BlurXterminator just got better

Stuart Taylor:
I must admit I remain genuinely baffled at the huge praise this is getting.

Here is a before/after on an image taken with an 80 mm scope. The difference is astonishing. The trick, I think, is low noise and oversampling.

I haven't (yet) tried the new variant but otherwise I wouldn't expect miracles from the thing. Let's see whether the new AI really doesn't mess with colours anymore.

@Stuart Taylor 
Try using NoiseXTerminator a couple times first to remove the smaller scale noise and then see if it has a more noticeable effect. I also have the sense that it helps to start with a low noise image.

I've also found BXT seems to work better (for me at least) when I partially stretch my image first. I shoot with an OSC - but I've tried it on a couple true color and false (narrowband) color images and it seems to have the effect I expect.

Ok, that's interesting. I understood that (slightly contrary to my expectations) BlurXT should be run before noise reduction. I also gathered it should be done on linear data (see here)

I'll try your suggestions.

Stuart Taylor:

@Stuart Taylor 
Try using NoiseXTerminator a couple times first to remove the smaller scale noise and then see if it has a more noticeable effect. I also have the sense that it helps to start with a low noise image.

I've also found BXT seems to work better (for me at least) when I partially stretch my image first. I shoot with an OSC - but I've tried it on a couple true color and false (narrowband) color images and it seems to have the effect I expect.

Ok, that's interesting. I understood that (slightly contrary to my expectations) BlurXT should be run before noise reduction. I also gathered it should be done on linear data.

I'll try your suggestions.

I know. I read both of those things too, but decided to wildly experiment I guess. All I can say is it seemed to work the best for
me by not following the instructions!

Arun H:
he trick, I think, is low noise and oversampling

Absolutely true.  Then given those two basic conditions it works superbly -- better than PI deconvolution under nearly all circumstances -  and no more or less of any problem with colours  than standard deconvolution if you follow the same rules and just  work with the luminance.

Having tested it on many images now I think that this software deserves all the praise that it is getting  and more.   Especially in combination with NoiseXterminator and StarXt.   Together all hese tools provide a step change in processing.  After applying them I am seeing images based upon just 6h of data aquisition that have similar SNR and better sharpness  than images with 18h of data but processed only using the old tools .

Honestly, I image at 2.02/px and BxT is insane on my images. My Pillar's became razor sharp, and I was able to release a 200% crop from an 11MP camera. Nuts.

The update is even more significant for MacOS users like myself.

Arun H:

Tim Hawkes:
Together all hese tools provide a step change in processing.

I've not found NoiseX to be too useful. I can get similar or better results with usual PI tools and workflow and better control.

But StarX and BluX are absolutely gamechangers. There was a fear when BlurX came out that it would make everyone's images look the same. No, it doesn't do that. Quality of source data matters a lot, like it did before. But it does allow you to do more with your data, as any good tool should.

I

I started using iterative NxT passes in amounts around .4 and it was better. I have a modified Rista-Noise method that has done me well for years, and I only find NxT does better if I do it in small passes at a time.

Yuxuan:
I think the issue you have is the image is too noisy. The SNR is too low other than at the stars. BlurX doesn't have a lot of signal to extract details from. Meanwhile Arun's "before" image was much smoother, and encoded a lot of fine details that are not easily visible by human eyes.

I agree. BlurX is not magic, it is just better deconvolution. Things that make deconvolution more effective will make BlurX more effective. Here is some useful detail from Zeiss's excellent guide to deconvolution:

"What are the general recommendations for image acquisition settings if the resulting images are meant to be processed with a deconvolution algorithm?

The key to best results for deconvolution is to present an optimal input image to the algorithm. And the optimal image is the one with the highest possible signal and the lowest possible noise. In addition, the sampling rate, which in a confocal system then determines the number of pixels in the resulting image, must be high enough. What is high enough? The sampling rate must result in an image where the size of each pixel is less than half of the size of the structure you want to resolve, provided that the system supports this sampling. The theorem of such a sampling is called Nyquist–Shannon criterion for sampling. This means, you need to know the resolution of the optical system to get the sampling right – especially if the sampling can be changed, like on a confocal system. Typically, your confocal system provides information on the current sampling rate for a given acquisition setting. The sampling can then be changed to comply with Nyquist or even go beyond (like 2 × Nyquist for Airyscan, as detailed below). When the acquisition is performed with a camera, the optical system is designed to comply with Nyquist and settings don’t have to be changed. However, you need to be careful when you increase the size of the pixels by binning. This method then requires special consideration, and you must know the resolution of your system to be able to predict potential consequences."

Tim Hawkes:

Stuart Taylor:
What am I doing wrong?

Just a possibility here?   You say that the sampling is 0.96 arcsec/ pixel but not what the resolution of the image already is that you are trying to deconvolute?

It could simply be perhaps that your image is already sharp and that there is nowhere to go - you have already done a great job?   So what is the estimate of FWHM of the linear image that you were trying Blur Xt  on?    If the FWHM is already down at say 1.7 or 1.8  arcsec or something then actually 0.96 is close to undersampling.  Under such good seeing it would be necessary to sample more for BlurXt to do much.

Another related possibility perhaps is that your image was drizzled - which again would mean that it might already be close to the limiting resolution that he sampling could provide?    The SNR also perhaps a bit low too to see the optimum effect?

Tim

I'm not sure I follow. The image scale is 0.96"/px (as in 3.76μm pixels and 808mm focal length). What other type of resolution are you asking about? (forgive my ignorance here, but I thought image scale is the same as resolution.

The mean FWHM is 2.6 px and the image is not drizzled.

andrea tasselli:

Stuart Taylor:
I must admit I remain genuinely baffled at the huge praise this is getting. All it does for me is reduce stars (and I already have many ways to do that). It doesn't do anything to nebulosity. Here is a before and after using the latest AI (v 2). I've even tried the suggestion of creating a PSF for the image and plugging in the FWHM value. Makes no difference.

What am I doing wrong?

The likely answer is: nothing. It is that BX isn't the panacea for all ills that some here think it is nor does it transform you images in Hubble-class ones. It works well for what it does (me thinks that it has the main advantage of not introducing too much noise when deconvolving compared to other methods) but some subjects don't lend themselves well to that kind of AI-powered improvements. But I fear I'm preaching to the wind here...

Thanks Andrea. This does sound likely.

Yuxuan:
I think the issue you have is the image is too noisy. The SNR is too low other than at the stars. BlurX doesn't have a lot of signal to extract details from. Meanwhile Arun's "before" image was much smoother, and encoded a lot of fine details that are not easily visible by human eyes.

But my understanding is that BlurXT should be done before de-noising.

Arun H:

I've also found BXT seems to work better (for me at least) when I partially stretch my image first

To the point about stretching before applying BlurX - here is why, from the same Zeiss article, deconvolution requires a linear image:

"Deconvolution wants to pick intensities from the observed image and re-assign them, by convolving the inverse of the PSF"

A nonlinear operation such as stretching arbitrarily and irreversibly changes the relationship between the intensities in the image. So the algorithm may work "better" in the sense of giving you a sharper image, but you've lost connection with the original image through the intermediate nonlinear operation.  Of course, you may or may not care about this...

Just for fun I extracted the PSF for a mid-range star before and after stretching (hear a mild HT followed by AS):

Before: fwhm-x=4.12, fwhm-y=3.39
After:    fwhm-x=10.52, fwhm-y=9.86

The effective consequences of using deconvolution on a stretched PSF is that the noise floor is much closer to the peak intensity and the overall FWHM is much larger thus creating the false effect that the amplitude of the signal is much weaker and at the same time the blur is much larger, IMO.