you to see that the data lives in entirely different intensity ranges. This alone would cause problems, even if all the rest would be fine.

I hope this helps a bit and maybe motivate you to have a look into our papers again. I'm sure N2V could be helpful for you, but you will need to train on images that are subject to pixel independent noises (e.g. shot noise) and then apply either to the same image or images that are similar in nature, not just some other images coming from the same microscope.

Best, Florian