View Full Version : Scanning DPI
I have a Epson 4990 for scanning my 4x5 negatives. I wondered regarding dpi how high should you go for the best quality prints for the biggest size prints. I mean it can go to 12800 dpi and produce a 16gb file but surely there is an over kill point?
Cheers,
Jamie
I have a Epson 4990 for scanning my 4x5 negatives. I wondered regarding dpi how high should you go for the best quality prints for the biggest size prints. I mean it can go to 12800 dpi and produce a 16gb file but surely there is an over kill point?
Cheers,
Jamie
Tests have shown (confirmed by my own experience) that these and the V700/750 scanners do not gather up any more detail than about 2000-2400 pixels/inch.
I scan at 2400 just because my computer needs the breathing room. I have scanned medium-format films at 3200 and then downsampled, but I haven't been able to detect any advantage to doing so.
One problem I have not yet licked is that the scanner seems to posterize the tone curve to some extent, though I haven't yet done enough with color in that scanner to figure out why.
Rick "still getting used to the V750" Denney
I have a 4990 and scan at 2400. As Rick points out, anything more, just gives you bigger files, with redundant pixels.
One way to optimize resolution with these scanners is to find the point of sharpest focus, just above the glass. If you search this forum, you'll find lots of discussion of film holders from BetterScanning.com, as well as making do-it-yourself holders.
One way to get more detail and fidelity - without purchasing either a high-end scanner or the services of a scanning service - is to move from 4x5 to 5x7. You get an immediate boost of 66%, and keep your affordable scanner. An 11x14 from a 5x7 is only a 2X enlargement: any scanner can handle that. With larger film, you'll probably never worry about scanner resolution again.
If you only plan to make images as large as 16x20 from 4x5, these Epsons will be good enough to please all viewers except the most demanding Large Format shooters. Many of them, frequent this forum ;)
AFSmithphoto
4-Dec-2009, 16:55
Tests have shown (confirmed by my own experience) that these and the V700/750 scanners do not gather up any more detail than about 2000-2400 pixels/inch.
I scan at 2400 just because my computer needs the breathing room. I have scanned medium-format films at 3200 and then downsampled, but I haven't been able to detect any advantage to doing so.
One problem I have not yet licked is that the scanner seems to posterize the tone curve to some extent, though I haven't yet done enough with color in that scanner to figure out why.
Rick "still getting used to the V750" Denney
That may be about the optical limit of the 4990, and as long as all the edges in your photograph are vertical or horizontal and land exactly on pixels, thats all you should scan at.
If thats not the case you may find you experience less aliasing and retain more fine detail at higher resolutions.
My tests with my 4990 show noticable improvments when scanning at 3200 dpi and VERY slight improvements over that at 4800.
However, I'm wet mounting with a hybrid betterscanning/scanscience solution which does improve the optical limits of my system so that influences my results to some degree.
Your best bet for determining the resolution YOU should scan at would be to scan several small swatches of the a few images, each at varying resolution, and compare them all at 100%.
Scan at the lowest resolution in which you can discern additional detail over the next lowest setting.
(That is assuming your system can handle the image generated. Scanning at a higher resolution and downsampling introduces a level of interpolation that may actually reduce details.)
Thanks for the replies I had a feeling it was around 2400.
The Epson software is bugged... I try to scan any bigger than 2400 I get a message saying "scanner could not allocate required memory" although I have 200GB worth of free space in my temp drive.
When I try and scan in SilverFast at these high dpi levels there i no problem.. weird.
How much RAM do you have ? What operating system ? Are you scanning b&w or color ? 16-bit, 48-bit ?
Have you cropped the image before scanning ? If you're trying to scan something the size of the bed at high color depth, perhaps you really are running out of room somewhere.
Well Ken I don't see how as the same settings in SilverFast work fine :S
4GB of RAM, on 32 bit system though so 3GB effective. Running Windows XP. Sanning B&W at 16bit.
I am cropping the image from the preview scan also.
Might try and older version of the Epson software :/
4GB of RAM, on 32 bit system though so 3GB effective. Running Windows XP. Sanning B&W at 16bit.
That's exactly how I'm scanning black and white 4x5 negatives, though I'm scanning at 2400. My system is less capable than yours--I only have 2G ram and the system is now 9 years old and running XP-Home without above 175,000 patches. I have been using VueScan, because I have two scanners on the same system and prefer to use one piece of software to operate them both. The version of Silverfast that came with my V750 will only operate the V750, not also my Nikon film scanner.
Rick "whose stock holders seem to be well focused and who has not yet experimented with wet mounting" Denney
Well tomorrow I shall dry a back dated version of the Epson software. Im not the keen on the SilverFast it seems to produce weird results when selecting the options in negafix, like the film and type. The image looks better when I select the wrong film yet Epson software is automatic and produces a lot better image than of SilverFast (tones, levels). So need a fix for the memory problem :/
Try VueScan. It should work without problems.
Nathan Potter
4-Dec-2009, 20:25
Relative to Kens comment about using larger format film; if you don't have the camera to achieve the larger format I suppose you could make a high quality enlargement to say 8X10 film from the 4X5 and scan that. While there would be some loss from the original 4X5 it might be small and manageable, but a significant gain over a 4X5 direct scan at 2400 spi. Has anyone done this and seen a real increase in detail?
Nate Potter, Austin TX.
Ben Syverson
4-Dec-2009, 21:12
make a high quality enlargement to say 8X10 film from the 4X5 and scan that.
If you're going to go to all that trouble, just shoot 8x10 in the first place!
I scan 4x5 at 4800 DPI on the 4990 (target: JPG), and then immediately downsize to 2400. (Photoshop > Image Size > 50% > Bilinear). The resulting 2400 DPI file is much cleaner than if you scan at 2400. I see zero image quality improvement by scanning at 16bit (as long as you do your color correction pre-scan), but a gigantic reduction in speed.
8x10 I scan at 2400 DPI. I'm not willing to wait for 4800 on 8x10, as the small improvement in noise is not worth the time. With a 16x20 lambda print, 8x10 is indistinguishable from 4x5 as scanned using my method... Though you can see it in the file.
AFSmithphoto
6-Dec-2009, 09:40
If you're going to go to all that trouble, just shoot 8x10 in the first place!
I scan 4x5 at 4800 DPI on the 4990 (target: JPG), and then immediately downsize to 2400. (Photoshop > Image Size > 50% > Bilinear). The resulting 2400 DPI file is much cleaner than if you scan at 2400. I see zero image quality improvement by scanning at 16bit (as long as you do your color correction pre-scan), but a gigantic reduction in speed.
8x10 I scan at 2400 DPI. I'm not willing to wait for 4800 on 8x10, as the small improvement in noise is not worth the time. With a 16x20 lambda print, 8x10 is indistinguishable from 4x5 as scanned using my method... Though you can see it in the file.
Out of curiosity do you see zero quality improvement of 16-bit at the print or just the screen?
I don't know of any monitors that display in a 16-bit color depth though I may be wrong.
Also if you're printing on light sensitive paper (RA-4 I believe the process is called), I'm pretty sure those machines are limited to 8-bit so you wouldn't see any difference there, assuming you did your corrections at the scan.
However, if the OP will be doing much tweaking in PS or printing from from an ink based printer, 16-bit is probably the way to go.
The advantages (http://www.kenleegallery.com/bronze.html#16-bit) of 16-bit are seen when you capture and adjust in the bigger color space.
You can always down-sample to 8-bit (or let the device drivers do it) when printing.
Ben Syverson
7-Dec-2009, 16:34
I'm a big skeptic about scanning to 16bit.
There is so much noise inherent to scanned film in the form of film grain and scanner noise that bit depth can become irrelevant.
See the attached. Three images: A, B, C. All three are from a single 4000 DPI Adobe RGB 35mm film scan. All have been converted to sRGB and had pretty aggressive Curve and Saturation adjustments (see screenshot).
One of the images is the original 32bit (8 bits per color channel) scan.
One of the images is a 15 bit version. Each channel has been exported to a 5 bit GIF (!) and re-imported. Each color channel has only 32 values instead of 256. The entire image has only 32,768 possible color values instead of 16.7 million.
One of the images is an 8 bit GIF re-imported. The entire image has only 256 colors!
All three were color corrected at 4000 DPI, then resized for this forum (650 pixels) using Bilinear interpolation.
Ben Syverson
7-Dec-2009, 16:38
Here are the 100% crops from the same, along with the "straight" version without color correction.
Can anyone figure out which is the 32bit, 15bit and 8bit image?
The difference would be far more subtle with a large format image, because the noise would be oversampled that much more.
I submit that if you can't tell the difference between a 16.7 million color image and a 256 color version even after applying color correction, then there's not much point in scanning at 16 bit.
Ben Syverson
7-Dec-2009, 16:48
Here are the histograms for the raw (non-color corrected) images, in order from 8 bit to 32 bit.
Despite massive gaps in the histogram, all three images look good, even at 100%. They would all make excellent prints.
The combination of inherent noise and oversampling when scanning LF eliminates any theoretical advantage of higher bit depths. Quantization error (which is what people are trying to avoid by scanning at 16 bits per color channel) is, in my opinion, totally irrelevant.
Keith S. Walklet
7-Dec-2009, 17:17
Ben, could you repeat the same test with continuous tone areas like sky and water that don't have inherent texture? My experience is consistent with Ken's post. Perhaps the advantage of the higher bit depth in your example is masked by the skin and brick texture in this image.
Mike1234
7-Dec-2009, 17:20
No offense but I think we need a second and third party to perform the same above tests and let's have a full gammut of colors with a full array of graduated tones and densities in the image. Also, is anyone available to view the final prints?
Ben Syverson
7-Dec-2009, 17:21
No offense but I think we need a second and third party to perform the same above tests.
Go for it
Mike1234
7-Dec-2009, 17:34
Nice sarcasm, Ben, and thanks for the challenge. But the only decent scanner I have access to is at work. I intended no offense but comparing a 256 color image to a 16M color is questionable. I say back up your claims with a more challenging image. Not that the one you used is bad... it's actually quite good artistically... just not as challenging as it could be for testing bit depth and gradation.
Ed Richards
7-Dec-2009, 17:37
As Ken says, get Vuescan. It knows how to virtualize the memory for big scans, and, unlike Siverfast, can downsample the file before it is saved to disk. It is the only program to use on limited hardware machines. If you are doing black and white, just scan one channel - red or green - to reduce noise and file size. Set it to save as 16-bit grey scale.
Ben Syverson
7-Dec-2009, 17:41
Ben, could you repeat the same test with continuous tone areas like sky and water that don't have inherent texture?
Here you go. Original un-color corrected image, then the same three scenarios. One is 32bit, one is 15bit, and one is 8bit. All three were quantized before color correction... So you are seeing the quantized images pushed via color correction.
Ben Syverson
7-Dec-2009, 17:43
...and the crops.
This round is a little easier to pick out (maybe because I know the answer), but the very subtle difference would be erased by making a print.
And to clarify, the point of these posts is not to say "256 colors is enough for any image!" My point is that if I can quantize images THIS much and still get something plausible, then 16 bit is truly overkill. The SNR of a film scan just obliterates the difference between 8 and 16 bit.
Mike1234
7-Dec-2009, 17:48
Ben, again, nice images but still limited in hue/tonality... mostly blues and greens with spots of gray and reds. What about some images with the full color spectrum with gradations? I'm skeptical but interested!! And don't forget to include the 256 color image. :) BTW, I'm seeing "graying" of colors in some of these to varying degrees. Could you please explain what I'm seeing?
Ben Syverson
7-Dec-2009, 17:50
I can't win! Not enough sky, too much sky... I don't take pictures of Macbeth color charts!!! :) :)
Ben Syverson
7-Dec-2009, 17:53
BTW, I'm seeing "graying" of colors in some of these to varying degrees. Could you please explain what I'm seeing?
I'm not sure. What areas?
Would this image be convincing?
Mike1234
7-Dec-2009, 18:02
Ben, I'm afraid I may have offended you. I didn't intend to but, if I did, I'm very sorry. It's just that I've dealt with digital images for a very long time and I completely disagree that 256 color images can be indistinguishable from 16M colors in all circumstances. With that said, I'm trying to remain open minded... so you CAN win. You just need more challenging examples to prove it to me. Others may already be convinced. My arguement is that images with limited hues and densities can look pretty good with limited bit depth but the deeper one goes the lousier they look. A solid gray image with no hue or density changes will look just as good with one color as it does with billions of colors. In fact a single color bitmap is just the ticket for such an image. Again, no offense intended.
Ben Syverson
7-Dec-2009, 18:08
It's just that I've dealt with digital images for a very long time and I completely disagree that 256 color images can be indistinguishable from 16M colors in all circumstances.
Mike, I never said this!! And you have not offended me. Re-read my last post. I'm not trying to say 256 color images are indistinguishable from 32bit images in all circumstances. I'm saying that if I can push my scans down to 256 and not see much of a difference most of the time, then surely there's no possibility that I'll see an improvement by switching to 16 bit. That's all I'm trying to say. These 15bit and 8bit examples are just for fun
One more round... original un-color corrected image in previous post
This article, 16-Bits Grayscale Scanning of Ansel Adams' "The Tetons - Snake River (http://www.inkjetart.com/4990/48bit/page4.html) is a pretty quick compelling illustration of why we prefer higher bit-depth:
Original scans may appear indistinguishable - on our 8-bit monitors - but as we continue to manipulate and adjust the images, the histograms of those with lower-bit-depth should start to reveal banding and gaps, faster and more obviously, than those with greater bit depth. Sooner or later, those gaps in the histogram manifest as a "digital" look.
It seems fairly logical to extrapolate the same principles from b&w, to color images.
As the article says (emphasis mine): "The reason the 16-bits scan file retains its quality is because it starts out with a possible 65536 tonal shades* per channel -- allowing one to throw out most of those values still leaves one with much more than 256 (the maximum number of values in an 8-bits per channel image file)".
Ben Syverson
7-Dec-2009, 18:12
...and the crops
Mike1234
7-Dec-2009, 18:15
Ben, there's so much grain (noise) that it's impossible to make a clear decision. Is this 135 film? What speed, 400? Again, please don't be offended.
Ben Syverson
7-Dec-2009, 18:16
This article, 16-Bits Grayscale Scanning of Ansel Adams' "The Tetons - Snake River (http://www.inkjetart.com/4990/48bit/page4.html) is a pretty quick compelling illustration of why we prefer higher bit-depth.
Ken, sorry to say, that article is not convincing at all. Scanning an 8x10 negative at any reasonable DPI introduces so much noise that banding would never be an issue. Never ever.
If you scan at high enough resolution, you could make a 1-bit (100% black or 100% white) scan of B&W and still have a gorgeous continuous tone final print. How much you would need to oversample (ie, how high the DPI would have to be) would depend on the print size.
Ben Syverson
7-Dec-2009, 18:18
Ben, there's so much grain (noise) that it's impossible to make a clear decision. Is this 135 film? What speed, 400? Again, please don't be offended.
In the liquor store image, you're also seeing the pebbly texture of the façade. That's not film grain... But you're right, there's plenty of grain overall, as well as noise from the scanner. That's my point!
The film speed of the provided examples varies... I don't have that info on-hand.
Mike1234
7-Dec-2009, 18:22
Well okay, Ben. I'll not argue the 135 stuff with you. I'll leave that to others. I have always used much larger film which may have a very different criteria.
Sorry for any confusion or vagueness. You might want to re-read my posting.
It's not the original scan we're interested in. It's what the image file looks like, after we have performed substantial adjustments. I may be wrong, but it seems that the article illustrates that distinction quite plainly.
Ben Syverson
7-Dec-2009, 18:28
Ken, all of my examples are after adjustments have been made. Look at the originals vs the finals. The images are quantized then color corrected.
Ben Syverson
7-Dec-2009, 18:31
Well okay, Ben. I'll not argue the 135 stuff with you. I'll leave that to others. I have always used much larger film which may have a very different criteria.
Larger film formats mean bit depth matters less! That's why I chose 135 for the example! In 4x5 or 8x10, you tend to enlarge less than 35mm, so this pixel-level stuff matters much less!
Mike1234
7-Dec-2009, 18:49
Well, Ben... you've given me something to study. But I'm VERY VERY VERY skeptical. :)
Either way, thanks for giving me something to consider.
Ken, all of my examples are after adjustments have been made. Look at the originals vs the finals. The images are quantized then color corrected.
You may be right. I'm certainly no expert.
Could it be that a single layer of color adjustment is not enough to introduce noticeable artifacts ? My impression is that photographers often perform more robust correction to their images. Perhaps I have seen too many Photoshop tutorials ;)
Mike1234
7-Dec-2009, 19:37
TO THOSE WHO WANT THE TRUTH..... few people know anything...
Ben Syverson
7-Dec-2009, 20:00
You guys, I have something really cool to post later, but I have to run at the moment.
I converted one of my 8x10 portraits to 3 bit color... That's right, 1 bit per color channel! When sized down to 300 DPI for an 8x10" print, it looks amazing. Coming soon.
Ben Syverson
7-Dec-2009, 22:32
Okay, so a 2400 DPI 8x10 scan... 19200 x 23834 pixels. First order of business: copy each channel out to a new document, and convert it to 1 bit (pure black & white) using the Indexed Color mode (see attached 100% view). Ah, it brings back fond memories of my first B&W Mac.
Next, I pasted those images into the RGB channels of a new document. A histogram won't help you much at this point -- look at the second image. There's nothing there between black and white! This is literally the lowest bit depth you can use while maintaining independent color channels. There are only 8 colors in the entire image! I saved it as a GIF. When was the last time you opened a 150 MB 8 color GIF? :eek:
Next, I resized to 300 DPI (1/8th size) for an 8x10 print (see attached 100% view). Because of the resizing interpolation, there are more colors now: a massive 16 values per color channel, or 4096 possible colors.
Next, I applied a radical color correction. I boosted the curves and pumped up the saturation. If this is what I wanted to print, it would be ready to go. You would never know that a few steps ago, this was an 8 color image! :D
Ben Syverson
7-Dec-2009, 22:33
For reference, here is the whole image. (All images are flipped horizontally, because I did this from the raw scan). I needed to convert this GIF to RGB to take the screenshot, because the previews for indexed images are not antialiased (see second image)
All of this is to prove that bit depth is NOT directly correlated to image quality. If you oversample enough, you can use anything down to 1 bit. That's how those high-end Korg 1 bit audio recorders work. Instead of sampling at 44kz, they do it at 5.8 MHz.
Even if you're not oversampling, lower bit depths can look as good as higher bit depths if your noise floor is high enough. When you increase noise, you lower your SNR, which has the effect of lowering your effective bitdepth. So banding may be a concern if you're scanning a color slide at 150 DPI. But if you're scanning at high DPIs, you will be introducing enough noise (thanks to the scanner noise and film grain) that 8 bits is more than enough.
16bit for a 2400 DPI or greater film scan will simply always be gross overkill.
You may be right. I'm certainly no expert.
Could it be that a single layer of color adjustment is not enough to introduce noticeable artifacts ? My impression is that photographers often perform more robust correction to their images. Perhaps I have seen too many Photoshop tutorials ;)
Ken,
You are absolutely correct. A single layer of color adjustment will not result in the loss of much information. However, if you do extensive post-scan tonal corrections of 8 bit files they will fall apart and the result will be posterization, no question about it. In fact, even if the scan is made in 16 bit it is advisable to make as many changes and corrections as possible in non-destructive layers.
Sandy King
Ben Syverson
7-Dec-2009, 22:45
You are absolutely correct. A single layer of color adjustment will not result in the loss of much information. However, if you do extensive post-scan tonal corrections of 8 bit files they will fall apart and the result will be posterization, no question about it.
See above. That was 1 bit. Posterization is a non-issue when it comes to high DPI film scans. You can see that for yourself!
You just have to be careful to do all of your adjustments on the full-size film scan, and only downsample to print size at the last possible moment. I didn't even do that in the above example, and it still turned out great!
See above. That was 1 bit. Posterization is a non-issue when it comes to high DPI film scans. You can see that for yourself!
I have made tonal corrections on hundreds of scanned files, most of them very high resolution. I know from my own experience that the danger of posterization is a very big issue, and nothing you could show would convince me otherwise.
I am sure that your experiment proves something but it does not prove that 8 bit files are as good as 16 bit files when a lot of post-scan processing is required.
Sandy King
Ben Syverson
7-Dec-2009, 23:12
I am sure that your experiment proves something but it does not prove that 8 bit files are as good as 16 bit files when a lot of post-scan processing is required.
Sandy, don't confuse a 16 bit working space with a 16 bit scan.
You can scan at 8 bit and then convert to 16 bit if you need to do extremely lossy color correction, and you will have the same results as scanning at 16. The important thing in that case is that you do the operations in 16 bit, not how the file was originated.
On the other hand, it would take some really incredibly radical Photoshop work to make a 2400 DPI 4x5 or 8x10 scan start to posterize. If you have an example of such a thing occurring, I would love to see it. I'm a little skeptical.
AFSmithphoto
7-Dec-2009, 23:26
Okay, so a 2400 DPI 8x10 scan... 19200 x 23834 pixels. First order of business: copy each channel out to a new document, and convert it to 1 bit (pure black & white) using the Indexed Color mode (see attached 100% view). Ah, it brings back fond memories of my first B&W Mac.
Next, I pasted those images into the RGB channels of a new document. A histogram won't help you much at this point -- look at the second image. There's nothing there between black and white! This is literally the lowest bit depth you can use while maintaining independent color channels. There are only 8 colors in the entire image! I saved it as a GIF. When was the last time you opened a 150 MB 8 color GIF? :eek:
Next, I resized to 300 DPI (1/8th size) for an 8x10 print (see attached 100% view). Because of the resizing interpolation, there are more colors now: a massive 16 values per color channel, or 4096 possible colors.
Next, I applied a radical color correction. I boosted the curves and pumped up the saturation. If this is what I wanted to print, it would be ready to go. You would never know that a few steps ago, this was an 8 color image! :D
Mind blowing stuff man. A True Eye opener. Now I must play . . .
Ken,
You are absolutely correct. A single layer of color adjustment will not result in the loss of much information. However, if you do extensive post-scan tonal corrections of 8 bit files they will fall apart and the result will be posterization, no question about it. In fact, even if the scan is made in 16 bit it is advisable to make as many changes and corrections as possible in non-destructive layers.
Sandy King
Yes absolutely we should all use non-destructive layers, but lets be clear that non-destructive is a bit of a mis-leading term. It is totatly non-destructive because it preserves all data from the original file. (The background layer.) However, a levels adjustment layer will create EXACTLY as much banding in the final product as identical changes on a levels adjustment would.
The advantage to adjustment layers is that if you change your mind about the degree to which you wish to make the change, you're not working from information that's already been messed with once. Photoshop still has all the original data and refrences that to re-caculate the lastest tweak.
bob carnie
8-Dec-2009, 07:13
I have a lambda output device, I am willing to output test strips at 30x40 magnification or larger of 16 bit and 8 bit capture.
Our printer is a 8 bit printing device which runs at 200ppi and 400ppi .
I would suggest using lambda flex which is high gloss and shows the most information. I also reccomend for this test running at 400ppi and at a larger magnification.
Those who are interested on this thread may get together , decide on the file criteria, decide who judges the prints and send them around.
I have done these tests and I have my opinions but seeing prints live and not on monitor may dispel or affirm some thoughts going on this thread.
Our needs for the file... Adobe 1998 , RGB , 8bit flattened tiff no channels and simple header name in the file.
I will not handle shipping costs but I will make a series of prints at n/c within reason.
Bob
Ben's point is coming to me, and a little different way of expressing it might help.
His point is that if you scan at high enough resolution, the elements being scanned are so elemental that gross approximations are precise enough to be accurate, as long as you intend to display at a lower resolution than what was scanned.
We do this in various forms of modeling all the time. If we consider an action to be modeled sufficiently disaggregated, the choice of action we are modeling eventually becomes binary. Either the action takes place or it doesn't. It is when all those microscopic actions are summed into larger samples that the micro-error of that assumption is averaged out, leading to an accurate larger sample.
So, color depth is more necessary the less the resolution of the capture.
This is why direct-digital capture is much smoother than scanned film. Film is disaggregated by grain an dye clouds. The scan either has to have high enough resolution to characterize those grain and dye clouds sufficiently, or it has to be low enough to average them into smooth colors. The former does not require much bit depth, but absolutely depends on subsequent aggregation, either by downsampling or by printing it fine enough so that the elements are blurred together in the viewer's eyes. The latter approach is what requires the bit depth. So, I might need more bit depth in my 13-megapixel digital camera than in my 500-megapixel scan. The former will be presented such that each of those pixels must be accurate in and of themselves, and the latter must be downsampled before being presented.
It makes an interesting possibility of scanning negatives at extremely high sample rates, and but at very low color depth, and then averaging to increase color depth.
Rick "who expresses new points back to make sure he understands what is being said" Denney
Ben Syverson
8-Dec-2009, 10:07
Rick "who expresses new points back to make sure he understands what is being said" Denney
Rick, you've got it!!!
I would underscore one additional point, which is the relationship between noise and bit depth. As you increase noise, you decrease effective bit depth even at 100% resolution of the scan. This page (http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html) is good if you want more info on that.
Every film scan above 2000 DPI I have ever seen has had enough noise in the form of film grain and scanner noise that the effective bit depth is far below 8 bits per color channel. So I cannot imagine any situation when scanning at high DPI and 16 bits would offer any advantage.
It's irrational to fear quantization error and banding when it comes to the kinds of film scans we do. 8 bit is more than enough, so 16 bit is overkill.
Keith S. Walklet
8-Dec-2009, 10:44
Fascinating. As described it makes sense. Sounds dangerous, but I'm going to try this at home and see what I come up with.
Thanks for citing the article (http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html).
Concerning noise, the author states:
"...the stepwise transitions of posterization... can't be discerned if the level [of noise] is on average randomly hopping around between any pair of adjacent pixels by more than the spacing between levels"
I see the point, but I wonder if our scans are really as noisy as that. The samples which the author shows, appear rather noisy. Wouldn't it also be true that when scanning at high resolution and higher bit-depth, noise is reduced ?
Ben Syverson
8-Dec-2009, 11:48
Here's a slightly more formal look using a tiny bit of math (bear with me). Say we scan our cleanest, smoothest 4x5, at both 8 bit and 16 bit, and 2400 DPI.
Looking at a patch of sky, we find that the average error is about 10 (Photoshop uses 8 bit values for display in both 8 and 16 bit). That's extremely clean. Look at the attached image. I've never been able to get a 2400 or 4800 DPI scan this clean at 100%.
SNR = 20 * Log10(signal / average noise)
SNR = 20 * Log10(255 / 10) = 28.13 dB
As per the calculation for SNR of a fixed point digital system (http://en.wikipedia.org/wiki/Signal-to-noise_ratio#Fixed_point) (20 * Log10(2^bitdepth)), we can invert the equation to find our effective bit depth.
Effective bit depth = Log2(10 ^ (SNR / 20))
Or, since we already know what 10 ^ (SNR / 20) is (it's 255 / 10), we can simplify:
Effective bit depth = Log2(255 / average noise)
Doing this for our hypothetical "clean" image gives us:
Log2(255 / 10) = 4.67 bits per color channel.
Note that this a measure of the relative signal in the image, so both the 8 bit and 16 bit image have an effective bit depth of 4.67 bits in this case. You could scan the 4x5 at 128 bit, but it would still have a "true" bit depth of 4.67.
If you think my example was too noisy, try it with a noise level of 5. That's fantastically clean. I've never seen a film scan that clean (see second image). But it's still only an effective bit depth of 5.67.
Film scans are simply too noisy to justify scanning at 16 bit. You can see that even 8 bit is overkill. Now you have both the real world examples I showed you with 35mm film scans, as well as the theory/math to back it up.
Thanks - You are a wonderfully patient person !
we find that the average error is about 10
10 what - Is it 10 out of 255 levels of brightness ?
How do get that number from Photoshop ?
Ben Syverson
8-Dec-2009, 12:02
10 what ?
10 8 bit values -- Photoshop's displays use 8 bit values in both 8 and 16 bit mode. So in other words, we're talking an average error of 10 / 255 in 8 bit or 1285 / 32767 in "16 bit" (Photoshop uses signed ints internally, so it's really 15 bit -- ssh, don't tell anyone). The relative error is the same (because the film grain stays the same, and most likely so does the scanner's SNR).
Noise levels will vary across different parts of the image because of the gamma transfer function, so it's best to measure from highlight (such as sky), midtone and shadow areas and take the average. Generally shadows will be noisier than highlights, because of how the "perceptual" gamma curve works.
How do get that number from Photoshop ?
You can use the eyedropper with a large average radius (see attached) to measure average values for a "flat color" area (ideally out of focus), write that down, and then use the "Point Sample" dropper in the same area to see the average error. Just drag the dropper across the area to see how much the RGB sliders in the Color palette jump around. The average amount they stray from your measured "target" sample is the noise level for that area. No point in getting overly exact about it -- just see if the values are straying by about 5, about 10, etc.
Heavens.
This suggests that the effective color depth of the scanner/film system, is less than 5 bits per channel... Oops.
It reminds me of the observation that the effective sampling resolution (in b&w) of so many scanners, is often less than advertised: you can make larger files, but you won't get any more usable data.
Here, you can capture at a higher color depth, but you won't get more colors.
Are these two themes, related ?
If we scan b&w film using the the green channel only - the sharpest sensor on most scanners - are we simply choosing the one with the least noise ? I guess we're always scanning in color, whether we want to or not.
Ben Syverson
8-Dec-2009, 12:27
It reminds me of the observation that the effective sampling resolution (in b&w) of so many scanners, is often less than advertised: you can make larger files, but you won't get any more usable samples.
Precisely! It's actually the same issue. Both cases highlight the diminishing returns of oversampling.
If we scan in b&w, using the the green channel only - the sharpest sensor on most scanners - are we simply choosing the one with the least noise ?
Technically the safest thing to do for B&W would be to scan RGB and do a "smart" monochrome conversion that biases green, includes plenty of red, and a smaller amount of blue. That way you're not leaning on any one of the scanner's sensors, which should give you a better SNR. Although it must be said that some scanners probably already do this internally when scanning grayscale, so testing would be in order.
Ben Syverson
8-Dec-2009, 12:37
Heavens.
Ha :)
percepts
8-Dec-2009, 12:55
considering this forum and its experts have been discussing scanning for and 16 bit over 8 bit for 10 years or more it is surprising this has only just raised its little head. Makes you wonder...:confused:
Its the all the fault of histograms you know...;)
considering this forum and its experts have been discussing scanning for and 16 bit over 8 bit for 10 years or more it is surprising this has only just raised its little head. Makes you wonder...:confused:
Its the all the fault of histograms you know...;)
The discussion here is interesting but mostly misses the point of why we scan in 16 bit. It is due to the fact that if we scan in 8 bit and then make a lot of tonal adjustments to the image file we will get posterization in areas of subtle tonal transition, say in the sky with clouds. If you don't make any changes to the 8 bit file it will be fine to print as is, but once you start to make tonal corrections it is very easy to lose so many values that you get posterization. This is not some kind of deep lost secret but something that people who do a lot of image processing see very frequently.
Scanning in 16 bit helps to prevent the loss of tonal values, but even with files of 16 bit scans if a large number of tonal corrections are made one may run into the posterization problem. If you scan in 8 bit it helps a lot to convert the file to 16 bit before doing any image processing, but doing so does not actually create any more tonal values, it only prevents losing some of them to rounding off when making tonal changes to the file. My Eversmart Pro does not save files in 16 bit so I have a lot of experience with making corrections to an original 8 bit file converted to 16 bit. It is a workable proposition but much greater care has to be observed to avoid posterization than with straight 16 bit scans.
Sandy King
"it is surprising this has only just raised its little head."
Bit depth is still important. Note posting #48 (emphasis mine):
"...don't confuse a 16 bit working space with a 16 bit scan....
The important thing ... is that you do the operations in 16 bit, not how the file was originated."
Ben Syverson
8-Dec-2009, 13:50
Ken's got it
David Luttmann
8-Dec-2009, 14:38
Sandy's got it....
With regards to both Sandy and Emil Martinec: "Momma didn't raise no fool", as the saying goes.
So if I've got anything at this point, it's only enough to prompt me do some testing !
Ben Syverson
8-Dec-2009, 15:21
The discussion here is interesting but mostly misses the point of why we scan in 16 bit.
Sandy, if you go back over the thread hopefully the "ah-ha" light will come on.
Whether you scan 8 bit or 16 bit, your file has the same amount of "real" information (signal), and that information is comfortably less than 8 bit. I encourage everyone out there who is skeptical to prove it to themselves with this workflow:
Scan some film at 2400 DPI or higher, at both 8 and 16 bit.
Open the 16 bit scan
Create a new Action, and start recording
Do all of your most extreme color correction to the 16 bit scan.
Stop recording, and open the 8 bit scan
Convert the 8 bit scan to 16 bit mode.
Run the new action
Compare the 8 bit and 16 bit scan.
You will find that they are identical. I promise. I've done it.
bob carnie
8-Dec-2009, 15:53
Fair enough
now make a 30 x40 of each file at high resolution on a glossy stock and see if they are the same.
Sandy, if you go back over the thread hopefully the "ah-ha" light will come on.
Whether you scan 8 bit or 16 bit, your file has the same amount of "real" information (signal), and that information is comfortably less than 8 bit. I encourage everyone out there who is skeptical to prove it to themselves with this workflow:
Scan some film at 2400 DPI or higher, at both 8 and 16 bit.
Open the 16 bit scan
Create a new Action, and start recording
Do all of your most extreme color correction to the 16 bit scan.
Stop recording, and open the 8 bit scan
Convert the 8 bit scan to 16 bit mode.
Run the new action
Compare the 8 bit and 16 bit scan.
You will find that they are identical. I promise. I've done it.
percepts
8-Dec-2009, 15:54
So I won't be buying a 4990 then. What's the SNR on a drum scanner?
Sandy, if you go back over the thread hopefully the "ah-ha" light will come on.
Whether you scan 8 bit or 16 bit, your file has the same amount of "real" information (signal), and that information is comfortably less than 8 bit. I encourage everyone out there who is skeptical to prove it to themselves with this workflow:
Scan some film at 2400 DPI or higher, at both 8 and 16 bit.
Open the 16 bit scan
Create a new Action, and start recording
Do all of your most extreme color correction to the 16 bit scan.
Stop recording, and open the 8 bit scan
Convert the 8 bit scan to 16 bit mode.
Run the new action
Compare the 8 bit and 16 bit scan.
You will find that they are identical. I promise. I've done it.
Ben Syverson
8-Dec-2009, 15:58
So I won't be buying a 4990 then. What's the SNR on a drum scanner?
It's not scanner-limited; grain limits SNR too. If you have grain that causes a 5 unit variation in a flat color area, it's the film itself which has a low SNR and which is bringing your effective bit rate below 6 bits / channel. So it will happen with any scanner.
Please god, will someone with a drum scanner please scan the same file twice, at 8 and 16 bit and run this workflow?
I know I'm going against the collective wisdom of the ages here, but the collective wisdom is wrong.
percepts
8-Dec-2009, 16:05
It's not scanner-limited; grain limits SNR too. If you have grain that causes a 5 unit variation in a flat color area, it's the film itself which has a low SNR and which is bringing your effective bit rate below 6 bits / channel. So it will happen with any scanner.
Please god, will someone with a drum scanner please scan the same file twice, at 8 and 16 bit and run this workflow?
I know I'm going against the collective wisdom of the ages here, but the collective wisdom is wrong.
So assuming you can get the SNR down very very low with a digital camera, then you can have much better colour fidelity with a digital camera than you ever can with film?
Ben Syverson
8-Dec-2009, 16:25
So assuming you can get the SNR down very very low with a digital camera, then you can have much better colour fidelity with a digital camera than you ever can with film?
I'm going to leave gamut out of the equation (that's a sensor / file format issue) and define color fidelity as the sum of the dynamic range of each channel. If that seems fair.
SNR is synonymous with "dynamic range." The greater the SNR, the more signal is above the noise floor, and thus the more levels of signal you can represent. In that case, for images with equal pixel dimensions, the image with greater SNR will have better color fidelity.
So in order to compare the digital camera with film, they must be the same pixel size. For example, my Canon 5D Mk II produces files with nearly the same pixel dimensions as my Nikon 35mm film scanner. In this case, the 5D wins, because after you factor in film grain and scanner noise, the digital camera has a far superior SNR, thus more DR, thus better color fidelity.
However, if you compare the 5D with an 8x10 image resized to the 5D's size (or vice versa), the 8x10 obviously wins in terms of color fidelity. While the SNR at 100% of the 2400 DPI scan is not great (maybe 30 dB if we're lucky), once we resize, the SNR shoots through the roof.
But in any case (35mm, 5D, 8x10) resizing for print will again alter the SNR. If we resize to a 300 DPI 8x10" print, the SNR of both the 5D and 8x10 (but not the 35mm, most likely) will jump beyond 48 dB (8 bit). Put another way, we are now in danger of banding. We can either do our color correction prior to print resizing, or we should convert to 16 bit.
So it's not such a simple question. SNR, being a relative measure, is only meaningful when comparing two signals of the same sampling frequency (ie, image size).
Don Hutton
8-Dec-2009, 17:35
Here's a tip (and you may find it crucial to your testing) - make the prints... then compare.
Clearly Bob has done it, I know I have, and I'll bet Sandy has. Why go to all this trouble is you're going to exclude the final product?
Ben Syverson
8-Dec-2009, 17:59
Contrary to popular opinion, nothing magical happens when you make a print. If there is no difference in the file, there is no difference in the print.
But if someone is willing to do prints, I'll send them two files and we can go from there.
I just scanned a negative (http://www.kenleegallery.com/html/still/2.html) with lots of areas of smooth tonality, at 2400 spi on my Epson 4990. I scanned in both 16 and 8-bit B&W modes. I converted the 8-bit file to 16-bits, ending up with 3 files.
To each image, I added the same 18 Curve adjustment layers, in the same order - and then merged the layers.
In comparing the 3 images at 100% size, I am unable to distinguish them.
Of even greater surprise to me, I can see no banding or posterization in any of them. Even though the histograms of the images with lower bit-depth look more "spikey", the images themselves are, for all intents and purposes, identical.
Heavens to murgatroid.
(The negative by the way, is 4x5 Kodak TMY developed in Pyrocat HD. The lens was an old "junker", a barrel-mounted 190mm Bausch&Lomb Tessar which came with the 5x7 Kodak 2A I bought at an auction.)
Ben Syverson
8-Dec-2009, 18:32
In comparing the 3 images at 100% size, I am unable to distinguish them.
Crazy, isn't it?! Thanks for posting your results, Ken!
percepts
8-Dec-2009, 18:34
Of even greater surprise to me, I can see no banding or posterization in any of them. Even though the histograms of the images with lower bit-depth look more "spikey", the images themselves are, for all intents and purposes, identical.
Its the all the fault of histograms you know...;)
Don Hutton
8-Dec-2009, 19:00
Contrary to popular opinion, nothing magical happens when you make a print. If there is no difference in the file, there is no difference in the print.
But if someone is willing to do prints, I'll send them two files and we can go from there.You've done prints? Because I have and I can tell straight up, that there are all sorts of issues which I cannot discern on my very nice and accurately calibrated monitors which do reveal themselves in the test prints I made when I did the same exercise from drum scans on a Howtek 4500 (which only has native 14 bit output which is interpolated up to 16 bit by software). I accept that the SNR on a drum scan could negate the result, except that I did the same test on my Epson 4990 about 4 years ago - and I did the whole exercise to print, because that it my final product... The different histograms Ken observed (the interpolated 8 to 16 bit histogram becomes spikey) actually can make become an issue... you may not see it on your monitor, but you may well see it pretty clearly when you make final prints. And of course, I found this out by doing a complete test from scan to print. Not sure what the point of a half test is...
This is actually an exciting development. It means I can scan at 3200--somewhat beyond the capability of my V750--at 8 bit to control file size. Then, I can convert it to 16-bit and downsample it before making big adjustments. Those pixels won't be that accurate in and of themselves, for two reasons: 1.) the optics and mechanics of the V750 won't support it, and 2.) the samples will be small enough to see significant noise in the film itself, even Velvia. I have seen the effects of both of those reasons in my scans.
I have seen banding in my V750, when scanning in 48-bit color at 2400. Maybe I need to scan at a higher pixel density to eke a bit more grain detail, and then downsample that to average out the noise it will pick up. Scanning in 24-bit color would make that a lot easier.
I'm reminded of my old Minolta Multi II scanner that I owned before buying a Nikon 8000. That scanner had a resolution of only 1128 finished pixels per inch with medium format roll film, but it made beautifully smooth and accurate scans as long as you didn't need really big prints. I would often upsample the image as needed and it would still look good. It was smooth and accurate because each sample integrated out all the noise in the film. If we sample at the noise level in the film, then our sampling method only has to have a signal/noise ratio greater by about 3dB to be basically unnoticeable.
Time for some experiments.
Rick "thinking the distinction between scanning depth and working depth are critical" Denney
Don Hutton
8-Dec-2009, 19:02
In comparing the 3 images at 100% size, I am unable to distinguish them.
Of even greater surprise to me, I can see no banding or posterization in any of them. Even though the histograms of the images with lower bit-depth look more "spikey", the images themselves are, for all intents and purposes, identical. Print them Ken - and have a careful look at tonal gradations - the stepping effects which cause posterisation are very difficult to discern on a screen, but are usually very clear on prints...
Will do. It will have to wait until tomorrow AM, but I'll let you know what I see.
Ben Syverson
8-Dec-2009, 19:57
Can't wait to hear the results... all of my color prints have been 8 bit, as my lab's lambda printer is limited to 8 bit sRGB.
Its the all the fault of histograms you know...;)
There is the thing, is it not? The histograms don't lie, but your eyes might!
Somebody might want to go back and look at some of my previous messages. I work all the time with 8 bit files that are converted to 16 bit for correction because I have to. My Eversmart Pro makes the analog to digital conversion in 14 bit but will only save the files in 8 bit. So I have learned to know how much I can push the corrections before I get posterization.
Now clearly I am able to do a fair amount of correction on the 8bit > 16 bit files before an artifact like posterization takes place and is visible on the print. But you can see it happen little by little in the histogram, and at some point, the posterization breaks through visually on the image. It is not a question of "if" you will get the posterization, but "when".
BTW, I scan mostly 5X7" B&W negatives with the Eversmart, and typically scan at 2540 spi RGB, which gives a file size of about 650 mb in 8 bit. I convert this to 16 bit RGB and then apply corrections. I can do a lot of corrections on the file, but if it is pushed far enough I will eventually get posterization. I fluid mount most of my scans and noise from grain is very low.
What I have learned from this thread is that I can probably get just as good a scan in Grayscale, which at 2540 spi 8 bit would reduce file size to 215mb, or 430mb converted to 16 bit for processing.
Sandy King
percepts
8-Dec-2009, 22:20
in PS people refer to curves but in trad printing changing the curve means adjusting contrast. All traditional printers know (at least they should) that if you increase contrast you lose detail and the transitions from one tone to another become more abrupt. Increasing the curve slope in PS does exactly the same. It is a destructive process (although undoable) but has finer control. You don't need a histogram to tell you that or at least you shouldn't.
Ben Syverson
9-Dec-2009, 00:40
in PS people refer to curves but in trad printing changing the curve means adjusting contrast.
Before I printed digitally, I printed photochemically (RA4 and B&W). Comparing "wet" RA4 to Lambda, I find that the digital process gives you more plasticity in the tonal range, at the expense of a very small amount of sharpness.
In other words, a photochemical RA4 print will be marginally sharper, but the tonal range is much harder to control. Your adjustments in the darkroom will either be global (subtract 1 point on yellow) or local and non-repeatable (dodge bottom left corner).
Overall, I am more than happy to trade that tiny bit of sharpness for the flexibility of the "Curves" tool. That is just my perspective. 16x20 is the smallest I print.
As for histograms, I think I have proven definitively that they have no direct relationship to image quality. You can have an "empty" histogram (nothing between 0 and 255) and still wind up with a gorgeous print. What matters is the SNR of the file you send to the printer.
bob carnie
9-Dec-2009, 06:45
Post #50
Contrary to popular opinion, nothing magical happens when you make a print. If there is no difference in the file, there is no difference in the print.
But if someone is willing to do prints, I'll send them two files and we can go from there.
David Luttmann
9-Dec-2009, 06:46
Am I missing something here? Scan in 8 bit, post process in 16 bit? May as well scan in 16 bit and avoid the discussion altogether. I've never had any problems with posterization with 16 bit scans....I have with 8 bit.....regardless of the resolution of the drum or Imacon Scan. Memory is cheap, hard drives are cheap.....scan to the best of your equipments ability and then all discussions like this become pointless.
Oh, and as I've seen posterization in B&W images from 4000ppi drum scans of 4x5 in 8 bit after adjustments, I know beyond the shadow of a doubt that it's better to scan in 16 bit to start.
Oh, and by the way.....scanners don't scan in "DPI"....they sample the film. So, more accurately, you have PPI or SPI. A scanner can't "dot" an image.....but it can sample it. Printers use dots. Sorry, a pet peeve of mine!
bob carnie
9-Dec-2009, 07:06
I agree with this.
I have a 8 bit Lambda printer, I use to scan in 8 bit and do my work all in 8bit.
I must be getting old but around 1000 prints ago we decided to scan in 16 bit, and then at the very last stage change into 8bit for printing.
I seem to recall the reason was because at high magnification on glossy prints we were getting some issues with images that had lots of curve adjustment , colour shifts, or over sharpening .** even though we did not see these issues on screen**
I also have a 16 bit ink jet printer, a all 16bit workflow from scan to print,vs a all 8bit workflow scan to print, magnified to 30x40 with little adjustment on glossy stock**Harmon FB inkjet** will ***not show any difference***.
I am not willing to bet the farm that the artifacting we use to see was from 8bit scanning , I am willing to print out bensyverson files. As his observations are interesting .
From practical application we have shifted to 16bit scan for critical prints , maybe the switch was not needed, It would save us a lot of time , but for now this is how we scan.
Most scanner operators I have spoken with have recommended a flat file to keep as much shadow and highlight detail.
Some clients prefer a very ***strong **** look to their prints,, this requires a fairly significant curve adjustment, if done in RGB without a perfect nuetralization a whole can of whoopass will occur.
Our eyes are funny devices and if left for hours on monitors will create magic and eventually everything looks good. Until the next day when all the artifacting shows its head on the print and you say to yourself wtf was I thinking.
I think if your imagery is very ***soft**** without a lot of adjustment then one would never see a difference on monitor or print.
in PS people refer to curves but in trad printing changing the curve means adjusting contrast. All traditional printers know (at least they should) that if you increase contrast you lose detail and the transitions from one tone to another become more abrupt. Increasing the curve slope in PS does exactly the same. It is a destructive process (although undoable) but has finer control. You don't need a histogram to tell you that or at least you shouldn't.
bob carnie
9-Dec-2009, 07:11
A lambda printer does not care what tag or colour gamut is applied to it.
We print 8bit adobe 1998, srgb , both work.
Can't wait to hear the results... all of my color prints have been 8 bit, as my lab's lambda printer is limited to 8 bit sRGB.
Greg Miller
9-Dec-2009, 07:58
Okay, so a 2400 DPI 8x10 scan... 19200 x 23834 pixels. First order of business: copy each channel out to a new document, and convert it to 1 bit (pure black & white) using the Indexed Color mode (see attached 100% view). Ah, it brings back fond memories of my first B&W Mac.
Your 1 bit file (only black pixels or white pixels; no gray pixels) may or may not exhibit banding. But you have thrown away true gray tonal values at the pixel level of the image, and you have lost control of tonal values and gradations to the interpolation process when you downsize.
Unless you are working with a less than capable computer, isn't it much better to scan and work in 16 bit and retain the true tonal values of the image?
"...isn't it much better to scan and work in 16 bit"
That's the question we're examining together. It's not as straightforward to all, is it is to some.
Ben Syverson
9-Dec-2009, 10:03
Am I missing something here? Scan in 8 bit, post process in 16 bit? May as well scan in 16 bit and avoid the discussion altogether.
Have you ever tried dealing with a 16 bit 4800 SPI (DPI vs SPI noted) scan of 8x10? Or even 2400?
When files get that large and slow, they really get in the way of working with an image. Scanning and color correcting at 8 bit gets around that issue, and since 8 bit is already overkill, there is literally zero justification to scan at 16.
I am not willing to bet the farm that the artifacting we use to see was from 8bit scanning , I am willing to print out bensyverson files. As his observations are interesting .
Bob, PM sent.
Unless you are working with a less than capable computer, isn't it much better to scan and work in 16 bit and retain the true tonal values of the image?
Yes, by all means, if the choice your scanner offers you is 1 bit or 16 bit, choose 16 bit, because you will not have to downsize your file in order to color correct it. However, if your scanner offers the option of 8 bit, you should by all means take it!
Greg Miller
9-Dec-2009, 10:34
"...isn't it much better to scan and work in 16 bit"
That's the question we're examining together. It's not as straightforward to all, is it is to some.
And I framed that question with a reason why 1 bit isn't as good as it may seem.
Greg Miller
9-Dec-2009, 10:37
Yes, by all means, if the choice your scanner offers you is 1 bit or 16 bit, choose 16 bit, because you will not have to downsize your file in order to color correct it. However, if your scanner offers the option of 8 bit, you should by all means take it!
I believe that's the philosophy most of us here work with. Scan and work at the highest bit depth possible for your work environment. Working with lower bit depth can, but not always, degrade image quality and integrity.
David Luttmann
9-Dec-2009, 11:16
Have you ever tried dealing with a 16 bit 4800 SPI (DPI vs SPI noted) scan of 8x10? Or even 2400?
When files get that large and slow, they really get in the way of working with an image. Scanning and color correcting at 8 bit gets around that issue, and since 8 bit is already overkill, there is literally zero justification to scan at 16.
If I recall, you mentioned scanning in 8 bit, but converting during the adjustment workflow to 16 bit in an attempt to maintain the integrity of the 8 bit file. Converting to 16 bit after the fact still means you have a huge file to work on. Maybe it was someone different in the thread....I've lost track.
By the way.....what software are you using to work on an 8 bit file 48,000 pixels wide? And why? Even a 40x50 at 360dpi only requires an 18,000 pixel wide scan. 60x75 needs 27,000. I think 48,000 is simply overkill for anything. With that in mind, a 24,000 pixel wide scan is sufficient....and in 16 bit is only 2.7gb.
When I use 4x5, and have a 16 bit scan at 3200 spi, it's a 1.2gb file. Current workstations (Vic 20's excluded) can easily handle a 1.2gb file. I do so with only 4gb of RAM. And this gives you a print larger than 32x40 at 360dpi....without any worry of posterization.
My point is simple. Scan in 16 bit and don't worry about it. Scan in 8 bit and you have to watch every step with the hope errors don't happen. I prefer not to worry about it and scan in 16 bit. But that's just me.
And finally, while it may be your experience that there is no benefit to scanning in 16 bit....as I've seen posterization occur in 8 bit scans with adjustments that didn't happen with the same adjustments in a 16 bit scan....my experience indicates there is a benefit.
If I recall, you mentioned scanning in 8 bit, but converting during the adjustment workflow to 16 bit in an attempt to maintain the integrity of the 8 bit file. Converting to 16 bit after the fact still means you have a huge file to work on.
I am a bit confused on this as well. My recollection is that the point that was made to me by Ben was that what mattered was that the file be in 16 bit work space, not that the scan be made in 16 bit?
Now, if one could scan in 8 bit and correct in 8 bit and still get great results there would be some point to all of this. But as you remark, if you scan in 8 bit and them convert to 16 bit for corrections, leaves you with a file the same size as if you had made the original scan in 16 bit, which seems pointless since scanners scan in 8 bit as fast as in 16 bit (at least the three scanners I own do).
Sandy King
Ben Syverson
9-Dec-2009, 11:48
I do color correct in 8 bit... I mentioned that some people may want to use a 16 bit working space if they're doing extremely lossy color correction, such as many Curves in a row, some of which cancel each other out.
99% of the work I see on this forum has relatively minimal color correction (or tonal adjustment in the case of B&W) applied, so it would be fine to do those operations in 8 bit. Remember that you're starting with ~5-6 bits at the most, so even 8 bit mode gives you plenty of overhead.
It's impossible for me to follow up with every point made in this thread. I will say this: a lot of people are making faith-based arguments. Until you test this out for yourself empirically, you are essentially subscribing to the "religion" of 16 bit, rather than the science of image processing.
Ben Syverson
9-Dec-2009, 12:09
Bob Carnie has very generously offered to make three 400 PPI 30 x 40" prints in the name of this experiment. My plan is to shoot an 8x10 studio portrait with a softly gradated background to test for banding. Do you guys want to see a Macbeth in there?
Mike1234
9-Dec-2009, 12:24
I would like to see a full range of colors included particularly if they can be graduated.
Ben Syverson
9-Dec-2009, 12:25
"Full range" meaning beyond Macbeth?
Mike1234
9-Dec-2009, 12:26
Macbeth charts are not graduated. :)
David Luttmann
9-Dec-2009, 12:40
It's impossible for me to follow up with every point made in this thread. I will say this: a lot of people are making faith-based arguments. Until you test this out for yourself empirically, you are essentially subscribing to the "religion" of 16 bit, rather than the science of image processing.
Based on what I've read here, many of us have found that posterization problems vanish with 16 bit scans. Thus, I'd say it's based upon our real world experience as opposed to faith in what we want to believe.
If I found that 8 bit scans never exhibited posterization after major adjustments, then I'd glady use it. As I've found different in real world printing, all I can say is that the highest quality is maintained with a 16 bit scan.
Ben Syverson
9-Dec-2009, 12:53
Based on what I've read here, many of us have found that posterization problems vanish with 16 bit scans. Thus, I'd say it's based upon our real world experience as opposed to faith in what we want to believe.
So you've made a 2400 or greater SPI scan at 8 bit and seen banding? I'll give you an FTP site to upload the file to. I am beyond skeptical.
Macbeth charts are not graduated. :)
They are if they're out of focus... :)
Here are some more revealing test results. Not printed out, but evident in a browser. I finally managed to induce some banding, by performing a pair of Curve adjustment layers, seven times: Lighter, darker, lighter, darker, etc., for a total of 14 layers.
See http://www.kenleegallery.com/html/tech/bits.html
Note that the only image which shows banding, is the one which was scanned at 8 bits, but not converted to a 16 bit workspace.
The 4th image in the series, shows that with only 2 adjustment layers, the 8-bit-only image displays no banding. Once we add another layer, banding starts to show.
(For the record, I never correct an image with this many layers: I try to get things right with Zone System exposure and development, and tweak in the scanning phase, and finally perform as few adjustments in Photoshop as possible - such as burning dodging, and spot removal.)
So you've made a 2400 or greater SPI scan at 8 bit and seen banding? I'll give you an FTP site to upload the file to. I am beyond skeptical.
They are if they're out of focus... :)
I have seen banding with scans at 5080 spi with 16 bit scans. But the result was not pleasing and I did not save it in anticipation that someone would question the fact.
It really boils down to how much correction you apply to the image. Whether the file is 8 bit or 16 bit you will get posterization with a great number of corrections. You just have a lot more headroom with 16 bit scans.
Instead of arguing with people who obviously are much more experienced with you in this area I would suggest you do the test yourself. Make two scans of an image that has a lot of delicate tonal range, one in 8 bit and one in 16 bit, then subject both of them to several dozen tonal corrections. If you apply enough manipulations you will eventually get posterization, first in the 8 bit file, and then if you have the patience to persevere, in the 16 bit file.
Sandy King
Ben Syverson
9-Dec-2009, 13:49
Instead of arguing with people who obviously are much more experienced with you in this area I would suggest you do the test yourself.
I guess 12+ years of image processing software development isn't enough experience? I write image compositing apps for fun. Believe me, I know my way around a bitmap.
Folks, my "point" here is not to say "You never need more than 8 bit! 8 bit is perfect!" So yes, if you do 14 conflicting Curves adjustments in a row on an 8 bit image, you will see banding. That's not in question. If you need to do that, you should be doing it in 16 bit mode -- though you can still scan at 8 bit!
I have two simple points. Pay close attention to the wording, otherwise you're arguing with a straw man.
There is no advantage to scanning film at 16 bit, provided you're scanning at ~2000 SPI or greater.
You can do 99% of the color correction you need to do in 8 bit mode, provided you do it at the full resolution of the scan.
bob carnie
9-Dec-2009, 14:00
Ben
When you send me the files , could you also send me the files with no adjustments.
I would like to include some adjustments that I do, on your raw files to test as well.
I will keep my workflow separate from yours and not add onto what you would usually do.
thanks
Bob
I guess 12+ years of image processing software development isn't enough experience? I write image compositing apps for fun. Believe me, I know my way around a bitmap.
Folks, my "point" here is not to say "You never need more than 8 bit! 8 bit is perfect!" So yes, if you do 14 conflicting Curves adjustments in a row on an 8 bit image, you will see banding. That's not in question. If you need to do that, you should be doing it in 16 bit mode -- though you can still scan at 8 bit!
I have two simple points. Pay close attention to the wording, otherwise you're arguing with a straw man.
There is no advantage to scanning film at 16 bit, provided you're scanning at ~2000 SPI or greater.
You can do 99% of the color correction you need to do in 8 bit mode, provided you do it at the full resolution of the scan.
Ben Syverson
9-Dec-2009, 14:01
Bob, sounds good!
Jeremy Moore
9-Dec-2009, 14:07
I have two simple points. Pay close attention to the wording, otherwise you're arguing with a straw man.
There is no advantage to scanning film at 16 bit, provided you're scanning at ~2000 SPI or greater.
You can do 99% of the color correction you need to do in 8 bit mode, provided you do it at the full resolution of the scan.
1. Absolutely NO advantage?
2. What's the 1% you can't do in 8-bit? I figure with such a defined number and your distaste for strawmen (such as appeals to artificial statistics) you have quantified this.
Ben Syverson
9-Dec-2009, 14:11
1. Absolutely NO advantage?
2. What's the 1% you can't do in 8-bit? I figure with such a defined number and your distaste for strawmen (such as appeals to artificial statistics) you have quantified this.
1. No advantage.
2. The 1% you can't do in 8 bit is stuff like 14 Curves adjustments in a row. If you want to get anal about it, "99%" is clearly a rhetorical colloquialism, not a statistic. For the non-native English speakers out there, it means "overwhelmingly, with few exceptions" (as in, "99% of the time, I don't need to explain this idiom" ;))
Greg Miller
9-Dec-2009, 14:25
choose 16 bit, because you will not have to downsize your file in order to color correct it.
There is no advantage to scanning film at 16 bit, provided you're scanning at ~2000 SPI or greater.
Huh?
Ben Syverson
9-Dec-2009, 14:33
Huh?
Please, this is difficult enough without people pulling my quotes WAY out of context. Here is the full quote, emphasis added:
Yes, by all means, if the choice your scanner offers you is 1 bit or 16 bit, choose 16 bit, because you will not have to downsize your file in order to color correct it. However, if your scanner offers the option of 8 bit, you should by all means take it!
You cannot color correct a 1 bit file. There are no values between 0 and 1 (255) to manipulate via Curves or other familiar tools. Thus, you need to resample it in order to manipulate it.
Greg Miller
9-Dec-2009, 15:17
If 16 bit is better than 1 bit, which it is, then it should also stand that 16 bit is better than 8 bit.
To state that there is never an advantage of 16 bit over 8 bit is certainly not valid. If you scan at 8 bit and convert to 16 you are still left with 256 discrete tonal values per channel, whereas scanning at 16 bit gives you 64k tonal values. Those extra values actually exist.
While it may be true that in many cases you will not see banding with 8 bit images, it is also true that scanning at 8 bit results in less fidelity in respect to tonal values than scanning at 16 bit. So stating that 16 bit is never better than scanning at 8 bit is not true unless you remove fidelity from the equation.
I guess 12+ years of image processing software development isn't enough experience? I write image compositing apps for fun. Believe me, I know my way around a bitmap.
[/LIST]
I don't mean to question your smarts or your knowledge, but sometimes even smart and very informed persons are wrong. That goes for all of us.
In fact, I wish you were right. An Eversmart Pro II scanner is work several thousand dollars more on the market than my Eversmart Pro, and the only difference is that with the Pro II you can scan and save in 16 bit, as opposed to 8 bit with the Pro. If I were able to convince people that saving in 8 bit gives results that are just as good as saving in 16 bit that would be worth real money to me.
However, as I have said before, the quality of an 8 bit scan, before any image processing, is just as good as that of the same same made in 16 bit. And one can do a lot of processing by converting an 8 bit scan to 16 bit. I do it all the time.
Sandy King
Ben Syverson
9-Dec-2009, 15:52
If 16 bit is better than 1 bit, which it is, then it should also stand that 16 bit is better than 8 bit.
First of all, that's some incredibly faulty logic. That's like saying "If 16 camera cases are better than 1 case for storing 5 cameras, then it should also stand that 16 camera cases are better than 8 cases." No; 5 cases would actually be fine, and 8 gives you room to move them around. There is no advantage to keeping 11 empty cases around.
Secondly, 16 bit isn't better than 1 bit. I never said that. I said if you have the choice of a 2400 SPI scan at 1 bit and 16 bit, 16 is better. 8 would also be better, but not inferior to 16! You can't make a blanket statement like "16 bit is better than 1 bit" unless you know the sampling frequencies and SNR of the two signals, and what is being sampled. If the 16 bit signal is 1 Khz, and the 1 bit signal is 1 MHz, the 1 bit will be superior -- but if what you're sampling is 1 hz, they're both overkill, so neither is "better!"
To state that there is never an advantage of 16 bit over 8 bit is certainly not valid. If you scan at 8 bit and convert to 16 you are still left with 256 discrete tonal values per channel, whereas scanning at 16 bit gives you 64k tonal values. Those extra values actually exist.
16 bit does not hold an advantage when scanning film at ~2000 SPI or higher. And yes, those extra values "exist," but they do not help you represent the film better than 8 bit. You could scan a 4x5 at 1,000,000 SPI, and those samples would "exist," but they would not improve your image quality.
While it may be true that in many cases you will not see banding with 8 bit images, it is also true that scanning at 8 bit results in less fidelity in respect to tonal values than scanning at 16 bit.
No, that is not true. I have shown that film scans have substantially less bit depth than 8 bit, so 16 bit offers no advantage. "Fidelity" means SNR, and 8 bit images have more than enough overhead above the SNR of scanned film.
Greg Miller
9-Dec-2009, 16:01
All things else being equal, 16 bit is better than 1 bit. Don't try to confuse things by throwing in other variables. If you have a single scanner that is truly 16 bit capable, the 16 bit scan will be more accurate than an 8 bit scan.
And 16 bit provides better fidelity than 8 bit. If I have a tone that is halfway between values 127 and 128 in 8 bit, the 16 bit scan will be more accurate in representing that tone.
Ben Syverson
9-Dec-2009, 16:19
If you have a single scanner that is truly 16 bit capable, the 16 bit scan will be more accurate than an 8 bit scan.
For what? Scanning film? No. The 16 bit scan will not be more "accurate" than the 8 bit scan.
And 16 bit provides better fidelity than 8 bit. If I have a tone that is halfway between values 127 and 128 in 8 bit, the 16 bit scan will be more accurate in representing that tone.
Actually, 8 bit represents the value 127.5 just fine. It does so by dithering, which comes naturally from scanner and grain noise. When you start pushing the file, you will get a gradation of values between 127 and 128.
If you're in 8 bit and keep adding operations one after the other, eventually you may see rounding errors creep up as image artifacts such as banding. If you have the need to torture your files that much, simply convert your 8 bit scan to 16 bit beforehand. But even then there shouldn't be a need to start off with a 16 bit scan.
Otherwise, if you're like most people and keep your color / tonal adjustments in the realm of sanity, 8 bit provides plenty of headroom.
Ken Lee's results seem rather compelling. The problem with 8-bit color depth occurs when making large tonal moves, not when scanning. So, when large moves are needed, convert the 8-bit scan to 16-bit, and then the make the moves--the color space will have sufficient depth resolution to move the tones up a down a lot without being forced into too few bins of tone as a result of the changes. The changes need a place to move the tones to, especially if they are going to be moved back again later.
What it suggests to me is that large tonal moves can be relocated in the workflow to the targeting process, rather than to the correction process. I might scan at 2400 and 8 bits, make a few small adjustments (unaffected by lack of color depth), clone out dust spots, perform any necessary retouching, and use that as my base image. And then, when I want to make a print, downsample appropriately to a smaller resolution as needed by the print, but at 16 bits. Then, I can make my big moves on a smaller file where it can happen in my lifetime. It takes a LONG time to perform some of these actions on my computer. Yes, my computer is old, but then so is my camera.
If I store those big moves as an action, I can then apply them after downsampling to the target at will.
Someone asked, why scan 8x10 at such high resolutions? It seems to be that scanning at high resolution improves the signal/noise ratio because it is able to see into the grain structure of the film. But some of that signal (from the scanner's perspective) is noise (from the film's perspective) in the form of grain. But from the photographer's perspective that film noise may not be noise at all. Having all that extra resolution allows us to control how the pixels are mixed when resampling, instead of expecting the scanner to do it (the scanner might well only collect samples from every other sensor site, instead of averaging two adjacent sites, which would be ore accurate).
Back to the theory--if you scan something at high enough resolution, each sample doesn't have to be that accurate.
Let's use sound as an example. If I digitize an audio source at 1 million samples/second, I could sample at only two settings--sound on and sound off. If I played the sound on a device that could click at 1 million Hz, then I suspect it could produce very finely graded sound--the mixing of those clicks going by so fast wold melt into continuous sound. I could average those 1-bit samples into much larger samples, and as long as those fewer larger samples had the needed depth, they would maintain the signal to noise ratio. So, I need the depth as I downsample, because when there are fewer samples, they must be more precise.
Or traffic flow data. If I try to measure a car's time-space trajectory once a second, my measure of the distance traveled needs to be pretty good--let's say within 20%--to avoid unrealistic departures. If I measure it 100 times a second, my individual sample can be off by 100% or even 500%, and all that noise will average out as I combine those micro-samples into larger samples. My once-per-second sample needs to be able to measure distance traveled to the nearest 5 or 10 feet, and even that can cause departures (such as a the miscalculation of queue delay). That means the sample needs to be in one of about 16 bins--four-bit samples. But the distance traveled in 1/100 of a second is less than a foot, so I can estimate it as either 1 foot or zero feet (i.e., 1-bit samples), and still be more accurate than I was above, once I combine the samples together. Some might say that if I could measure distance to the nearest 1/1000 of a foot, why wouldn't I? There is always a price for unneeded precision--it may be inconsequential and it may not be, but it's never free. This is settled science in the modeling field, as far as I can tell, though people still discuss the ramifications of it.
The question is what constitutes high enough resolution, and that's where I intend to conduct some tests.
Rick "not doubting that Ben's principles are correct, but questioning the thresholds in a context relevant to me" Denney
Ben Syverson
9-Dec-2009, 16:33
Let's use sound as an example. If I digitize an audio source at 1 million samples/second, I could sample at only two settings--sound on and sound off.
How about 5.8 million samples/second? That's how Korg's 1 bit products (http://www.korg.com/Product.aspx?pd=284) work.
Excellent post, Rick! I would suggest doing a test -- on one file, keep it at 100% and 8 bit, do all your color correction, then downsample for printing. On the other file, convert to 16 bit, downsample for printing, and then do the same color correction as the first file. View and print both -- you'll be surprised.
David Luttmann
9-Dec-2009, 17:13
How about 5.8 million samples/second? That's how Korg's 1 bit products (http://www.korg.com/Product.aspx?pd=284) work.
Excellent post, Rick! I would suggest doing a test -- on one file, keep it at 100% and 8 bit, do all your color correction, then downsample for printing. On the other file, convert to 16 bit, downsample for printing, and then do the same color correction as the first file. View and print both -- you'll be surprised.
I have been as well. Making large tonal corrections in 8 bit showed posterization....in 16 bit it didn't.
Anything else?
Greg Miller
9-Dec-2009, 17:34
For what? Scanning film? No. The 16 bit scan will not be more "accurate" than the 8 bit scan.
Actually, 8 bit represents the value 127.5 just fine. It does so by dithering, which comes naturally from scanner and grain noise. When you start pushing the file, you will get a gradation of values between 127 and 128.
If you're in 8 bit and keep adding operations one after the other, eventually you may see rounding errors creep up as image artifacts such as banding. If you have the need to torture your files that much, simply convert your 8 bit scan to 16 bit beforehand. But even then there shouldn't be a need to start off with a 16 bit scan.
Otherwise, if you're like most people and keep your color / tonal adjustments in the realm of sanity, 8 bit provides plenty of headroom.
But even after dithering it still has to store individual pixels as either 127 or 128. So then adjacent pixels then have to alternate between 127 and 128 to render and appearance of 127.5 (whereas in 16 bit mode the effective value of 127.5 can actually be stored). And with color you have to dither 3 different channels. I'm thinking the LF community that is largely infatuated with the ability of the larger format to capture smoother tonal gradations and finer detail isn't so willing to compromise about that if/when the option exists to use 16 bits and preserve as much of that detail as possible.
It also sounds like you scan big and print small. But there is a significant population that scans big and prints bigger. They will not reap the apparent benefits of down sampling that you do.
Ben Syverson
9-Dec-2009, 17:35
I have been as well. Making large tonal corrections in 8 bit showed posterization....in 16 bit it didn't.
Anything else?
What exactly was your methodology? Are you talking about a > 2000 SPI file? If so, forgive me for disbelieving you. I would like to see that file.
Ben Syverson
9-Dec-2009, 17:54
But even after dithering it still has to store individual pixels as either 127 or 128. So then adjacent pixels then have to alternate between 127 and 128 to render and appearance of 127.5 (whereas in 16 bit mode the effective value of 127.5 can actually be stored). And with color you have to dither 3 different channels.
What you're referring to is rounding error, and it's an error of <= (0.5 / 255) per operation. In order to become visible as banding or noise, the errors need to add up enough to exceed the noise level. So if your noise is (5.0 / 255), you would need to apply 10 image-wide operations (such as Curves) before the rounding errors would have a chance of becoming visible.
It is far more likely that any of those operations would introduce an artifact than that you would be affected by rounding errors. It's relatively easy to create banding using Curves. All you need to do is apply one Curve which flattens out the image, making it very low contrast, followed by a Curve which boosts the contrast. Your 8 bit values will be smooshed together, and then pulled apart. Instant banding. If you're irrational enough to do that, by all means, do it in 16 bit. But still scan at 8 bit!
It also sounds like you scan big and print small. But there is a significant population that scans big and prints bigger. They will not reap the apparent benefits of down sampling that you do.
Greg, re-read my 2 claims. I'm not sure where you got the idea that I was espousing the benefits of down sampling. I'm not. I'm saying there is no benefit to scanning film at 16 bit beyond 2000 SPI, regardless of print size. I don't print below 16x20, not that it's relevant in the slightest.
Ben Syverson
9-Dec-2009, 18:31
But even after dithering it still has to store individual pixels as either 127 or 128.
Okay Greg, here is a very clear visual example of why you don't need to worry about banding.
Two images. One is 16 bit, the other is 8 bit. Both are gradients, from 64 - 192. They've had 2% noise added to simulate a VERY clean scan. On the right-hand side of each image, I've applied Levels to boost the contrast. You would expect the 8 bit image to start banding, right? Wrong. See attached.
You can also see the histograms. Look at all the holes in the 8 bit histogram after I apply Levels! But it doesn't matter. The images are visually equivalent.
Ben Syverson
9-Dec-2009, 18:44
Here's an even more extreme example. A 120-130 gradient with only 1% noise, pulled to "full range."
Now look at the histogram! The 8 bit image has only 10 levels, and we can see that in the histogram. Yet there is no banding, thanks to the noise (which would be there in 16 bit anyway).
Noise is the great equalizer. Noise is your friend.
Greg Miller
9-Dec-2009, 19:06
What you're referring to is rounding error, and it's an error of <= (0.5 / 255) per operation. In order to become visible as banding or noise, the errors need to add up enough to exceed the noise level. So if your noise is (5.0 / 255), you would need to apply 10 image-wide operations (such as Curves) before the rounding errors would have a chance of becoming visible.
It is far more likely that any of those operations would introduce an artifact than that you would be affected by rounding errors. It's relatively easy to create banding using Curves. All you need to do is apply one Curve which flattens out the image, making it very low contrast, followed by a Curve which boosts the contrast. Your 8 bit values will be smooshed together, and then pulled apart. Instant banding. If you're irrational enough to do that, by all means, do it in 16 bit. But still scan at 8 bit!
Greg, re-read my 2 claims. I'm not sure where you got the idea that I was espousing the benefits of down sampling. I'm not. I'm saying there is no benefit to scanning film at 16 bit beyond 2000 SPI, regardless of print size. I don't print below 16x20, not that it's relevant in the slightest.
You keep bringing everything back to banding, as if that's the only significant part of an image. I'm talking, and have been for the last several posts, about inaccuracy of tonal vales. An error is still an error, even if it is < 0.5. You cannot get around the fact that the value for that pixel is more accurate when it is stored in 16 bit. With 8 bits, you have only 256 discrete values per channel, and if that was adequate, no-one would ever have bothered moving to 16 bit.
Greg Miller
9-Dec-2009, 19:07
Here's an even more extreme example. A 120-130 gradient with only 1% noise, pulled to "full range."
Now look at the histogram! The 8 bit image has only 10 levels, and we can see that in the histogram. Yet there is no banding, thanks to the noise (which would be there in 16 bit anyway).
Noise is the great equalizer. Noise is your friend.
See my latest post. There are other things besides banding to care about.
Ben Syverson
9-Dec-2009, 19:09
An error is still an error, even if it is < 0.5.
So it's a matter of principle? Why aren't you railing against film grain, which introduces errors far in excess of 0.5? Or the scanner, which also introduces much larger errors? I guess what I'm saying is, why pick on 8 bit encoding, which is the least of your worries?
Look at the images -- the 8 bit image looks the same as the 16 bit image, even after a massive manipulation. The 16 bit image has more data, but they both have the same amount of real visual information.
Trust your eyes, not the histogram.
J.Medlock
9-Dec-2009, 19:32
I don't have a digital scanner yet (maybe one day soon), but I'm really impressed with the discussion in this thread.
Ben is the new Galileo -- keep going man -- I think I'm understanding. :)
David Luttmann
9-Dec-2009, 19:57
What exactly was your methodology? Are you talking about a > 2000 SPI file? If so, forgive me for disbelieving you. I would like to see that file.
Sorry to be the one to point this out....but the resolution has nothing to do with posterization. As I mentioned, I've seen enough 3200 and 4000 spi scans of 4x5 at 8 bit....and still seen posterization in print after large scale tonal corrections.
And as I said, you can avoid this by scanning in 16 bit depth. Really, it's so simple, I'm not certain why this is even being debated. As well, most methodology is based upon experience for most of us....not some religious fervor.
I think enough has been said on this topic. If you feel 8 bit scans are sufficient, then use them. I've seen otherwise countless times regardless of scanning resolution....so I'll stick to 16 bit scans.
Regards,
You keep bringing everything back to banding, as if that's the only significant part of an image. I'm talking, and have been for the last several posts, about inaccuracy of tonal vales. An error is still an error, even if it is < 0.5. You cannot get around the fact that the value for that pixel is more accurate when it is stored in 16 bit. With 8 bits, you have only 256 discrete values per channel, and if that was adequate, no-one would ever have bothered moving to 16 bit.
Don't confuse accuracy with precision. In engineering, accuracy is whether a value is correct, precision is how finely you can measure it. If the film being sampled does not have more than 8-bits of signal/noise ratio, then representing it as a 16-bit image does not increase accuracy. It does increase precision. The 127.5 value is not more accurate if the film is incapable of being more accurate than +/-1 (as an example of a little less than 1% noise). If the grains of the film clump in such a way that one sample is at 126 and another is at 130, then measurement precision that sees to the nearest 0.5 is precision in search of accuracy. If the original has no more than 8 bits of accuracy, the sampling it at more than 8 bits adds precision but not accuracy.
The key is that the noise is higher than the sampling precision. If the sampling precision is less than the noise, then the sampling will magnify error rather than get lost in it.
Rick "who has the difference between accuracy and precision drilled into him" Denney
Ben Syverson
9-Dec-2009, 20:24
And as I said, you can avoid this by scanning in 16 bit depth.
Avoid what? I haven't had any problems with 8 bit scans, so I haven't needed to use 16 bits. It's other people such as yourself who seem to have had bad experiences with 8 bit.
Resolution matters greatly when it comes to banding, because resolution is related to the sampling rate. If you downsample a 4x5 to 300 SPI, you can push the SNR above 48 dB, and then you will have to worry about banding. The solution is to do your color correction at full resolution.
bob carnie
10-Dec-2009, 07:35
Dan Margulis would start yelling at me or anyone who looked at the histogram when image editing ( I spent 12 days with this man over two year period and he never relented.. I could never figure out why.
He would always say are we NOT printing historgrams BUT images.
I am really interested about some of the various observations on how banding,posterization occurs.
For example , who here selectivley paints in local contrast, on a scene. I refer to finding a global contrast that one likes and then looking at the image and deciding too boost the quarter tones, mid tones and 3/4 quarter tones to affect things like sides of white buildings, gravel roads, deep shadow areas.
Is this problematic??
Also in colour if you do not nuetralize the rgb curves and then move to all types of curve adjustment, sharpening or large colour moves, one will see artifacting.
Are these the dreaded things that cause most of our problems. I seem to have made every mistake in the book with my editing learning curve.
Ben -- I have to agree with most here that editing in 16 bit is better than editing in 8 bit, it has been my experience(I do believe your computer skills blow mine away) that I get less problems in 16 bit.
But I am willing to admit that maybe the problems that I attributed to an all 8 bit workflow may something else, like experimenting too much and creating problems unwittingly on files.
I am only speaking about my own experience here and not for others on this forum but I think you will have to send the files to convince me live on prints.
So it's a matter of principle? Why aren't you railing against film grain, which introduces errors far in excess of 0.5? Or the scanner, which also introduces much larger errors? I guess what I'm saying is, why pick on 8 bit encoding, which is the least of your worries?
Look at the images -- the 8 bit image looks the same as the 16 bit image, even after a massive manipulation. The 16 bit image has more data, but they both have the same amount of real visual information.
Trust your eyes, not the histogram.
AFSmithphoto
10-Dec-2009, 12:11
Benny, thank you.
My system can easily handle a 16bit 4800 dpi scan up to 6x12, but at 4x5 it comes to a crawl. My tests have only indicated a slight improvement with my workflow from 3200 to 4800, so UNTIL NOW, I have scanned 4x5 at 3200/16 because I thought I needed the bit depth more than the resolution. Note that I assumed and never tested. (Lazy). I now see that even the slight improvement of the resolution is more valuable than the bit depth. (ESPECIALLY considering that with sheet film my tone adjustments are quite modest as the developing has already got me most of the way there.)
I will have to do more testing, but I will probably continue scanning smaller formats at 16 because hard drive space is cheap, my system still sings at 6x12, and I'm stubborn.
Thank you for your continuing posts and responses to mis-interpretations of your statements. I for one, was happy to learn that I had made some poor assumptions based on an incomplete set of facts. [My understaing of the mathmatics of signal to noise ratio was quite weak, and is now, well, just north of weak. :-) ] I'm sorry that some others are more interested in finding holes in your logic than taking it as a whole.
Adam "Has some testing to do and loves Rick's sign-offs" Smith.
Ken Lee
10-Dec-2009, 12:38
In that spirit, it's time to do some quick testing to determine how high the resolution actually goes on my humble scanner. If I see anything worth sharing, I'll post it.
sanking
10-Dec-2009, 13:58
In that spirit, it's time to do some quick testing to determine how high the resolution actually goes on my humble scanner. If I see anything worth sharing, I'll post it.
I owned a 4990 for a few years and tested it several times with a resolution target. Depending on how optimistically I evaluated the target the effective resolution was between 30-35 lp/mm, or about 1600-1800 ppi.
Sandy King
Ken Lee
10-Dec-2009, 14:17
Thanks Sandy. That's what I'm afraid of :)
This encourages me to stick with... 5x7 !
Ben Syverson
10-Dec-2009, 15:26
Indeed, that's why I shoot 8x10 vs 4x5. Much easier to scan!
I'm getting all the pieces together for the test shot. It will be another couple days before I'm able to get everything scanned and sent off to Bob.
tgtaylor
10-Dec-2009, 15:52
I don't have a digital scanner yet (maybe one day soon), but I'm really impressed with the discussion in this thread.
Ben is the new Galileo -- keep going man -- I think I'm understanding. :)
Me too. But,
if 2400 pixels per inch is the maximum optical resolution of the scanner, then...that's 94 pixels per millimeter. Two of those side-by-side would make a "linear pair" and that would mean...47 LP/mm. However the resolution of my negatives are anywhere from 60 lp/mm to 200 lp/mm. Am I losing anything in the scan?
Thomas
Ken Lee
10-Dec-2009, 16:34
Am I losing anything in the scan?
Of course. We lose data and gain noise whenever we copy an image, by either analog or digital means.
It's like going to a currency exchange ;)
Brian Ellis
10-Dec-2009, 17:56
. . . I am really interested about some of the various observations on how banding,posterization occurs.
For example , who here selectivley paints in local contrast, on a scene. I refer to finding a global contrast that one likes and then looking at the image and deciding too boost the quarter tones, mid tones and 3/4 quarter tones to affect things like sides of white buildings, gravel roads, deep shadow areas.
Is this problematic?? . . .
I do. What's the point?
I don't know why I've been following this thread, I have nowhere near the technical knowledge of many here and haven't been following a lot of it. Maybe it's like the old story about the general who asked the young paratrooper if he liked to jump out of airplanes. The kid said no. The general asked why, if he didn't like to jump out of airplanes, did he join the paratroopers? The kid said he liked to be around people who liked to jump out of airplanes.
Ken Lee
10-Dec-2009, 19:23
"I owned a 4990 for a few years and tested it several times with a resolution target. Depending on how optimistically I evaluated the target the effective resolution was between 30-35 lp/mm, or about 1600-1800 ppi".
Your numbers sound just right for my test results. I don't have a target, but I see a modest increase when going from 1200 to 2400 - but nothing after that, except more pixels.
So roughly 4x is around my max enlargement - which for 4x5 and 5x7, is good enough.
Francisco Disilvestro
10-Dec-2009, 20:36
This has been a very interesting thread so far. There is some useful information, but there are some wrong interpretations that I would like to address.
What is correct: I agree that the practical bit depth will be affected by noise, so it may be that there is no point in scanning at 16 bits if 8 is more than enough for your SNR. I cannot confirm if your value for noise is the correct or not, just that the concept is valid. Just in case, you can actually configure Photoshop to show 16 bit values instead of 8 in the Info palette. If you display 8 bit values the lower value is 1 and the equation will give you 8 bit depth.
What I think is wrong? The assumption that the higher the SPI the lower the bit depth needed, saying that if you scan at a SPI high enough you can use only 1 bit.
This is fundamentally wrong and the example of a supposedly 1 bit per channel 8 color image is not such.
Why? Well one bit per channel images are used everyday in the printing industry. When you prepare separations for offset printing what are you doing? You are preparing 4 (CMYK) negatives or plates that are 1 bit depth and you get a nice print that resembles continous tone. What the example show is that, is the equivalent of doing separation, this time 3 (RGB)
The fundamental issue is that the color information is not in the bit depth anymore, it is in the dithering of each separation. Saying that the composite image of the 3 channels is an 8 color image is just incorrect.
When you downsample, the algorithms just perform a lot of operations and practically convert it to a continuos tone image.
What about the Korg 1bit? It is a similar approach. What it does is very close to Pulse Width Modulation (PWM) where the input signal is compared to a reference signal and the output is either high or low. What you get in the end is a train of pulses of the same level and varying widht. So the information is not in the level of the signal but in the pulse widht. Calling it 1 bit sample is misleading. The big advantages are a reduction in the noise since it is not anymore in the audible frequencies and you don´t need a D/A converter to reconstruct the signal, just a low pass filter, so that´s why you get better quality, This is the same way that high end class D amplifiers work.
Going back to the photo, since the information is in the dithering distribution and not in the sample value you get that empty histogram. This is because the information is not in the sample value. It would be possible to construct an histogram based on the dithering information and it would show you all the information.
Ben Syverson
10-Dec-2009, 23:24
However the resolution of my negatives are anywhere from 60 lp/mm to 200 lp/mm.
In LF? Not likely. Diffraction is probably limiting your results.
But yes, your film & lens are probably outperforming your scanner, but by a relatively small margin. If your flatbed gets 35 lp/mm, you probably get 45 lp/mm on film if you're lucky. A drum scan will get that extra few lp/mm for giant prints when it matters.
bob carnie
11-Dec-2009, 07:27
Ok so the point is many curves on top of each other within the same scene was pointed out as a reason for artifacts.
I was curious as I have seen problems with too many curves.
Jeez Brian you are a pretty funny guy, I almost spilt my milk over your paratrooper story.
I do. What's the point?
I don't know why I've been following this thread, I have nowhere near the technical knowledge of many here and haven't been following a lot of it. Maybe it's like the old story about the general who asked the young paratrooper if he liked to jump out of airplanes. The kid said no. The general asked why, if he didn't like to jump out of airplanes, did he join the paratroopers? The kid said he liked to be around people who liked to jump out of airplanes.
Ken Lee
11-Dec-2009, 10:22
"Ok so the point is many curves on top of each other within the same scene was pointed out as a reason for artifacts."
Many curves - or anything that results in a lot of lossy adjustments. There are many ways to arrive there.
My guess is that color photographs require more adjustments in general, because of the problems inherent in producing images that appear to have realistic colors. Even so, some b&w shooters apply dozens of adjustments, so nobody is immune.
One way to avoid the issue, is to "Just say No" to scenes which are technically challenging. Amateurs have that luxury. :rolleyes:
Ben Syverson
11-Dec-2009, 12:33
What I think is wrong? The assumption that the higher the SPI the lower the bit depth needed, saying that if you scan at a SPI high enough you can use only 1 bit.
This is fundamentally wrong and the example of a supposedly 1 bit per channel 8 color image is not such.
I can't really give a personal signal processing class to every person on the internet, but "no, I am not wrong." If you oversample a signal enough, you can drop all the way to 1 bit with absolutely zero loss in fidelity. In fact, you could create a full color image with only 1 bit per pixel (not 1 bit per color channel, 1 bit per pixel!). You would just need to sample through a Bayer filter.
Your example about halftones is all mixed up. Halftones are not 1 bit. They vary in size. The bit depth of a digitally generated halftone is dependent on the DPI of the platemaking machine, the LPI of the halftone, and potentially the grain of the film. The bit depth of a photochemical halftone would be dependent on the LPI, the grain of the film, and the resolving power of the film.
You are right that the real information is in the dithered image, not the histogram. That is precisely what I was trying to show. People tend to freak out when they see a holey histogram, but it's not always a sign of lost information. The point is that you shouldn't trust a histogram implicitly, because it has no direct correlation to image quality.
percepts
11-Dec-2009, 12:42
Its the all the fault of histograms you know...;)
Ben Syverson
11-Dec-2009, 12:58
Percepts, that's become your catchphrase :)
Francisco Disilvestro
11-Dec-2009, 13:38
Bensyverson, I know something about signal processing, anyway this may be getting off topic.
High sampling rate and one bit depth is a different method than sampling at a frequency just above the Nyquist limit and several bith depts, It is like Pulse Width Modulation vs. Pulse Amplitude modulation. In the first the amplitude does not matter because the information is in the pulse width. In the latter, the information is in the amplitude and that is why you need several bit depth (practically limited as you say by SNR)
My point is that there is not an in between. It is not like as you increase sampling rate you can lower bit rate in a linear way. They are 2 different methods. PAM needs a sampling rate above twice the max frequency to be able to reconstruct the signal, PWM needs a rate over 1 order of magnitude higher than max frequency of your input signal or more. That's why the Korg 1 bit sound recorders use 2.8 or 5.6 MHz.
In PAM you measure the amplitude of the imput signal, in PWM you compare it to a reference signal. It is a different approach.
The way to reconstruct the signal is also different. With PAM you need a Digital to analog converter, with PWM a low pass filter will do (or an integrator).
You are right that you could create a full color image with 1 bit per pixel trough a Bayer filter, the limitations of the human eye will do as a low pass filter
I also agree that histogram do not correlate directly to image quality. It is just that in the 1 bit high frequency sampling approach, a histogram to be useful should be constructed from pulse width data and not from signal amplitude.
Ben Syverson
11-Dec-2009, 13:58
I also agree that histogram do not correlate directly to image quality. It is just that in the 1 bit high frequency sampling approach, a histogram to be useful should be constructed from pulse width data and not from signal amplitude.
No argument there -- talk to Adobe!
In the meantime, it's important for photographers to know that they shouldn't rely on the histogram alone.
sanking
11-Dec-2009, 15:53
In the meantime, it's important for photographers to know that they shouldn't rely on the histogram alone.
Do you believe this is a wide spread problem? I have personally never known a single photographer who relied on the histogram alone!!
Sandy
Ben Syverson
11-Dec-2009, 15:59
It is a widespread problem -- nearly every photographer I have known has had an irrational fear of histogram gaps.
sanking
11-Dec-2009, 16:03
It is a widespread problem -- nearly every photographer I have known has had an irrational fear of histogram gaps.
You need to document this and get it published in a good peer journal. It could be the first step in federal recognition of the problem, which could lead to funding for more research and support groups.
Sandy King
Mike1234
11-Dec-2009, 16:05
I don't use histograms at all.
Francisco Disilvestro
11-Dec-2009, 16:23
Histograms are useful to get additional information about the image, not to evaluate it.
Are those photographers who fear gaps expecting a particular shape from the histogram? Just tell them to convert to LAB mode and check the histogram in the a & b channels. I hope they won't try to adjust levels to bring back a nice shape
Ken Lee
11-Dec-2009, 17:19
"Its the all the fault of histograms you know"
"It is a widespread problem -- nearly every photographer I have known has had an irrational fear of histogram gaps."
"You need to document this and get it published in a good peer journal. It could be the first step in federal recognition of the problem, which could lead to funding for more research and support groups."
I actually had my histograms removed when I was a young man - my parents insisted !
Don't be shocked - back in those days, it was fairly common where I lived. Lots of guys got histogramectomies. If I had to do it today, I'm not sure I could go through with it.
Thanks Sandy - I feel better, just being able to talk about it !
sanking
11-Dec-2009, 18:41
[I][COLOR="DimGray"]
I actually had my histograms removed when I was a young man - my parents insisted !
Don't be shocked - back in those days, it was fairly common where I lived. Lots of guys got histogramectomies. If I had to do it today, I'm not sure I could go through with it.
Ken,
I am guessing that your disease was diagnosed back in the days when most everyone was using MAC systems for image processing and the histogramectomy was the only recommended cure. With the popularity of PCs cryotherapy is proving a popular alternative to the histogramectomies. Several versions of Windows have proven to be highly efffective at freezing the historgram, which of course renders it totally inoperative, and saves the the ordeal of removing the disease from the host organism.
Sandy
Jim Graves
11-Dec-2009, 21:07
It is a widespread problem -- nearly every photographer I have known has had an irrational fear of histogram gaps.
Well, as one of the probably 98% of the population of this board that understands about 2% of the technical discussion herein ... I'd have to agree with Bensyverson ... when I see one of those gapped histograms, I just assume I've completely screwed up my image ... why else would it look like that? ... and that way too much of the original information is now gone and were I to try to print it there would be a huge decrease in detail and tonal values. So, I back up and start over.
This discussion thread has been very enlightening on that and several other valuable points.
sanking
11-Dec-2009, 22:27
Well, as one of the probably 98% of the population of this board that understands about 2% of the technical discussion herein ... I'd have to agree with Bensyverson ... when I see one of those gapped histograms, I just assume I've completely screwed up my image ... why else would it look like that? ... and that way too much of the original information is now gone and were I to try to print it there would be a huge decrease in detail and tonal values. So, I back up and start over.
This discussion thread has been very enlightening on that and several other valuable points.
You might think of the hole in the histogram in terms of the final print as an early indicator of lung cancer in the human body. Early on there is an irritating cough that indicates something is amiss, but the body is able to continue with its normal activities. Then the hole in the histogram gets larger and larger and eventually the image implodes on itself, and the cough becomes more and more incessant as the cancer grows and the body tries to reject it. Ultimately the result is visual disaster with the print, death to the body.
Sandy King
Tiziano
12-Dec-2009, 16:31
This thread was a great reading. Thanks to All.
adonis_abril
12-Dec-2009, 17:54
What's the results on the prints? I have a Colorgetter drum scanner running on a software than can only scan 8-bit. The upgrade software is almost 1k..and that's for having the capability of 16 bit. The information I've ingested here might just save me the 1k :D
Brian Ellis
12-Dec-2009, 18:28
You might think of the hole in the histogram in terms of the final print as an early indicator of lung cancer in the human body. Early on there is an irritating cough that indicates something is amiss, but the body is able to continue with its normal activities. Then the hole in the histogram gets larger and larger and eventually the image implodes on itself, and the cough becomes more and more incessant as the cancer grows and the body tries to reject it. Ultimately the result is visual disaster with the print, death to the body.
Sandy King
And I thought darkroom chemicals were risky! : - )
Brian Ellis
12-Dec-2009, 18:38
"I owned a 4990 for a few years and tested it several times with a resolution target. Depending on how optimistically I evaluated the target the effective resolution was between 30-35 lp/mm, or about 1600-1800 ppi".
Your numbers sound just right for my test results. I don't have a target, but I see a modest increase when going from 1200 to 2400 - but nothing after that, except more pixels.
So roughly 4x is around my max enlargement - which for 4x5 and 5x7, is good enough.
I seem to recall the late and sorely missed Ted Harris testing a 4990 and he concluded that 2100 give or take 100 was the maximum it could resolve.
chacabuco
1-Feb-2011, 08:09
So were the prints ever made? I'm curious to hear/see how this all turned out over a year later.
Powered by vBulletin® Version 4.2.5 Copyright © 2024 vBulletin Solutions Inc. All rights reserved.