Getting plate-solving status, please wait...
AI denoised Hoag's object, Benoit Blanco

AI denoised Hoag's object

Getting plate-solving status, please wait...
AI denoised Hoag's object, Benoit Blanco

AI denoised Hoag's object

Equipment

Loading...

Acquisition details

Loading...

Description

I reprocessed this Hubble project from scratch using conventional processing and a deep learning denoising method called Deep Prior.

The trendy way to denoise images with machine learning is to train GANs (like the well-known DeepArt) with a lot of examples. It will be very efficient for smartphone pictures, but very dangerous for scientific data as it can introduce structures from another dataset.



I used a deep convolutional network called DeepPior to denoise the Hoag's object. This algorithm does not search for the answer in the image space like a trained GAN would do, it looks for the answer in the space of the neural network parameters. It is like a brain that trains itself to recreate that image without noise. There is no training, no apriori information so it's safe for the data.

I added Nasa version for comparison. (C)

I also performed a SuperResolution (D)

Processing sequence in Pixinsight (Originale) :

Crop, Stamp correction around background galaxies, manual dynamic alignement of background galaxies, stack all data in Luminance, Stack 814, Stack 450, Stamp correction cleaning of each stack, RGB (814, 606, 450), DBE, Assisted color calibration, Stretch, Luminance Merge, Deep Prior restoration of Luminance 11000 iterations, LRGB, Local Histo, Deconvolution, resample, Camera Raw and cosmetics in Photoshop.

This image was featured in Nasa APOD : https://apod.nasa.gov/apod/ap191127.html

Comments

Revisions

  • Final
    AI denoised Hoag's object, Benoit Blanco
    Original
  • AI denoised Hoag's object, Benoit Blanco
    C
  • AI denoised Hoag's object, Benoit Blanco
    D

C

Description: Nasa Version

Uploaded: ...

D

Description: My reprocessing

Uploaded: ...

Histogram

AI denoised Hoag's object, Benoit Blanco