M9 noise reduction at medium ISO?

[jaapv]

Advertisement (gone after registration)

Hey Thanks, Jamie. Works great!

[Advertisement]

The Leica Q3 is available for pre-order at B&H Photo Video and Adorama

[scott kirkpatrick]

It's hard to be sure just what changes occur in the processing of a DNG uncompressed file when ISO increases if you compare images taken of different subjects at different ISOs. The best test shots I've seen to date are Sean Reid's images of vegetables shot with tungsten illumination and fixed camera position, at all values of ISO, constant aperture. But vegetables don't have detail that you can be sure is real, they have texture.

 

So I did the following test this afternoon. Shot a nice fairly uniform surface of books in their shelves (smallest details are text on the book spines), in strong indirect sunlight to keep all images as sharp as possible. I used tripod, M9, Summicron 35 @ f/5.6, at a distance of about 2 m, keeping the back of the camera as parallel to the bookshelves as I could. Below I show 100% crops from slightly left of center of the scene. In order, the scenes are ISO 160, 320, 640, and 1250.

 

ISO 160 1/60

ISO 320 1/125

ISO 640 1/250

ISO 1250 1/500

 

What I think I can see in these examples is continuously decreasing contrast in the finest lines in each image. Helvetica text of medium size ("William Christenberry," "SLIDE SHOW," "Behold Man") is unaffected by ISO, but the thick-and-thin fonts are affected visibly. Look at "Bulfinch" (publisher of the Kertesz book) or "Yale University Art Gallery" (second from the left, the Nick Nixon book). This supports Noah's inital observation. The decreased contrast occurs in both high key and low key regions of the image, so it has no connection with the black point setting.

 

scott

[Jamie Roberts]

{snipped}

What I think I can see in these examples is continuously decreasing contrast in the finest lines in each image. {snipped}The decreased contrast occurs in both high key and low key regions of the image, so it has no connection with the black point setting.

 

And yet what is totally clear to me is how much *increase* of contrast there is in the progressive shots around the black point as ISO increase:) Look at (measure) the dark books; they get darker (and the white type stays the same) as ISO increases; therefore you have increasing contrast and less detail.

 

I don't know if that has anything to do with black point and compression but I suspect it does; it certainly has to do with a loss of shadow detail (and if the compression scheme is trying to take the RAW file's black point into account at some certain "wrong" point, then it's conceivable that the loss of detail (the loss of "contrast" Scott is mentioning is at least partly due to the encoding before compression.

 

Personally, I think this increase in contrast is partly an M9 quirk--I've been using the shadow recovery slider on my M9 files where I never touched that with the M8. OTH, I couldn't shoot the M8 reliably above ISO 640 either

 

OTH it could just be simple loss of DR as gain increases, right? In any case, as noise inevitably rises, detail will get destroyed.

 

I'm still a little surprised, though, at the change round ISO 400-640...

[scott kirkpatrick]

I forgot to say these were all shot in uncompressed DNG, processed in C1V5.1, white balanced on the white book in the center.

 

Jamie, maybe one can distinguish the effects of changes in the blacks and near-blacks from high key changes by looking at examples in the set above. Over on the right, "Manzanar" and "Westerbeck and Meyerowitz" both depend on shadow to midtone contrast, but the roles of object and surround are reversed. I find that "Manzanar" remains readable until 1250, while "Westerbeck and Meyerowitz" starts blurring out immediately in the 320 example. When Sandy exhibited the blackpoint shift seen in M9 files with ISO, his values showed that the shift is very small until you get to 1250 and above. That's consistent with "Manzanar" retaining some color and luminance contrast until the example shot at 1250. In "Westerbeck," the text is the part in black, and I don't see how its smearing would be the result of pulling down the black point, plus it happens more rapidly with increasing ISO.

 

Also on the right, the title "American Odyssey," rendered as silver on grey, seems to disappear between 160 and 320.

 

scott

[mjh]

A comparison of the histograms of the ISO 160 (top) and 1250 versions (bottom) shows that the higher ISO shot is a bit darker and has slightly less contrast. The same is true for the ISO 320 and 640 versions, only less so.

 

 

This is global contrast, but how about local contrast in fine detail? Again these samples were taken from the ISO 160 (top) and 1250 shots (bottom):

 

 

As expected, the ISO 1250 sample has more noise which would account for a loss of fine low contrast detail – it gets obliterated by the noise. There are also some strange colours in the letter shapes, easily explained as chroma noise affecting the demosaicing algorithm. But is there any evidence of noise reduction killing detail? Frankly I don’t see any. Noise reduction would destroy low contrast detail by blurring, so we would expect to see a loss of sharpness, of definition. Still, even when parts of the glyphs (especially “f”, “i”, and “c”) don’t manage to surface among the noise, the lettering is almost as sharp at ISO 1250 as it is at ISO 160. What we see is noise and its inevitable effects (also a somewhat darker image with slightly less global contrast), but not much in the way of noise suppression artefacts.

[scott kirkpatrick]

Michael, thanks for the Photoshop analysis. The slight darkening of the 1250 example is evident to the eye, and from your histogram it appears to reflect the black point moving to the right. What do you see causing the differences between the 320 and 640 examples and the starting point at ISO 160?

 

scott

[mjh]

Advertisement (gone after registration)

What do you see causing the differences between the 320 and 640 examples and the starting point at ISO 160?

Well, how dramatic are these differences? This time I have normalized contrast so the histograms are aligned (as far as possible); the samples are scaled to 200 percent (“Bulfinch” was scaled to 400 percent):

 

 

 

Again what I see is a gradual loss of local contrast due to the image being superimposed by increasingly more prominent noise, exacerbated by the demosaicing that at high ISO settings is at a loss to correctly interpret the incomplete pixel data. The strategies to accurately guess the missing values without sacrificing sharpness crucially depend on reliable data, and with increasing levels of noise the data get less and less reliable.

[jaapv]

Is there a difference between RAW converters here? ( I assume the default denoising has been switched off)

[scott kirkpatrick]

I set all sharpening off in C1. And Michael is plotting excerpts from the jpegs that I produced. C1 does have three or four noise reduction defaults that I will have to recheck, to see that they were each zeroed. I did not adjust them when making the examples. And then there is always the question of whether a raw converter really turns off one of its routines when a slider is zeroed. With C1, setting sharpening to zero does not make it zero, but there is a check box that does that. You have to mess with slanted edge charts and Imatest to find out that sort of stuff, which goes beyond this discussion.

 

In the examples extracted above, notice that "American Odyssey" starts to degrade immediately in the ISO 320 version, while "Westerbeck et al." holds up until the 1250 example. Whatever the underlying mechanism, that is worth knowing as you think of how to get impact from a very detailed scene.

 

scott

[Jamie Roberts]

Well, how dramatic are these differences? This time I have normalized contrast so the histograms are aligned (as far as possible); the samples are scaled to 200 percent (“Bulfinch” was scaled to 400 percent):

 

{snipped}

 

Again what I see is a gradual loss of local contrast due to the image being superimposed by increasingly more prominent noise, exacerbated by the demosaicing that at high ISO settings is at a loss to correctly interpret the incomplete pixel data. The strategies to accurately guess the missing values without sacrificing sharpness crucially depend on reliable data, and with increasing levels of noise the data get less and less reliable.

 

Thanks Michael for doing the leg work Obviously the exposure changed somewhat over the course of the shots, and that accounts for the apparent change in black level.

 

The rest is totally what I'd expect from noise increase as ISO goes up.

[Jeff S]

And it's not a stupid question at all... a lot of people don't know the importance of this when developing. Also, IIRC, ACR / LR does this in a totally different way algorithmically than C1, and it's one of the reasons I like C1's results better.

.

 

In fact, Jamie, setting the black and white points in developing was one of the key recommendations of the combined Capture One / Lightroom approach discussed in Luminous Landscape last year...A Workflow Combining Capture One and Lightroom

 

I thought of trying this approach, but to keep things simple, I decided to stick with Lightroom since I like its workflow and file management. Based on threads here, I realized I probably was giving up some improvements, but I'm not the computer/ color management whiz that many here are.

 

In the current LR, I don't know of specific settings for the black point; rather, I'm aware of how to use the black slider to avoid black clipping. Is the LR algorithm built into LR, or is there a way to change individual settings other than the black slider?

 

And, do you know if LR3 will close the gap with Capture One on this issue?

 

I appreciate any feedback.

 

Jeff

[Jamie Roberts]

{snipped}

 

In the current LR, I don't know of specific settings for the black point; rather, I'm aware of how to use the black slider to avoid black clipping. Is the LR algorithm built into LR, or is there a way to change individual settings other than the black slider?

 

And, do you know if LR3 will close the gap with Capture One on this issue?

 

I appreciate any feedback.

 

Jeff

 

Hey Jeff, I'm sorry I'm not more up on my Lightroom, but as far as I know it doesn't do levels (or assign black-point) the same way C1 does, and using the black slider is the way LR does it. I used to know the details but have frankly forgotten them.

 

The LR Adobe crowd is a notoriously individualistic bunch, and they do things their very own way. I wouldn't be at all surprised to hear from them that it's C1 that has the catching up to do in all regards I will say that the files I've seen from LR v3 beta look much improved to me in terms of colour; I haven't done enough with detail to see if they're producing as nice a file as C1, since C1's is so nice I don't feel the need to check right now, to tell the truth.

[Jeff S]

Thanks, Jamie. I'm having too much fun lately making photos to experiment much on the PP side.

 

I guess the good news...for both of us... is that each product seems to get better with every version.

 

Jeff

[scott kirkpatrick]

Jaap pointed out something that I should have looked at before. C1 by default introduces some noise smoothing. In the Focus panel, any M9 image defaults to a luminance smoothing value of 25, and a color smoothing setting of 40 (ISO 160), or 45 (ISO 320), or 50 (ISO 640), or 55 (ISO 1250) or 60 (ISO 2500). There is also a Moire pattern noise cancellation setting which defaults to 8. I set all these to zero for all values of ISO, except for Moire to 4 (it doesn't go any lower). I ran everything through C1 again with the noise settings zeroed.

 

The effect is visible in the size of the cropped jpegs. The ISO 160 file increased 4% in size. The ISO 320 by 7%, the ISO 640 by 12%, the ISO 1250 by 18% and the ISO 2500 by 25%. Howvever, if you put the previous and the new files side by side, or flick from one to another in the same viewing frame, you can't tell them apart, looking at the crop presented above. Just to make sure, I show below the "American Odyssey" book title for each of the ISOs, 160 to 1250 top to bottom:

 

 

I'll upload the five DNG files to YouSendit and post the link in case anyone wants to try another RAW development environment on them.

 

scott

[scott kirkpatrick]

Here are the DNG files used above:

 

YouSendIt: Online File Sharing and collaboration with FTP Replacement - Send Large Files and Email Attachments with Managed File Transfer Solution

 

scott

[mn4367]

I forgot to say these were all shot in uncompressed DNG, processed in C1V5.1, white balanced on the white book in the center.

 

Jamie, maybe one can distinguish the effects of changes in the blacks and near-blacks from high key changes by looking at examples in the set above. Over on the right, "Manzanar" and "Westerbeck and Meyerowitz" both depend on shadow to midtone contrast, but the roles of object and surround are reversed. I find that "Manzanar" remains readable until 1250, while "Westerbeck and Meyerowitz" starts blurring out immediately in the 320 example. When Sandy exhibited the blackpoint shift seen in M9 files with ISO, his values showed that the shift is very small until you get to 1250 and above. That's consistent with "Manzanar" retaining some color and luminance contrast until the example shot at 1250. In "Westerbeck," the text is the part in black, and I don't see how its smearing would be the result of pulling down the black point, plus it happens more rapidly with increasing ISO.

 

Also on the right, the title "American Odyssey," rendered as silver on grey, seems to disappear between 160 and 320.

 

scott

 

Wow, this is real pixel peeping and I thought I was already extreme in this discipline...

I wonder if this has a real relevance. To me the sample shots show that the M9 can deliver realy good results at least up to ISO 1250. I've downloaded the DNGs and ran them through LR 3 Beta 2. Attached are the results with the default sharpening and denoising LR settings and additionally with all settings set to zero (which is unusual in LR). Apart from that I adjusted WB and brightness to achieve a similar look on the samples.

 

Michael

Welcome, dear visitor! As registered member you'd see an image here…

Simply register for free here – We are always happy to welcome new members!