Is there an objective comparison of proprietary noise reduction techniques with published algorithms? - Forums

Topic starter

Aug 24, 2022 · Edited: Aug 24, 2022

TopazDenoise and NoiseXterminator seem like the denoising flavour of the month at the moment.

I've recently been looking at a range of modern high performance denoising algorithms (non-local means (NLM), block matching and 3D filtering (BM3D), K-SVD, profuse clustering technique (PCT) etc.) with a view to using them for astro image denoising. These algorithms are well studied academically and have numerous papers analysing their performance in terms of standard denoising measures (peak SNR, SSIM, FOM). Some even have open source implementations rather than just algorithm descriptions.

However I've not been able to find any scientific studies comparing the denoising performance of Topaz and NX with these academically characterised algorithms. Does anyone know of some papers that I've missed? I know people love Topaz and NX subjectively, but I would have thought that if these proprietary algorithms were actually better than anything else available, their creators would want to shout about it and publish performance comparisons.

Well Written Insightful Respectful Concise Engaging

Frédéric Auchère

Aug 24, 2022

Hi Adrian,

Interesting question, and I don't have an answer.
I'm still using a relatively old-school wavelet-based 'à la' Starck & Murtagh denoising scheme in my DIY processing S/W. I'm not up to speed with the most recent stuff and should definitely catch up with the literature.

Frédéric

Well Written Respectful

Andy Wray

Aug 24, 2022 · Edited: Aug 24, 2022

Not scientific, but I paid the full license for TopazDenoize a year ago and used to use it for astrophotography. I will continue to use it for terestrial photography. However, NoiseXterminator is far, far better for astrophotography denoising. Using Topaz you will always be fighting with the stars in an image, whereas NoiseXterminator knows just how to handle them.

They are both AI-based, but one has been trained for normal photography (Topaz) and the other for astrophotography (NoiseXterminator).

One last point: the default 0.9 value in NoiseXterminator results in a bit too creamy result for my liking … 0.5 to 0.6 is the sweetspot for most of my images.

Helpful Concise

Jared Holloway

Aug 24, 2022

Andy Wray:
Not scientific, but I paid the full license for TopazDenoize a year ago and used to use it for astrophotography. I will continue to use it for terestrial photography. However, NoiseXterminator is far, far better for astrophotography denoising. Using Topaz you will always be fighting with the stars in an image, whereas NoiseXterminator knows just how to handle them.

They are both AI-based, but one has been trained for normal photography (Topaz) and the other for astrophotography (NoiseXterminator).

One last point: the default 0.9 value in NoiseXterminator results in a bit too creamy result for my liking ... 0.5 to 0.6 is the sweetspot for most of my images.

This 100%

Die Launische Diva

Aug 25, 2022 · Edited: Aug 25, 2022

Yet we don't know if NoiseXterminator does something novel, or it was just trained by an expert with years of knowledge using simpler techniques. Maybe, under the hood it is just one traditional denoising algorithm applied in a starless version of the input image and packed in a convenient GUI. For most of astrophotographers, unintuitive GUI controls and lack of real-time previews is PixInsight's Achilles heel. We have powerful algorithms in our avail but the lack of real-time previews and plethora of parameters can lead to sub-optimal results. In other words, a poor denoising result may just be the result of human error after using a complex GUI. That's why we need objective comparisons.

Well Written Insightful Respectful Concise Engaging

Tom Dinneen

Nov 5, 2022

My guess is that NoiseXTerminator / DenoiseAI are Encoder-Decoder based neural networks. Just like Starnet++. Free implementations exist. There’s nothing novel in there use for denoising. However it requires configuring, a lot of tweaking and even more training to do anything useful. That’s where the real value add is and what you are paying for. On top of that NoiseXTerminator adds noise back in to where the star is removed to better blend it in with the rest of the background. It does that very very well. Almost seamless. That’s something completely unique to NoiseXTerminator.

Tom

Helpful Insightful Respectful