downsampling reduces noice! - uh noise! - con't.

Started Nov 16, 2011 | Discussions
crames
Regular MemberPosts: 188
Like?
downsampling reduces noice! - uh noise! - con't.
Nov 16, 2011

Continuing that interesting thread that just hit the max post limit...

Here is a crop from a plenty-noisy Imaging Resource D3s 102400 ISO shot. It has been processed as follows:

1 Original

2 Lowpass filtered, everything above Nyquist/2 zeroed using DFT. Not downsampled.

3 Original downsampled by factor 2, without filtering. Equivalent to "Nearest-neighbor" or true "decimation by 2." Includes aliases due to lack of lowpass filter.

4 Lowpass filtered as in 2, then downsampled as in 3. "Properly downsampled."

5 Image 4 subtracted from image 3 to reveal noisy aliases resulting from lack of lowpass filter. (offset 128 has been added)

Unless I'm going blind, the "properly downsampled" image 4 looks a lot less noisy and even sharper than the original. Of course it's smaller, but that's part of the definition of downsampling.

Cliff

Nikon D3S
If you believe there are incorrect tags, please send us this post using our feedback form.
hjulenissen
Senior MemberPosts: 1,593
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to crames, Nov 16, 2011

Thank you, you saved me from making such images myself.

May I suggest that you also resample the "properly downsampled" image back up to the higher resolution so that people will have the possibility to compare original/processed image without image scale complicating things?

If we define noise to be an "image error" quantifiable as e = (y_ref - y)^2 or something similar, it would perhaps be illustrating to present an image of the same scene using e.g. A) iso100, B) iso1600 and C)iso1600 processed by down/upsampling. The error of B vs A compared to C vs A would be interesting.

-h

Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to crames, Nov 16, 2011

crames wrote:

Unless I'm going blind, the "properly downsampled" image 4 looks a lot less noisy and even sharper than the original.

It isn't. No need to guess...just resize it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

As all software labeled "noise reduction" reduces the apparent noise while maintaining the resolution and practically all the apparent image detail (which is also the expectation of the people using such software,) I'd say that any discussion on noise reduction should hold to the same. I think it would be dishonest to redefine "noise reduction" now just to have it fit a theoretical argument with no value.

And #4 does not fit the bill.

.

Reply   Reply with quote   Complain
bobn2
Forum ProPosts: 27,835
Like?
Nicely done
In reply to crames, Nov 16, 2011

Thanks, bookmarked for future reference
--
Bob

Reply   Reply with quote   Complain
hjulenissen
Senior MemberPosts: 1,593
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

It isn't. No need to guess...just resize it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

Of course real scene detail will be lost. You have been told so many times. I have asked you several times: do you know what a lowpass filter is?

The point is that noise also is removed, quite the oposite of the view that you repeatedly have claimed without empiry or arguments.

As all software labeled "noise reduction" reduces the apparent noise while maintaining the resolution and practically all the apparent image detail (which is also the expectation of the people using such software,) I'd say that any discussion on noise reduction should hold to the same. I think it would be dishonest to redefine "noise reduction" now just to have it fit a theoretical argument with no value.

Do you understand the difference between the term A) "noise reduction" and the term B) "to reduce noise"? Should any scientific definition of what "noise" is, be defined on the basis of what some photography-specific software choose to do with noisy images? Don't you think that is kind of backwards?

-h

Reply   Reply with quote   Complain
hjulenissen
Senior MemberPosts: 1,593
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

The approach below will of course be scene (signal) dependant, and work best for a smooth scene with few spatial details. That is because my simple error metric will quantify the effects of both noise filtering and linear signal filtering. If the signal part have only a 0-frequency component ("DC" or flat wall), then any filter that leaves DC untouched will not change the signal, and its effect on noise alone can be easily quantified.

For parts of the scene that have little details (e.g. sky), lowpass filtering or downsampling by an appropriate factor might work ok. For most real-world scenes, the loss of high-frequency detail in in-focus detailed features (e.g. hair and eyes) would be untolerable if you want to suppress any significant amount of noise. I believe that a very broad characterization of dedicated image noise reduction is that it does similar smoothing dependant on assumed local signal/image features, and possible dependant on known camera behaviour (iso setting, explicit estimates etc).

hjulenissen wrote:

May I suggest that you also resample the "properly downsampled" image back up to the higher resolution so that people will have the possibility to compare original/processed image without image scale complicating things?

If we define noise to be an "image error" quantifiable as e = (y_ref - y)^2 or something similar, it would perhaps be illustrating to present an image of the same scene using e.g. A) iso100, B) iso1600 and C)iso1600 processed by down/upsampling. The error of B vs A compared to C vs A would be interesting.

-h

Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

hjulenissen wrote:

Graystar wrote:

It isn't. No need to guess...just resize it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

Of course real scene detail will be lost. You have been told so many times.

Having said it doesn't make it acceptable.

As all software labeled "noise reduction" reduces the apparent noise while maintaining the resolution and practically all the apparent image detail (which is also the expectation of the people using such software,) I'd say that any discussion on noise reduction should hold to the same. I think it would be dishonest to redefine "noise reduction" now just to have it fit a theoretical argument with no value.

Do you understand the difference between the term A) "noise reduction" and the term B) "to reduce noise"?

Yes. There's no difference.

Should any scientific definition of what "noise" is, be defined on the basis of what some photography-specific software choose to do with noisy images? Don't you think that is kind of backwards?

It's only backwards to someone who is desperate to have the term "noise reduction" mean something it doesn't.

.

Reply   Reply with quote   Complain
hjulenissen
Senior MemberPosts: 1,593
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

Do you understand the difference between the term A) "noise reduction" and the term B) "to reduce noise"?

Yes. There's no difference.

Thank you.

It's only backwards to someone who is desperate to have the term "noise reduction" mean something it doesn't.

First you claimed that downsampling did not reduce noise. Now you try to discuss semantics. Any informed reader should now understand that you are not doing this community a service.

http://forums.dpreview.com/forums/read.asp?forum=1018&message=39785288&problem=max

Graystar wrote:

Downsampling does not reduce noise. It’s provable experimentally because it’s proven mathematically.

Still waiting for that mathematical proof.

-h

Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

hjulenissen wrote:

First you claimed that downsampling did not reduce noise.

It doesn't.

Now you try to discuss semantics.

What do you mean by "now"? That's been the issue from the very beginning.

Any informed reader should now understand that you are not doing this community a service.

Any informed reader will expect a noise reduction process to leave the image resolution and detail as is.

Anyone who is proposing a noise reduction scheme that doesn't meet the expectations of the community is the one who's not doing this community a service.

Still waiting for that mathematical proof.

That's because you weren't paying attention when I gave it. I showed how binning reduces the noise variance by half, which equals the loss of resolution.

And I'm still waiting for the images demonstrating that downsampling reduces noise.

.

Reply   Reply with quote   Complain
hjulenissen
Senior MemberPosts: 1,593
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

Any informed reader will expect a noise reduction process to leave the image resolution and detail as is.

That would be a poorly informed reader since that is (generally) impossible. Any reader that is actually interested in photography should know from experience that there is a trade-off involved between removing noise and preserving image detail. Any reader that is interested in signal processing should know that separating signal from noise blindly is at least very hard, and probably generally impossible. So what is you interest?

Anyone who is proposing a noise reduction scheme that doesn't meet the expectations of the community is the one who's not doing this community a service.

You still dont get it? I have said multiple times that I am not proposing a noise reduction scheme. But I think it is in the interest of the reader to know what happens to the image IF she choose to do it anyways. And I think that your explanations so far have been naiive, misleading and lacking proof.

And I'm still waiting for the images demonstrating that downsampling reduces noise.

Look at the top of this post.

In case you still find it hard to read my english:

Image filtering/resampling will typically alter image noise and image signal. Noise reduction will typically alter image noise and image signal. If what you want is a "good" looking image (or one that closely resemble that image that you could have taken at a lower ISO using some sensible similarity measure), you will generally be more satisfied using dedicated noise reduction algorithms than using a linear filter/image resampling.

-h

Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

hjulenissen wrote:

Graystar wrote:

Any informed reader will expect a noise reduction process to leave the image resolution and detail as is.

That would be a poorly informed reader since that is (generally) impossible.

First, not only is it possible to leave resolution as is, it's a requirement. Second, I've seen many examples of noise reduced images where image detail was preserved to the point where losses are practically imperceptible...very unlike the OP’s image #4.

Anyone who is proposing a noise reduction scheme that doesn't meet the expectations of the community is the one who's not doing this community a service.

You still dont get it? I have said multiple times that I am not proposing a noise reduction scheme.

I don’t care what YOU are saying. I care about what was proposed in the opening posts of the threads I’m responding to. And the opening posts of both threads are claiming to reduce noise through downsampling, and that’s not going to happen if we keep to the photo-editing definition of reducing noise (which is all that matters to people.)

But I think it is in the interest of the reader to know what happens to the image IF she choose to do it anyways. And I think that your explanations so far have been naiive, misleading and lacking proof.

I don’t need to prove anything. YOU are the one that needs to prove your claims. And so far, no images have been produced to support your claim. Once again, put up, or shut up.

And I'm still waiting for the images demonstrating that downsampling reduces noise.

Look at the top of this post.

The opening post shows a reduced sized image that has significant losses in detail, and so doesn’t meet the expectation of a noise reduction process.

In case you still find it hard to read my english:

Image filtering/resampling will typically alter image noise and image signal. Noise reduction will typically alter image noise and image signal. If what you want is a "good" looking image (or one that closely resemble that image that you could have taken at a lower ISO using some sensible similarity measure), you will generally be more satisfied using dedicated noise reduction algorithms than using a linear filter/image resampling.

Of course you’d be more satisfied using dedicated noise reduction software when downsampling provides zero satisfaction.

.

Reply   Reply with quote   Complain
crames
Regular MemberPosts: 188
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

crames wrote:

Unless I'm going blind, the "properly downsampled" image 4 looks a lot less noisy and even sharper than the original.

It isn't. No need to guess...just re-size it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

Image #2 is #4 up-sampled back to the original size. But since #2 is no longer down-sampled, what is the point of comparing it to the original? #2 has been effectively lowpass filtered with a brick wall filter and contains no frequencies above half-Nyquist.

However image #4 has frequencies right up to Nyquist. Referring to Emil's illustrations in the original thread, it is likely that the image contrast near Nyquist is higher on the down-sampled image than on the original, which would explain increased sharpness. (Down-sampling moves Nyquist to the left on the frequency axis.)

As all software labeled "noise reduction" reduces the apparent noise while maintaining the resolution and practically all the apparent image detail (which is also the expectation of the people using such software,) I'd say that any discussion on noise reduction should hold to the same. I think it would be dishonest to redefine "noise reduction" now just to have it fit a theoretical argument with no value.

I would say that down-sampling can offer a trade-off between noise/sharpness and image size. It seems that you want to define noise reduction to exclude any change in image size and therefore exclude down-sampling. Image #4 has less visible noise, yet you would not say that it has undergone "noise reduction?"

And #4 does not fit the bill.

I think you're talking about #2, which isn't down-sampled and certainly doesn't fit the bill or the topic.

Image #4 is both less-noisy and sharper due to down-sampling.

Cliff

Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to crames, Nov 16, 2011

crames wrote:

Graystar wrote:

crames wrote:

Unless I'm going blind, the "properly downsampled" image 4 looks a lot less noisy and even sharper than the original.

It isn't. No need to guess...just re-size it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

Image #2 is #4 up-sampled back to the original size.

Your note for #2 says "Not downsampled." So it's not #4 upsampled. You are saying it is because you believe that it is equivalent to #4 upsampled based on your process. It is not.

But since #2 is no longer down-sampled, what is the point of comparing it to the original?

If you don't believe that the process of noise reduction has anything to do with preserving the signal, then I guess none whatsoever.

.

Reply   Reply with quote   Complain
theswede
Veteran MemberPosts: 3,936Gear list
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

Any informed reader will expect a noise reduction process to leave the image resolution and detail as is.

No, because that is not possible. Any such expectation is unrealistic. All noise reduction will reduce spatial resolution and detail to some degree. The effort in sophisticated noise reduction algorithms is to minimize this loss, but since some of the detail is noise this is impossible.

Anyone who is proposing a noise reduction scheme that doesn't meet the expectations of the community is the one who's not doing this community a service.

Then all noise reduction software manufacturers are doing this community a disservice. Clearly an unreasonable stance, since most noise reduction algorithms improve on binning to some degree despite them all losing detail and resolution when removing noise.

Still waiting for that mathematical proof.

That's because you weren't paying attention when I gave it. I showed how binning reduces the noise variance by half, which equals the loss of resolution.

The noise is reduced by half, exactly as expected. You even prove it yourself, yet claim it doesn't happen. Quite an interesting stance.

Jesper

 theswede's gear list:theswede's gear list
Fujifilm FinePix X100 Konica Minolta Maxxum 5D Sony SLT-A37 Sony 50mm F1.4 +2 more
Reply   Reply with quote   Complain
Graystar
Veteran MemberPosts: 8,373
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to crames, Nov 16, 2011

Print your #4 and #1 at the same size and you'll see that all you accomplished was to blur detail while the noise seems virtually the same.

I'm done with this ridiculous argument. Believe what you like.

.

Reply   Reply with quote   Complain
crames
Regular MemberPosts: 188
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

crames wrote:

Graystar wrote:

crames wrote:

Unless I'm going blind, the "properly downsampled" image 4 looks a lot less noisy and even sharper than the original.

It isn't. No need to guess...just re-size it back to the original size. The noise isn't as fine but #4 isn't anywhere near as sharp or as detailed as the original.

Image #2 is #4 up-sampled back to the original size.

Your note for #2 says "Not downsampled." So it's not #4 upsampled. You are saying it is because you believe that it is equivalent to #4 upsampled based on your process. It is not.

Yes it is. By "not downsampled" I meant that it has the same number of samples as the original.

Here is the process so that you can see the equivalence:

Image 2: Original -> DFT -> zero all frequency samples above Nyquist/2 -> IDFT
Image 4: Original -> DFT -> remove all frequency samples above Nyquist/2 -> IDFT

Which for Image 2 is the same as taking the frequency samples of image 4 and surrounding them with zeros out to the original size (this performs sinc interpolation in the frequency domain).

But since #2 is no longer down-sampled, what is the point of comparing it to the original?

If you don't believe that the process of noise reduction has anything to do with preserving the signal, then I guess none whatsoever.

Do you believe that the effect of down-sampling can be seen without actually looiking at the down-sampled image?

Note: I forgot to mention the the STD numbers on the images are the standard deviations measured within the selected corners (while the images were linear). The STD numbers show that the noise was reduced, whether or not you can see it.

Reply   Reply with quote   Complain
crames
Regular MemberPosts: 188
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to Graystar, Nov 16, 2011

Graystar wrote:

Print your #4 and #1 at the same size and you'll see that all you accomplished was to blur detail while the noise seems virtually the same.

In other words you want to up-sample the down-sampled image, so that it is no longer down-sampled, and draw some kind of conclusion about down-sampling from that?

I'm done with this ridiculous argument. Believe what you like.

I believe what I see!

Reply   Reply with quote   Complain
John Sheehy
Forum ProPosts: 15,958
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

hjulenissen wrote:

Thank you, you saved me from making such images myself.

May I suggest that you also resample the "properly downsampled" image back up to the higher resolution so that people will have the possibility to compare original/processed image without image scale complicating things?

You still have inequitable sharpness if you do that; it is less sharp than both the full size, and the full-size with LP filtering. This is where you philosophy falls flat on its face; #4 only looks less noisy when you upsample it because it is less sharp than the originals; this is getting absurd.

What you are doing is no more "reducing noise" than turning down the treble on a stereo reduces hiss; it just reduces the treble; it doesn't reduce the hiss. Reducing the hiss would leave the cymbals sounding crisp.

To call downsampling a method of NR is absurd; it is a game of semantic bait-and-switch, where you use the reduction (or elimination) of high frequencies of everything, and call it a reduction of one thing (noise).

Downsampling does not reduce noise; it eliminates the high end of the spectrum of everything . LP filtering at full resolution reduces the contrast of everything that is high-frequency; not just noise. You only reduce noise when you reduce it relative to something else at the same frequency.
--
John

Reply   Reply with quote   Complain
John Sheehy
Forum ProPosts: 15,958
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to hjulenissen, Nov 16, 2011

hjulenissen wrote:

Of course real scene detail will be lost. You have been told so many times. I have asked you several times: do you know what a lowpass filter is?

Low-pass filters do not lose detail. They lower the contrast of high-resolution details, including noise details. The SNR is still the same at every frequency, despite the lowering of high-frequency contrast.

The point is that noise also is removed, quite the oposite of the view that you repeatedly have claimed without empiry or arguments.

That's a completely worthless point, as it is just a mild version of removing all noise by powering off your monitor. Hiding noise, and reducing noise, are two completely different things.

-- hide signature --

John

Reply   Reply with quote   Complain
John Sheehy
Forum ProPosts: 15,958
Like?
Re: downsampling reduces noice! - uh noise! - con't.
In reply to theswede, Nov 16, 2011

theswede wrote:

Graystar wrote:

Any informed reader will expect a noise reduction process to leave the image resolution and detail as is.

No, because that is not possible.

It's not possible to do it perfectly, but it is possible to do less damage to desirable detail than to noise.

Anyone who is proposing a noise reduction scheme that doesn't meet the expectations of the community is the one who's not doing this community a service.

Then all noise reduction software manufacturers are doing this community a disservice. Clearly an unreasonable stance, since most noise reduction algorithms improve on binning to some degree despite them all losing detail and resolution when removing noise.

The best ones don't lose detail as much as binning. You can have your transient between original lines 4 and 5, and also between original lines 5 and 6; you can't do that with downsampling or binning.

The noise is reduced by half, exactly as expected. You even prove it yourself, yet claim it doesn't happen. Quite an interesting stance.

Nonsense. The noise is reduced to 50% in the Z axis. It now covers 4x the percentage of image area, and is recovered with rescaling.

-- hide signature --

John

Reply   Reply with quote   Complain
Keyboard shortcuts:
FForum MMy threads