Luminous Landscape Forum

Raw & Post Processing, Printing => Digital Image Processing => Topic started by: fike on December 21, 2009, 10:54:12 am

Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 21, 2009, 10:54:12 am
I have been noticing noisy skies in some of my dramatic black and white conversions.  Last week I reprocessed some of these images (they were 24-image panos) using a 16-bit workflow.  The mottled skies were substantially improved.  

Does the increased noise in the more densely packed sensors make a 16-bit workflow more valuable?

Can the extra bit-depth of a 16-bit image ameliorate the negative impact of chroma noise?  

(I am not implying that 16-bit workflow reduces noise.)

I am suggesting that the small extra quality in color/light gradations may improve image quality enough to reduce the perception of noise.  

This image is one of the images where I saw some mottled sky effects that I thought could be improved with 16-bit workflow:
Trailpixie.net (http://www.trailpixie.net/canaan_valley/dusk_on_camp_70.htm)
(http://www.trailpixie.net/Camp-70-50D-8417-10-HDR-Panorama-S.jpg)
Title: Noisy Sensors and 16-bit Workflow
Post by: Jonathan Wienke on December 21, 2009, 11:22:37 am
What you're seeing is not "noise", it is posterization. And using an all-16-bit workflow will prevent or greatly reduce it. If it was really noise, it would still be there regardless of 8-bit or 16-bit workflow.
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 21, 2009, 12:58:52 pm
Quote from: Jonathan Wienke
What you're seeing is not "noise", it is posterization. And using an all-16-bit workflow will prevent or greatly reduce it. If it was really noise, it would still be there regardless of 8-bit or 16-bit workflow.

I understand and believe what you are saying, but I'm not accustomed to seeing posterization in a mottled look.  Normally it manifests itself across gradients.  though it isn't noise, it looks a bit like noise.  As a result, I wonder if low-level noise added to this sort of mottled posterization makes it even worse.
Title: Noisy Sensors and 16-bit Workflow
Post by: ckimmerle on December 21, 2009, 01:56:10 pm
I think the problem is that you are, in an attempt to give the sky drama, over-stretching the tonal range. The levels of gray in most skies is limited to just a couple of stops, so when you (we) try to expand that range into full-blown black to white it adds noise. As you have seen, it can be pretty bad in 8-bit images, and much better working 16-bit, but even that has it's limits.

You can greatly reduce the noise by working in RAW (if you're not) and creating two digital files: one standard and one toned dark for the sky. Use layer masks to overlay the darkened sky over the standard image. Using this method, you have a greater beginning tonal range.

Also, you can selectively add noise removal in only the sky area as it's primarily the darker and flatter areas that have the biggest problem.

Without seeing a full rez version as well as the RAW/original file, I can only guesstimate. However, I've witnessed those same issues from time to time (especially when trying to convert to grayscale using a heavy red filter), so I'm pretty sure that's at least part of the problem.

Chuck
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 21, 2009, 02:18:49 pm
Quote from: ckimmerle
...(especially when trying to convert to grayscale using a heavy red filter)....

Chuck

yep, that's it.  On another version, I applied noise reduction to the skies with some success.  I do need to redo this image in 16-bit, but this is an HDR pano, so there are 33 images that need to be reprocessed, stitched, and HDR processed.  I need a couple of hours to work on this one.
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 21, 2009, 10:01:21 pm
Here is a very quick redo (cropped to show the relevant area) done in 16-bit.  The same mottled texture is apparent.  It isn't caused by the red channel, more like a combination of the blue and cyan channels.  These were and are RAW images, converted to TIFF, stitched in PTGui and HDR done by exposure fusion.

(http://www.trailpixie.net/temp-image/Camp-70-50D-84172-HDR-16Bit-Panorama.jpg)

While in color the problem isn't nearly as noticeable.  

(http://www.trailpixie.net/temp-image/Camp-70-50D-84172-HDR-16Bit-Panorama-color.jpg)

Next thing I need to look at is my PTGui HDR work.  I have been using exposure fusion, and that could be causing the issue.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 22, 2009, 06:51:00 am
Quote
It isn't caused by the red channel, more like a combination of the blue and cyan channels.
Cyan channel does not exist and blue channel is clean.

Red channel is noisy. Green and blue channels are good.
You can denoise the red using a good denoiser, like PhotoDenoising.

In the color image, noise is not visible as you "see much more green channel".

Building black and white, you have to take that in acount. If you were able to reduce the red noisy you can use as much red you like, otherwise the red must be kept at minimum level.


Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: tokengirl on December 22, 2009, 07:42:18 am
fike,

What software are you using to make your B&W conversions?
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 22, 2009, 09:39:32 am
I am using PS CS4 for B&W conversion. I am using the channel mixer for the conversion.

As for the red channel, I had suspected that red wasn't the culprit because moving the red slider up and down in the B&W channel mixer had no effect on the apparent noise level.  Moving the blue and cyan sliders does increase the effect, particularly when either blue or cyan is adjusted more aggressively.  Adjusting them together seems to reduce the magnitude of the effect but not sufficiently.  

But, to ensure that I was right, I did Neat Image noise reduction to the red channel.  There was not apparent change in the pattern.  The noise pattern in this image is much smaller than these mottled blotches I'm looking at.

I don't think it is noise because the pattern is too large and noise reduction has minimal or no effect.
I don't think it is posterization because it remains after changing the workflow to all 16-bit.

I am using Exposure Fusion with PTGui, so that is my next are of investigation. I will try to make a real HDR image and tone map that image.  

Unless anyone else has other suggestions or insight.
Title: Noisy Sensors and 16-bit Workflow
Post by: tokengirl on December 22, 2009, 12:25:10 pm
Quote from: fike
I am using PS CS4 for B&W conversion. I am using the channel mixer for the conversion.

You might want to download the free trial of Nik's Silver Efex Pro and fool around with it a bit.  I find that I am much more able to control this type of noise in the skies using Silver Efex Pro as opposed to using CS4 or Lightroom for the conversion.  HINT:  Rather than applying a red filter in Silver Efex Pro (which can cause that same type of noise), choose Film Types>Sensitivity and reduce the blue (and maybe the cyan a bit too).

This software is not cheap if you decide to buy it, but I have found it to be worth every single penny.  I do a lot of B&W.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 23, 2009, 04:30:00 am
Quote
I don't think it is noise because the pattern is too large and noise reduction has minimal or no effect.

It is noise, on the red channel. The pattern is not too large.

Noise reduction have to be performed and viewed on red channel. In color image the red noise is about invisible (this is a "green channel clean" effect, as I said).

[attachment=18837:Red.jpg]

This is the red channel.

Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 23, 2009, 08:56:06 am
Quote from: jbrembat
It is noise, on the red channel. The pattern is not too large.

Noise reduction have to be performed and viewed on red channel. In color image the red noise is about invisible (this is a "green channel clean" effect, as I said).

[attachment=18837:Red.jpg]

This is the red channel.

Jacopo

I tried again to reduce noise on the red channel.  No luck.  I used the PS CS4 noise reduction filter and Neat Image noise reduction.  I did noise reduction on the combined RGB channel and on just the red channel.  The mottled pattern remains.  

I also made a standard HDR output from PTGui and tone mapped it with Photomatix and the noise remains. So it doesn't appear that the mottled appearance is caused by exposure fusion.

To summarize what I have tried:
...so I wondered if there were more sophisticated approaches.

One other note.  This image was taken at ISO 800 on a 50D.  Not an ideal situation, but the shadows are clean.  I would expect noise to show up there before it rears its ugly head in clear, blue skies.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 23, 2009, 09:19:33 am
Quote
noise reduction on red channel: didn't work

Incorrect. With a good denoiser, you can get a very good red channel.

Red channel after fast denoising using PhotoDenoising:

[attachment=18847:Red_denoised.jpg]

Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 23, 2009, 10:38:57 am
Quote from: jbrembat
Incorrect. With a good denoiser, you can get a very good red channel.

Red channel after fast denoising using PhotoDenoising:

[attachment=18847:Red_denoised.jpg]

Jacopo

Can you tell me a little bit more about your tools and procedure?  At first glance, your processing looks very good.
Title: Noisy Sensors and 16-bit Workflow
Post by: tokengirl on December 23, 2009, 11:24:29 am
Quote from: fike
One other note.  This image was taken at ISO 800 on a 50D.  Not an ideal situation, but the shadows are clean.  I would expect noise to show up there before it rears its ugly head in clear, blue skies.

I experienced this problem sometimes with my old 40D, and from what I've seen, the 50D had even more noise problems with all those extra pixels crammed on the sensor.  Upgrading to the 5DMkII fully solved the problem though.  
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 23, 2009, 01:50:44 pm
Quote from: tokengirl
... Upgrading to the 5DMkII fully solved the problem though.

hmmmm...no.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 23, 2009, 02:02:26 pm
Quote from: fike
Can you tell me a little bit more about your tools and procedure?  At first glance, your processing looks very good.
Simple, I used PhotoDenoising and performed the action only on red channel.
A demo of PhotoDenoising can be downloaded  here (http://www.photoresampling.com/index_eng.php)

Jacopo

Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 23, 2009, 02:28:18 pm
Quote from: jbrembat
Simple, I used PhotoDenoising and performed the action only on red channel.
A demo of PhotoDenoising can be downloaded  here (http://www.photoresampling.com/index_eng.php)

Jacopo

I will try downloading it and give it a shot.  I wonder what PhotoDenoising is doing that Neat Image is not doing?
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 23, 2009, 02:43:54 pm
Quote
I wonder what PhotoDenoising is doing that Neat Image is not doing?
Why?

If you download PhotoResampling (from the same web site) you will find things that PhotoShop is unable to do.

Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 23, 2009, 05:44:48 pm
Quote from: jbrembat
Why?

If you download PhotoResampling (from the same web site) you will find things that PhotoShop is unable to do.

Jacopo

I guess you are involved in the creation of this interesting program, so I my comments are going straight to the source.  Great!

I have downloaded the program, but it is unable to handle this large a file. It says "Insufficient memory for dib Denoise." The image's dimensions are approximately 12,000 7,000 pixels making a 343MB file.  the 16-bit was even bigger.  I have 8GB of RAM in my system. Will this software handle 16-bit?  I converted my image to 8-bit for working.  What is the max file size it supports?

I reduced the image size to 8,000 x 4,500 pixels with 8-bit color making a 140MB file and it runs now.  The "demo" text is annoying, but I guess this is a free trial.  How do I save?  Unless I am missing something, I don't see how to save and try the rest of my workflow in CS4.  

Your software looks very interesting and promising.  It has done what Neat Image couldn't do.  It needs much improved documentation, though. At $18 USD, it is almost in impulse-buy to try it out.  I wish I could save the file to inspect the results more thoroughly in the rest of my workflow.  It isn't as if all those 'demo' words across my image is going to let me use the image elsewhere. I am also disappointed that it can't handle larger images.  I do mostly panoramic images and 12,000 x 7,000 pixels or larger isn't uncommon for me.

thanks for the advice.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 24, 2009, 05:20:03 am
Quote
I guess you are involved in the creation of this interesting program
My involvement is very marginal: some contribute to web site and to help.

Quote
I have downloaded the program, but it is unable to handle this large a file.
To give you the best answer, I asked to the authors:
1- 16 bit images are converted to 8 bit, the authors claim the results are very very similar
2- No more than 3 GB are available for a 32-bit Windows application. There is not a true limit to the image size, the limit is the memory operative system can allocate.
3- Last, but not least, why do you work on a stich? Denoising is the first (or second) operation to be performed on a photo

Quote
The "demo" text is annoying, but I guess this is a free trial.  How do I save?
If you purchase the product, "demo" is no more present and you can save.
 
Quote
It needs much improved documentation, though.
I'm interested on this point. What do you think is not clear?

Jacopo

Title: Noisy Sensors and 16-bit Workflow
Post by: Jonathan Wienke on December 24, 2009, 06:21:21 am
Quote from: jbrembat
3- Last, but not least, why do you work on a stich? Denoising is the first (or second) operation to be performed on a photo

I agree with this; you should do as much processing as possible (noise reduction, color corrections, tone curves, etc) before stitching. the only thing you should have to do after stitching is manual blending touch-up as needed to hide seams, and possibly a bit of sharpening to deal with interpolation softening.
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 24, 2009, 09:55:15 am
Quote from: Jonathan Wienke
I agree with this; you should do as much processing as possible (noise reduction, color corrections, tone curves, etc) before stitching. the only thing you should have to do after stitching is manual blending touch-up as needed to hide seams, and possibly a bit of sharpening to deal with interpolation softening.

I understand your  point, but it is also a bracketed HDR stitch.  I have experimented with doing the HDR processing first and then stitching, but Photomatix does some odd things with the image dimensions (depending on the config) and it kills the metadata that the stitcher uses to configure itself.  More importantlu, it is difficult to see how exactly you want to process the HDR image parts when you can only see 1/10 parts of the whole image at a time.  As a result it can be difficult to get dark foreground and bright sky the way you want it in one shot.  

So, what I do is to use PTGui's HDR output or exposure fusion.  Then, if it is HDR,  I use Photomatix to process the output file.  

I am not sure about the file size restrictions being based on the memory available.  Many programs are able to handle these big images (pthotomatix and neat image for example).  I believe what they do is break the image into smaller parts and work on them individually.  

Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 24, 2009, 11:08:56 am
Quote from: fike
I understand your  point, but it is also a bracketed HDR stitch.  I have experimented with doing the HDR processing first and then stitching, but Photomatix does some odd things with the image dimensions (depending on the config) and it kills the metadata that the stitcher uses to configure itself.  More importantlu, it is difficult to see how exactly you want to process the HDR image parts when you can only see 1/10 parts of the whole image at a time.  As a result it can be difficult to get dark foreground and bright sky the way you want it in one shot.  

So, what I do is to use PTGui's HDR output or exposure fusion.  Then, if it is HDR,  I use Photomatix to process the output file.  

I am not sure about the file size restrictions being based on the memory available.  Many programs are able to handle these big images (pthotomatix and neat image for example).  I believe what they do is break the image into smaller parts and work on them individually.

This doesn't explain why you want to denoise the final image.
For HDR, you should get  noise reduction, so the problem should be mitigated.
Are you sure the effect you see is noise?

I'm not sure. You posted a color image with some noise on the red channel. But I don't understand if and how that image is related to panoramic HDR work.

Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: fike on December 24, 2009, 03:42:14 pm
Quote from: jbrembat
...For HDR, you should get  noise reduction, so the problem should be mitigated.
Are you sure the effect you see is noise?...

no, I am not sure it is noise.  That is the point of this whole thread.  It looks a bit like noise, so maybe it doesn't matter and noise reduction can help me.  

This image is made up of 33 images.  there were two rows.  5 images in the bottom row.  6 images in the top row.  Each of the 11 images was bracketed with three exposures.  I hope that clarifies.  Yes, I thought that HDR wouldn't have this sort of noise.  In color, the noise isn't too noticeable.  In B&W, the noise blossoms to be a bit distracting.
Title: Noisy Sensors and 16-bit Workflow
Post by: jbrembat on December 25, 2009, 05:40:22 am
After reading your previous post, I found:
Quote
This image was taken at ISO 800 on a 50D

Do you use this ISO value?
If so, that may explain the noise. Your "zero" exposure level is underexposed (noisy).

If you take multiple exposures, you have to put ISO at low level (no or minimum amplification).
To get more light, the exposure have to be increased (you have to change shutter speed).

When you increase ISO, the signal is amplified, but the same thing is true for noise.

If you increase ISO, you get more noise. Noise may be acceptable for color image, but may be nasty on some channel (generally blue and/or red are more noisy).


Jacopo
Title: Noisy Sensors and 16-bit Workflow
Post by: Guillermo Luijk on December 25, 2009, 12:36:07 pm
Quote from: jbrembat
When you increase ISO, the signal is amplified, but the same thing is true for noise.

If you increase ISO, you get more noise. Noise may be acceptable for color image, but may be nasty on some channel (generally blue and/or red are more noisy).

It all dependes on how much exposure time and aperture you can afford, or in other words it depends on how much light you can afford entering the camera.

If you are shooting on a tripod without exposure time limitations, it is clear that the lowest real ISO will produce the least visible noise in the final image.

But if you are handheld shooting in such low light conditions that you cannot get a proper RAW exposure at base ISO, raising ISO will reduce visible noise in the final image. The reason for this improvement is that ISO amplification only amplifies noise produced in the early stages before the ISO circuitry, but it does not amplify noise added afterwards (A/D converter).

Regarding the optimum mix of the RGB RAW channels to have a lowest noise B&W image, I did some calculations time ago and concluded that if a channel has a lower RAW exposure than other in the RAW file, there is always an optimal mix value depending on the RAW exposure of the channels to be mixed that will maximise SNR improvement. Of course this mix is independent of the desired channel combination to obtain the preferred contrast and luminance representation in the final B&W image, we are just talking about minimising visible noise in the final B&W image

Regards