Pages: [1]   Go Down

Author Topic: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations  (Read 3138 times)

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com

Nobody testing DxO PureRAW seems to care about how much detail the neural network "invents", they just throw at it a noisy RAW file and look at the result -> WOW.
To find out we need a noiseless version of the scene and compare it with the result output from DxO PureRAW (DeepPRIME) when applied to a noisy version of the same scene.

Scene:
http://guillermoluijk.com/misc/dxoprimeescena.jpg

100% crops test (LEFT: noiseless capture, CENTRE: noisy capture, RIGHT: processed noisy capture):
http://guillermoluijk.com/misc/dxopureraw.jpg

Just comparing the output (right) with the noisy unprocessed capture (centre), the result is awesome, but when we look at the original scene some considerations have to be made:
  • Text masked by noise is cleaned, but its lines and traces are not recovered, as expected.
  • This is the most interesting spot: the neural network interprets and creates non existent edges and shaded facets, which are feasible looking at the noisy image but didn't exist in the scene
  • The neural network tends to simplify complex structures: the carvings on the leather mask are a series of curves but are interpreted more like linear shapes
  • Fine detail in the scene, completely lost in the noisy shot, is lost and interpreted as a plain colour area (with some fine gaussian-like grain at pixel level)
  • Flat colours are very well recovered, as would be with most noise reduction procedures

In the real world, this kind of software will be used without being able to compare with the real detail, so most photographers will consider valid all the fake detail recreated by the NN.
But watch out! those clean of noise feathers could belong to some other bird, not exactly the one you photographed.

Regards

« Last Edit: February 25, 2022, 03:17:48 pm by Guillermo Luijk »
Logged

digitaldog

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 20630
  • Andrew Rodney
    • http://www.digitaldog.net/
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #1 on: February 24, 2022, 07:03:50 pm »

Interesting test.
Indeed, the results look pretty good.
Indeed, not producing such noisy images is (when possible) a better approach.
GIGO:Garbage In Garbage Out still applies but I can see as you suggest, how people would say WOW.
Logged
http://www.digitaldog.net/
Author "Color Management for Photographers".

Rajan Parrikar

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3950
    • Rajan Parrikar
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #2 on: February 25, 2022, 02:42:48 am »

Did you disable the other corrections? My sense is that DxOPureRAW applies some sharpening even with those corrections unchecked. I use DxO Photolab instead for better control (to apply only DeepPrime for noise).

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #3 on: February 25, 2022, 05:19:15 am »

Did you disable the other corrections? My sense is that DxOPureRAW applies some sharpening even with those corrections unchecked. I use DxO Photolab instead for better control (to apply only DeepPrime for noise).
I did

http://guillermoluijk.com/misc/dxoprimesharpening.png

I have a feeling there is no hiden process here. The neural network could be as complex as it wants to be, but the output process is unique. Other programs like Topaz Denoise have many more options, which in my opinion are a disadvantage rather than an advantage, specially when its results are worse than those obtained with DxOPureRAW (the tiny texts for instance get highly corrupted in the output).

Regards

Rajan Parrikar

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3950
    • Rajan Parrikar
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #4 on: February 25, 2022, 07:23:24 am »

Thanks, Guillermo.

For less extreme cases of high ISO than the one you have analyzed, the DxO DeepPrime is outstanding. I hardly use Topaz now.

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
vs Topaz Denoise AI
« Reply #5 on: February 25, 2022, 12:12:16 pm »

Comparison with Topaz Denoise AI (RAW mode with default noise reduction and detail enhancement):

http://guillermoluijk.com/misc/dxopurerawvstopaz.jpg

I would stay with DxO PureRAW. Topaz seems to have a more black-box neural network output (I could see some Google Tensorflow libraries loading during the installation), going beyond in texture recreation but for the same reason making more mistakes and creating more fake detail. In the text test the result of Topaz is very bad, and the colour patches are noisier.

Regards

digitaldog

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 20630
  • Andrew Rodney
    • http://www.digitaldog.net/
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #6 on: February 25, 2022, 02:48:02 pm »

Downloaded the free demo.
Only tried one raw (DNG converted), the results are pretty impressive. The UI is super simple and I like that there are not a lot of options.
I have to assume the saved DNG is a linear DNG, it is much larger but based on the fact I'd only use this for some images and can still do a lot of work in Lightroom Classic or ACR, not a big deal.
I noticed a color difference between the original and new DNG. Well, the product somehow popped the wrong DCP profile onto the image! This was a Sony raw and when I opened the processed DNG, it had a Canon DCP profile. Toggle in Develop the correct profile, original color appearance as back.
Made a new DXO folder below the original which is fine I guess, makes it easier to move in LR to the folder above it I want. I see I can change that after processing.
Pretty slick, maybe worth $129. Maybe a company coupon will make me buy. I'll play for 30 days and see.
Logged
http://www.digitaldog.net/
Author "Color Management for Photographers".

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #7 on: February 25, 2022, 03:07:33 pm »

The UI is super simple and I like that there are not a lot of options.
I have to assume the saved DNG is a linear DNG, it is much larger
If low light pictures were my daily routine, I'd  happily pay those $129. The price of a cheap plastic lens to dramatically improve all my images at a click. I also consider an advantage not having to choose between any options; the algorithm just works, safely and without intervention (Topaz Denoise offers the classical bunch of options from which you need to choose).

The output DNG is linear and large because it's a demosaiced TIFF file, with the classical greenish colour just like RAW data. This means it has been demosaiced but not white balanced nor colour profiled, which is good.

http://guillermoluijk.com/misc/iso25600DEEP.jpg

Regards

digitaldog

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 20630
  • Andrew Rodney
    • http://www.digitaldog.net/
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #8 on: February 25, 2022, 03:11:09 pm »

Agreed, not something I'd use that often. But I tried this on one of the Sony captures at ISO 800, the results were impressive and for those few images, assuming I was going to do anything like print them, I'd apply the product happily.
Cheaper (but of course not superior) to an expensive fast prime lens.
Logged
http://www.digitaldog.net/
Author "Color Management for Photographers".

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
DxO PureRAW (DeepPRIME) DR enhancement
« Reply #9 on: February 28, 2022, 06:13:58 pm »

I wanted to numerically measure the DR enhancement provided by DxO PureRAW. I have shot an IT8 at ISO25600, measuring SNR over the 24 gray patches (I just used the G channel of a neutral RAW development done with DCRAW vs the linear extraction over DxO PureRAW's DNG):

http://www.guillermoluijk.com/misc/it8dxopureraw.jpg

The numbers say that all patches improve SNR (and hence DR) by aproximately 1,25EV. That means noise is 40% as large as in the original RAW files.
DR (SNR=12dB criteria) for the Sony A7 II at ISO25600 is enhanced by 1,9EV, from 4 stops to 6 stops:



I have to say I expected more, but this is what the calculations say (I repeated the numbers using Photoshop's info. window and got the same result). I think visually the result looks better for two reasons: DxO PureRAW eliminates all colour noise which is very annoying, but specially because of its closer to a uniform distribution (RAW noise is gaussian). This means noise in DxOPureRAW's DNG is very dense but never reaches deviations above a threshold. Looking at the histograms provided they look like truncated gaussian distributions; or three tiny ghosts as well:



Regards
« Last Edit: March 16, 2022, 06:56:35 am by Guillermo Luijk »
Logged

MauriceRR

  • Newbie
  • *
  • Offline Offline
  • Posts: 45
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #10 on: April 16, 2022, 05:51:07 am »

Nobody testing DxO PureRAW seems to care about how much detail the neural network "invents"
Hi,
I agree with this sentence. But here it's exactly what the step is supposed to do. And it's also what does every demosaicing algorithm.
On Pureraw, the whole demosaicing operation is done by the nn, and they include sensor-specific noise in the training (and maybe some artifacts, like fringing, glare, etc).
The denoising from deepprime is very well placed in the raw pipeline, where it works with a maximum of efficiency, already starting to deal with the noise before any demosaicing operation.
Maybe you already know this link :
https://blog.dxo.com/denoising-technology/

In practice, By experience, it's close to what you measured, 2EV increase in DR compared to a more regular algorithm like Adobe CR. The difference is also visible at all the iso stages, but high iso images is clearly where the algorithm outperform everything i saw on the market. One of the main drawbacks from a photographer's experience is that it can produce a kind of color quantization in challenging conditions or dark areas, which can be difficult to manage then in post production.
But clearly, it is worth the try, artifacts/drawbacks remain very well controlled
In practice, I also noticed the drastical gain that can produce the deepprime on old camera models (It made me change my mind on old images I never used considering noise).

in comparison, topaz tools (denoising, gigapixels, sharpen) are impressive too but are more prone to change the structure of the image and introduce some small artifacts. Ok, they deal with bitmap images, and it's another deal than raw datas.
I would recommend these tools at the end of the workflow, or at least when the final destination is known, in order to avoid the upsampling of these artifacts when upsampling a file.

« Last Edit: April 16, 2022, 06:29:54 am by MauriceRR »
Logged

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #11 on: April 20, 2022, 01:10:37 pm »

Maybe you already know this link :
https://blog.dxo.com/denoising-technology/
Good document, I have to read it closely. But if I got it right and DxO works always in the Bayer domain, I see no reason to produce demosaiced DNG's. Unless they are pretty sure their algorithm can perform the demosacing better than anyone else.

Regards

MauriceRR

  • Newbie
  • *
  • Offline Offline
  • Posts: 45
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #12 on: April 24, 2022, 06:40:54 am »

Good document, I have to read it closely. But if I got it right and DxO works always in the Bayer domain, I see no reason to produce demosaiced DNG's. Unless they are pretty sure their algorithm can perform the demosacing better than anyone else.

Regards

They are, and actually, it's far from everything else I saw on the market. The pipeline is very interesting. It's not a  matter of moire/aliasing/fine details as the algorithm from adobe, but they integrate noise reduction in their training.
I think they have several kind of raw and map of noise for each sensors/cameras (like dark noise/ current read noise, etc in their database). Its a clever way to leverage their database asset of raw pictures.
I think it's not working only with bayer, I think they also made versions for fuji x-trans. Maybe adaptable to other specific sensor (RWWB ?), but no much of tese sensors are really publc consummer products.
Logged

Guillermo Luijk

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2005
    • http://www.guillermoluijk.com
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #13 on: April 24, 2022, 08:51:49 am »

They are, and actually, it's far from everything else I saw on the market. The pipeline is very interesting. It's not a  matter of moire/aliasing/fine details as the algorithm from adobe, but they integrate noise reduction in their training.
I think they have several kind of raw and map of noise for each sensors/cameras (like dark noise/ current read noise, etc in their database). Its a clever way to leverage their database asset of raw pictures.
I think it's not working only with bayer, I think they also made versions for fuji x-trans. Maybe adaptable to other specific sensor (RWWB ?), but no much of tese sensors are really publc consummer products.
I read the document and indeed, they denoise and demosaice with the same NN (at least that's it what they claim). I don't see any difficulty in adapting their already trained algorithm to other 2x2 structures such as RWWB, RYYB,... but Fuji X-TRans has a completely different espatial pattern so I gues it would require specific training.

I'd like to create some synthetic RAW files to push DxO PureRAW to the limits in terms of moire,... but unfortunately my trial licence already expired. I couldn't process my Smartphone DNG files with it though, no idea about the reason since they are standard DNG Bayer files.

Regards

MauriceRR

  • Newbie
  • *
  • Offline Offline
  • Posts: 45
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #14 on: April 24, 2022, 03:23:09 pm »

You can try DXO photolab for a few days, it's not the same licence try, and actually, it even does a better job than pureraw, pureraw is a bit limited: you can control a bit better the "sharpenning" with photolab.
After that, the license is not so expensive. I would suggest dxo photolab license instead of pureraw for the reason above, even for DNG conversion only.
Logged

Rajan Parrikar

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3950
    • Rajan Parrikar
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #15 on: April 26, 2022, 02:46:34 pm »

You can try DXO photolab for a few days, it's not the same licence try, and actually, it even does a better job than pureraw, pureraw is a bit limited: you can control a bit better the "sharpenning" with photolab.
After that, the license is not so expensive. I would suggest dxo photolab license instead of pureraw for the reason above, even for DNG conversion only.

Yes, the PureRAW sharpening and/or detail enhancement is quite aggressive even after turning off the lens correction & sharpening options. I don’t know if this has been fixed in the new release which is a paid upgrade.

DeepPrime noise algorithm plays well within DxO Photolab 5 where the sharpening is optional. The sharpening in Photolab 5 (tailored to the body+lens combination) is very good but even here I generally have to pull the slider back from the default.

kers

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 4388
    • Pieter Kers
Re: DxO PureRAW (DeepPRIME) test on ISO25600 image and NN considerations
« Reply #16 on: July 02, 2022, 07:47:35 am »

chiming in on Pure Raw vs2...
I have used it now and indeed it works very well ; especially over 3200 asa are the results leaps better.

Working on a Nikon d850 is have some remarks and questions...

remark: it is a pity one cannot choose to have vignetting uncorrected apart from the sharpening.
It is distortion corrected and/or sharpening + vignetting as i see.


is this correct?:
The Pure Raw workflow results in a corrected DNG.  it will be the new starting point for use in Lightroom.
Corrections made in Lightroom on the DNG result in the same outcome ( apart from pure raw corrections) as if one would make those on the original NEF

So now i have two RAW files to keep: the pure Raw DNG and the original Nef...

Is there a way to extract the lightroom XMP form the DNG as a side car? ( it is now included in the DNG)
« Last Edit: July 02, 2022, 07:54:55 am by kers »
Logged
Pieter Kers
www.beeld.nu/la
Pages: [1]   Go Up