Ad
Ad
Ad
Pages: [1] 2 3 »   Bottom of Page
Print
Author Topic: Noise reduction by channel  (Read 6353 times)
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« on: January 19, 2013, 04:03:30 PM »
ReplyReply

I have tested a new method (as far as I know) for Noise Reduction. Unless I hear its published already I will call it Arthur's method.  Wink

Here is a link from Sony A99 @6400 as published in RAW by Imaging Resource. All rights to the RAW are theirs. I have done a conversion using NR by RGBL channel.
http://fs02n2.sendspace.com/dl/a9f06e0a4aedbad8c9510ed701cd51d6/50fb12b408dba234/cibx6j/CVT_AA99hSLI06400NR0%20NR%20Curves.jpg

This file has a lot less noise than their jpg version. Basically the idea is to treat each of the 4 channels as a 3 dimensional space for Noise Reduction. If you end up with mild patterns like the attached conversion you can break the original into CMYL do noise reduction on each channel then save. Combine both results in layers or by math operation average.

Your noise reduction software running on a powerful computer is far more capable than a simple de-bayer routine. The raw converters available are very good on low ISO. They seem to produce a lot more noise than this method on high ISO files. I did my work in ImagesPlus, a scientific imaging program

I am not in the imaging business so I am not going to try to develop this somehow. Anyone can use it. If someone puts it in their blog, book, document, whatever, I expect it to be called Arthur's method referencing Lula and Dyxum. I worked on it for a couple days posting on LuLa and Dyxum as FineArt.

People can easily verify that the method works.

A99 ISO6400 @ 50% attached.


« Last Edit: January 19, 2013, 07:01:06 PM by Fine_Art » Logged
bill t.
Sr. Member
****
Offline Offline

Posts: 2693


WWW
« Reply #1 on: January 19, 2013, 04:29:43 PM »
ReplyReply

That link is to an .exe file.  No way.  Could you please just post a full sized crop or something?
Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #2 on: January 19, 2013, 05:46:49 PM »
ReplyReply

That link is to an .exe file.  No way.  Could you please just post a full sized crop or something?

Sendspace is pushing their toolbar and ads. Ignore that and look at the link lower center in the blue box. If you hover your mouse over it it wall show a .jpg file.

Edit: For those that dont have bandwidth here is a screenshot of the RED channel seperated. Left is high ISO before NR. Right is the same file after NR. This is repeated for RGB and L channels. All channels have channel specific noise.


« Last Edit: January 19, 2013, 06:57:06 PM by Fine_Art » Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1153



WWW
« Reply #3 on: January 19, 2013, 09:15:27 PM »
ReplyReply

Quote
I am not in the imaging business so I am not going to try to develop this somehow. Anyone can use it. If someone puts it in their blog, book, document, whatever, I expect it to be called Arthur's method referencing Lula and Dyxum. I worked on it for a couple days posting on LuLa and Dyxum as FineArt.

I wouldn't worry about anyone giving you credit for your method, I doubt anyone's going to use it anyway.

Too much work for what little results it brings using such an obscure piece of software. Besides you haven't posted any instructions on how to do it EFFICIENTLY so photographers can apply it to the multitude of shots they have to process.

And BTW, don't know if you're aware of this but your noise reduction sample shots you've been posting have random glitter like white dots throughout the image. It's almost has a crystalline effect. Weird.
Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1153



WWW
« Reply #4 on: January 19, 2013, 09:35:15 PM »
ReplyReply

Just took a shot of my display showing the jpeg you posted on your original response which shows the spots up close. The B&W red channel sample you posted doesn't have this, but other similar RGB composite shots your've posted in other threads on the subject show the spots as well.
« Last Edit: January 19, 2013, 09:37:35 PM by tlooknbill » Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #5 on: January 20, 2013, 02:10:11 AM »
ReplyReply

Just took a shot of my display showing the jpeg you posted on your original response which shows the spots up close. The B&W red channel sample you posted doesn't have this, but other similar RGB composite shots your've posted in other threads on the subject show the spots as well.

I cant see much past the screen door effect of your LCD. Just reference the post if you want more info.

This?
« Last Edit: January 20, 2013, 02:22:59 AM by Fine_Art » Logged
thierrylegros396
Sr. Member
****
Offline Offline

Posts: 652


« Reply #6 on: January 20, 2013, 02:54:22 AM »
ReplyReply

Impressive results !

As already said, it seems to produce a lot of small bright dots.

But I imagine these are easier to remove than lower frequency dots.

Thanks for sharing.

Thierry
« Last Edit: January 20, 2013, 02:57:55 AM by thierrylegros396 » Logged
stamper
Sr. Member
****
Offline Offline

Posts: 2699


« Reply #7 on: January 20, 2013, 03:22:01 AM »
ReplyReply

If my ageing memory is correct then Dan Margulis has pioneered these methods in one of his books. I will have a look later to confirm this. It is very doubtful if you were the first. I would take a step back and think before making any more doubtful claims. Sad
Logged

stamper
Sr. Member
****
Offline Offline

Posts: 2699


« Reply #8 on: January 20, 2013, 05:12:31 AM »
ReplyReply

Quote Fine_ Art.

I have tested a new method (as far as I know) for Noise Reduction. Unless I hear its published already I will call it Arthur's method.

Unquote.

Dan Margulis covers noise reduction in channels in his Professional Photoshop book 1st Edition and in his book Photoshop LAB Color. He recommends a trip to the lab colorspace and blurring the A and B channels and then going back to RGB, It is well known that the blue channel is the noisiest and this method reduces it. Sharpening in the blue channel is not recommended. I don't think Arthur's method is a new concept?
Logged

thierrylegros396
Sr. Member
****
Offline Offline

Posts: 652


« Reply #9 on: January 20, 2013, 09:40:44 AM »
ReplyReply

But perhaps your method is more effective.

So we stay tuned to see how you do in detail !

Thierry
Logged
BartvanderWolf
Sr. Member
****
Online Online

Posts: 3640


« Reply #10 on: January 20, 2013, 10:37:00 AM »
ReplyReply

I don't think Arthur's method is a new concept?

Not necessarily, but it's completely different from what Dan Margulis was preaching. Dan is suggesting the blurring of the Chroma channels, while Arthur is focusing on removing noise at the Raw stage before demosaicing (which normally would turn single pixel noise into multiple pixel blobs of color). That's what Astrophotography tends to do, improve the signal to noise ratio as much as possible before demosaicing.

The challenge is in reducing the noise (not only chromatic noise), but not affect the capability to demosaic actual resolution at the same time.

Raw converters like DxO Optics Pro, RawTherapee 4.x, and Capture One Pro 7, offer noise reduction and Luminance detail control that works at the Raw data level, while being Color Managed. The program Arthur uses (ImagesPlus) is not color managed, and also a bit too specialized for regular photographic use. Some of the noise reduction functions of ImagesPlus are adaptive (edge preserving), and can be tweaked at the Bayer CFA level, before the demosaicing, and before leaving the linear gamma space.

A very high quality (and Color Managed !) alternative for ImagesPlus, is PixInsight. However, that's also not very easy to use, mainly because it's documentation is not available for many of the functions (although a lot can be learned from their forum members), and it is more than just an image processing application. It's also a program development environment for image processing, mainly focusing at the needs of astronomers, but it is very well programmed by a small team of dedicated programmers/astronomers.
 
Cheers,
Bart
Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #11 on: January 20, 2013, 11:58:41 AM »
ReplyReply

Quote Fine_ Art.

I have tested a new method (as far as I know) for Noise Reduction. Unless I hear its published already I will call it Arthur's method.

Unquote.

Dan Margulis covers noise reduction in channels in his Professional Photoshop book 1st Edition and in his book Photoshop LAB Color. He recommends a trip to the lab colorspace and blurring the A and B channels and then going back to RGB, It is well known that the blue channel is the noisiest and this method reduces it. Sharpening in the blue channel is not recommended. I don't think Arthur's method is a new concept?

Ok thanks. I will have to get his book.

Forget "Arthur's method" call it "Dan's method"! I just re-invented the wheel.
Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #12 on: January 20, 2013, 12:42:56 PM »
ReplyReply

Not necessarily, but it's completely different from what Dan Margulis was preaching. Dan is suggesting the blurring of the Chroma channels, while Arthur is focusing on removing noise at the Raw stage before demosaicing (which normally would turn single pixel noise into multiple pixel blobs of color). That's what Astrophotography tends to do, improve the signal to noise ratio as much as possible before demosaicing.

The challenge is in reducing the noise (not only chromatic noise), but not affect the capability to demosaic actual resolution at the same time.

Raw converters like DxO Optics Pro, RawTherapee 4.x, and Capture One Pro 7, offer noise reduction and Luminance detail control that works at the Raw data level, while being Color Managed. The program Arthur uses (ImagesPlus) is not color managed, and also a bit too specialized for regular photographic use. Some of the noise reduction functions of ImagesPlus are adaptive (edge preserving), and can be tweaked at the Bayer CFA level, before the demosaicing, and before leaving the linear gamma space.

A very high quality (and Color Managed !) alternative for ImagesPlus, is PixInsight. However, that's also not very easy to use, mainly because it's documentation is not available for many of the functions (although a lot can be learned from their forum members), and it is more than just an image processing application. It's also a program development environment for image processing, mainly focusing at the needs of astronomers, but it is very well programmed by a small team of dedicated programmers/astronomers.
 
Cheers,
Bart

It's a bit different, not quite what Dan did, from your explanation and not quite what you describe. I must have failed to communicate it properly.

This method does not really let me remove noise before de-mosaicing. I am doing my steps after de-mosaicing. What I guess my method shows is that the de-bayer algorithms are nowhere near as advanced as the latest de-noise algorithms. I am also using them in different dimensions. Thinking this way is more natural to someone who has matrix math training and/or someone who has OLAP database training. It's more of a jump for others. It can still be done.

I am using the power of a modern computer with modern noise reduction in 12 dimensions. In camera processing cannot touch this. For example my software is 64bit multi-threaded on a quad core 2.8 GHz with 8 Gigs RAM. Forget tiny chips in your camera for JPGs.

What we assume our de-bayer method does is fill in holes from missing pixels on each color to re-create full image by color layer just as the old film color layers did. I would say for Low ISO whatever they are doing seems to work. The problem comes to High ISO when we see very harsh jumps in pixel values (noise) or wandering color patterns. Clumps of colors without NR. This is a processing failure not necessarily a shot S/N problem.

So here it is again. My process uses the best NR programs on each color channel. I happen to use Images Plus for a very clean (random) noise conversion. I'm sure other programs like what Bart mentions work well too. The TEST for working well is leaving a fine random pattern of noise. There should be no clumps of color. When you seperate out the channels you should still see fine random noise that looks like fine film grain.

Attached is a screenshot of Blue, Green and Red channels shown in B/W. The noise is strongest on Red, then Blue, then Green. Your favorite N/R programs noise ninja, topaz de-noise, neat image, whatever will do wonders on this. Use it on RGBL. The noise also looks the same in CMYL. Do it again. Re-combine each split RGBL, CMYL. Combine both in layers or using average functions.

What you have done is take slices through a 12 dimension matrix, wiped out pixel to pixel variances in each dimension, leaving a very natural looking file with real depth. Heavy NR as your raw converter programs do it leave a very artificial looking image. With their methods your choice is leave noise in or look artificial.

So by Bart's description Dan's method is not the same as Arthur's method. The name Arthur's method stays.

Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #13 on: January 20, 2013, 12:56:34 PM »
ReplyReply

Looking at that screenshot in the prior post, it is easy to understand that smoothing functions will create smooth waves from those fine speckles. That is what most converter software is doing. When re-combined those waves will not line up. Areas that randomly have slightly higher values in 3x3, 5x5 or 9x9 for example will be different coordinates on each channel. Recombining those leads to lumps of color. You see artificial patterns of wandering red, green, blue. Clearly that was not in the real world. Clearly it was not at the bayer level which is single pixel level noise. It is from smoothed waves that are out of phase by channel.

If your software does this it is failing you. Use something else.
Logged
opgr
Sr. Member
****
Offline Offline

Posts: 1125


WWW
« Reply #14 on: January 20, 2013, 01:33:43 PM »
ReplyReply

This method does not really let me remove noise before de-mosaicing. I am doing my steps after de-mosaicing. What I guess my method shows is that the de-bayer algorithms are nowhere near as advanced as the latest de-noise algorithms.

You're still not making sense and contradicting yourself.

Bayer data gives you only one of R, G, or B for each pixel, and you therefore need to come up with a value for the other two channels per pixel. Applying noise-reduction prior to this stage and/or after this stage, on difference-signals of any kind is the standard method of denoise. You are currently deluding yourself (and others) by making it a feindishly complex combination of everything and then claiming nonsense about RAW converters and dedicated DSPs.

A dedicated denoise program will have more parameters to control noise reduction because the RAW converter generally makes compromises for the sake of speed. The most effective additional parameters in dedicated noise-reduction are the separation of the blur radius per difference level. Any which way you turn it, the general method is to take some kind of average to determine what constitutes noise and what constitutes signal. This average signal is what causes the large blobs of color. It is the average signal. It is not some kind of crap that the RAW converter is introducing. If you would look at a live-view of the image, these blobs would move even though the actual image would be stationary.

A dedicated noise-reduction program will have several levels of average signal and perhaps even colorprofiles for strength etc. You could keep the noise at the pixel level, and remove it at the larger average levels. That would give you your preferred fine-grained colornoise. LR 4 seems to have introduced new pyramid levels for the image processing which also allows both more sophisticated noise reduction as well as local contrast control which also works with these difference levels.

I think the only thing you are currently doing is using a really round-about way of saying that modern noise-reduction programs and algorithms do a better job than what is available in most RAW converters. Well, that is exactly why there are dedicated software options. What you should define is what it is exactly that you expect from the RAW converter to feed into the denoise-software. Simple interpolation like nearest-neighbor or linear?

Additionally, if you think yourself well-versed in matrix math, then you should know that you can matrix-convert an RGB triplet all day long, into as many dimensions as you like, but that ain't gonna do sh*t for the RGB triplets on the output side. If it can be described in matrix form, then it isn't doing anything. It is the additional math that introduce non-linearities and averaging that is making the difference.








Logged

Regards,
Oscar Rysdyk
theimagingfactory
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #15 on: January 20, 2013, 02:04:17 PM »
ReplyReply

You're still not making sense and contradicting yourself.

Bayer data gives you only one of R, G, or B for each pixel, and you therefore need to come up with a value for the other two channels per pixel. Applying noise-reduction prior to this stage and/or after this stage, on difference-signals of any kind is the standard method of denoise. You are currently deluding yourself (and others) by making it a feindishly complex combination of everything and then claiming nonsense about RAW converters and dedicated DSPs.

A dedicated denoise program will have more parameters to control noise reduction because the RAW converter generally makes compromises for the sake of speed. The most effective additional parameters in dedicated noise-reduction are the separation of the blur radius per difference level. Any which way you turn it, the general method is to take some kind of average to determine what constitutes noise and what constitutes signal. This average signal is what causes the large blobs of color. It is the average signal. It is not some kind of crap that the RAW converter is introducing. If you would look at a live-view of the image, these blobs would move even though the actual image would be stationary.

A dedicated noise-reduction program will have several levels of average signal and perhaps even colorprofiles for strength etc. You could keep the noise at the pixel level, and remove it at the larger average levels. That would give you your preferred fine-grained colornoise. LR 4 seems to have introduced new pyramid levels for the image processing which also allows both more sophisticated noise reduction as well as local contrast control which also works with these difference levels.

I think the only thing you are currently doing is using a really round-about way of saying that modern noise-reduction programs and algorithms do a better job than what is available in most RAW converters. Well, that is exactly why there are dedicated software options. What you should define is what it is exactly that you expect from the RAW converter to feed into the denoise-software. Simple interpolation like nearest-neighbor or linear?

Additionally, if you think yourself well-versed in matrix math, then you should know that you can matrix-convert an RGB triplet all day long, into as many dimensions as you like, but that ain't gonna do sh*t for the RGB triplets on the output side. If it can be described in matrix form, then it isn't doing anything. It is the additional math that introduce non-linearities and averaging that is making the difference.


Modern NR programs do not use averaging. Images plus has several NR functions that do not smear out edges or fine detail.

You can blab all you want about how it cant work. Anyone who does it on their computer picking the best NR by channel they can find, will see a big jump in image quality.
Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1153



WWW
« Reply #16 on: January 20, 2013, 02:32:18 PM »
ReplyReply

I cant see much past the screen door effect of your LCD. Just reference the post if you want more info.

This?

Yes, those tiny white specs. What causes that and can it be fixed?

The test shot you posted of the proportion wheel is very impressive with regard to detail as seen in the proportion wheel, but I don't see your methods attaining these results as functional for regular photographers seeing we aren't matrix mathematicians.

I'm assuming you've already shown this to photo astronomers, but I'm trying to understand why you'ld think regular photographers would find any use in this. Or maybe you're just wanting to toot you're own horn at finding a better mousetrap and if that's the case you might consider contacting competing Raw converter vendors of your newly discovered method and sell it to them.

I'm not familiar with the photo astronomy community, but I'm guessing that they're very highly technical individuals who can examine the math behind what you're doing and figure it out for themselves. Also I can't believe they'ld be resorting to CFA Bayer methods of recording the cosmos and manipulating data like this to support the idea that what they're photographing is actually factual and expecting us to believe it knowing those methods employed. I thought photo astronomy would be using a lot more sophisticated technology rather than simple bayer sensors and software interpretation.
« Last Edit: January 20, 2013, 02:34:38 PM by tlooknbill » Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #17 on: January 20, 2013, 02:32:47 PM »
ReplyReply

Nikon D800 ISO3200 NR0 converted in Images plus. Note the fine random nature of the noise.

The second screenshot is with NR by channel. It gets some faint color grouping from the NR. The noise does not start to form into a pattern.

Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1153



WWW
« Reply #18 on: January 20, 2013, 02:41:40 PM »
ReplyReply

I would think the forensic video recovery investigators of law enforement who have to make out license plates and facial features within noisy crime scene video would be interested in your methods if it could be applied in the same manner.
Logged
Fine_Art
Sr. Member
****
Offline Offline

Posts: 1087


« Reply #19 on: January 20, 2013, 02:48:57 PM »
ReplyReply

I would think the forensic video recovery investigators of law enforement who have to make out license plates and facial features within noisy crime scene video would be interested in your methods if it could be applied in the same manner.

They widely use deconvolution sharpening technology. So would the military on their recon imaging. Do they use NR by channel? Probably not, most of their work is B/W for the higher sensitivity of no color filters.
Logged
Pages: [1] 2 3 »   Top of Page
Print
Jump to:  

Ad
Ad
Ad