This page is devoted exclusively to problems when attempting to profile a 35mm professional film scanner using an IT8.7 reference target and the program 'SilverFast Ai 6'. The quality of profiles generated by SilverFast and Scarse are evalutated, and examples for a particularly problematic image are shown using different color management methods. The conclusion is that the current algorithms are not capable of producing satisfactory results with full range slides — actually, they destroy shadow details beyond rescue and creates new really nasty color casts!
[First published 2003-12-29]
The setup
I have a package consisting of a film scanner Microtek ArtixScan 4000tf plus scanning software SilverFast Ai 6 (versions 6.1.0r5 and 6.1.0r7 have been used in these tests) from LaserSoft Imaging (http://www.SilverFast.de) connected via FireWire to a computer running Mac OS X 10.2. Besides a target from Kodak (Q-60E3), which I will not use here, I have an IT8.7 target made by Wolf Faust from http://www.targets.coloraid.de. For studying the results, just about any imaging application can be used, such as Adobe Photoshop, Photoshop Elements, PaintShop Pro, Canvas, Gimp, ..., but for the sake of simplicity, I will sometimes use the product name Photoshop in place of any of these applications, or simply say "our image application".
As for monitor calibration, I use a 22" Mitsubishi Diamond Pro 2045u calibrated with PhotoCAL 2.7.7 from ColorVision, using gamma 2.2 (Windows users can therefore rest assured that this will give the same appearance as on a calibrated PC, despite this is a Macintosh). The 2.7.7 version of PhotoCAL corrects a very serious color cast bug in 2.7.6 (reported by me), and version 2.7.6 in turn corrects a minor white point problem with version 2.6. So now everything seems to be okay as far as monitor profiles is concerned.
The raw scan
The problematic image I will show as an example, is a somewhat under-exposed Fuji Sensia 100 slide, scanned at 4000 ppi, so the total size is about 5700x3700 pixels. At 48 bits per pixel, the tiff image is about 130 MB — too much to show here, and we do not need to see the whole image. I will only show you a shadow part of the image, whose size is about 1400x1000 pixels (6.6% of the total area). Although the slide is somewhat under-exposed, I can assure you that there are other parts of this image that are bright, so it would absolutely not be a conceivable solution to just crank up the lamp intensity to compensate for the under-exposure, since that would completely wipe out the highlights.
Actually, I have already optimized the lamp intensity, which can be done from within SilverFast. I have set the lamp intensity to precisely that level where GS0 (clear film patch) of the IT8 target reaches RGB = 253-254. By this method, we have as high setting as we can get without clipping the highlights, while getting as good shadow performance as we can get. The exact lamp setting is: +8 units overall, plus an additional +6 and +4 units respectively for green and blue.
The "neutral" gamma setting of this particular scanner is about 2.0. This means that at gamma 2.0, the scanner scans patch GS10 of an IT8 target to a "brightness" that is about the same as an IT8 calibrated scan would produce. In other words: at gamma 2.0 the image gets a brightness that is consistent with what an IT8 calibrated scan would produce. This setting is sometimes regarded as the optimal setting (e.g advocated by Ian Lyons at Computer-Darkroom). However, as we shall see, because of the way it is implemented in SilverFast Ai, this is far from optimum. (For your information, the "limit gamma slope" check box is deselected.)
The following image is scanned with the 48 bit HDR setting in SilverFast, using no color management at all, but with a gamma correction. So it is as clean and raw as you can get. It shows the problematic shadow area of the slide. To make it "browser compatible" for your convenience, it is converted to 8 bits per channel, jpg-compressed only gently (Photoshop-setting 11), and with no profile embedded: click to see in a separate browser window (264 kB).
First of all: turn bright lights off in your room! You may absolutely not have any lights that illuminate your screen directly, since this image is a bit dark. You will need to look at it for a while to get adapted. As you can see now, it is dark and noisy, but there are details there! And for many of the details you know that they are white or neutral grey things, plus the reddish tiled roof. The noise is caused by the CCD element in the scanner, despite this scanner is supposed to be the best you can get in this respect (apart from drum scanners of course)! Nevertheless, you can also see film grain. The fact that it is converted to 8 bits/channel and jpg-compressed might have you suspicious, but I can assure you that the difference in visual appearance is not very big — it is actually very small. If you really want to play with the original, okay here it is (6.5 MB 48 bit tiff image). (Curious where the picture is taken? It is shot from the Galata Tower in Istanbul.) We will soon assign a profile to this image, but first a few words about how the profile is made:
Using profiles generated by SilverFast
The profile is made using SilverFast Ai 6.1, according to instructions and also according to a few hints by Ian Lyons (see Computer-Darkroom). To be on the safe side, color management was turned off while making the profile. Also, it is important to use the same gamma setting and the same lamp setting. Actually, I made several profiles for different gammas, as we soon will see.
To see what happens when the raw image gets color managed, we first assign a profile to it. The rendering intent should be relative or absolute colorimetric (and Black Point Compensation should not really make any difference). I assign it the profile I made for gamma 2.0. The first thing you see then is a terrible mess of the shadows!! Absolutely ridiculous! It even inverts some dark areas!
This doesn't seem to be possible to show directly on the web, however, so to see that you must either download the image (which contains the profile) and open it in a icc-aware image application, or assign the profile yourself to the non-managed "raw" image you have just seen above. Click here to download the image that contains the scanner profile (688 kB, but note that it will not show up in your browser as it should since your browser is most likely not icc-aware!) The scanner profile I used is here (428 kB).
However, that appearance is anyway not the final one, since we must convert it to our working space, such as 'Adobe RGB (1998)'. This can, as far as I have tested, be shown pretty well on the web: If I convert it to Apples version of sRGB, and then save without profile, the appearance in a browser on a gamma 2.2 monitor is actually very similar to the true 48 bit 'Adobe RGB (1998)' version! Click here to see it in a separate browser window! (780 kB)
The end result is that it doesn't look quite as ridiculous as the version before working-space conversion, but the shadows are completely destroyed and a fierce colored noise has appeared instead, and the brighter areas have come to get an awful greenish/cyan cast that certainly is not in the real slide!
The above procedure is not the normal working flow with SilverFast; instead, you would set up a proper color management, so that SilverFast automatically assigns the scanner profile and converts it to your working space. I provided the above procedure to be able to follow exactly what happens. The following image is a web adapted version of what SilverFast produces when all color management is set according to all rules: Click here to see it in a separate browser window (596 kB). As above, this is converted from 'Adobe RGB (1998)' to sRGB and jpg-ed with very mild compression (Photoshop quality-setting 11). The appearance in a browser at gamma 2.2 is actually very similar to the real 48 bit tiff image.
If we compare this last one with the image we made by assigning and converting profiles in our image application, we see that they are reasonably similar (but not exactly equal — SilverFast makes something in the conversion a bit different). Actually, I find the latter "SilverFast-converted" image even worse than the one we converted in our image application, since it is even greener and even more shadowless. I have tried different variations, but they all come out the same, i.e: saving as 48 bit HDR; saving as 48 bit Color (not HDR); with and without "limit gamma slope"; different rendering intents — they all give the about same result.
THE FINAL COMPARISON!
As mentioned in the beginning, the slide was a bit dark (which is very common with slides!), so it really needs (as most of my slides) to have some 'curves' tweaked in the lower region to lighten it up. It is not until this stage we really can see how well the color profiles are working — any flaws in the profile will be grossly amplified when you try to lift the shadows a bit! So I take the previous image, i.e the one produced by SilverFast in a properly setup color managed work flow, increase the 'curve' a significant notch (RGB curve input value 26 mapped to 51, so we clearly can see what is going on). Do the same with the completely clean non-managed image I showed you in the beginning, so we can compare them and see just how much damage the profile did to our image. (JPG compression here is Photoshop-setting 8 — higher would be pretty meaningless for this comparison):
Now that is a comparison! A completely and totally useless image!! Well beyond rescue! I am pretty sure, that no matter how good a Photoshop artist you are, you simply cannot fully repair the damage in the latter image! That's just outright impossible!
But color management was supposed to make life easier! Well, obviously not if the profiles are as bad as the ones generated by SilverFast Ai 6 and you expect usable shadow detail...
You think your setup is better?
I am pretty sure you now think that your profiles are much better and that you do not have the problems I have(?) To show that my IT8 image looks as good as you would normally expect, I show you the IT8 image here! Note, however, that it is in 'Adobe RGB (1998)' space, so it will not show up in your browser as it should (since your browser is most likely not icc-aware)! I couldn't find a conversion strategy that rendered the IT8 image as good in a browser as it does in the imaging application. The closest I got is an sRGB conversion (untagged) that you can see here! Although jpg-ed and sRGB-converted, it nevertheless shows that there is nothing drastically wrong with it.
It is also possible to compare the patches with the data file (characterization file) to ensure that it is at least not wildly in error (which I will do later in this article). The IT8 image can either be obtained directly by SilverFast in a color managed flow, or by the first method above, i.e scanning a raw HDR image that later is assigned a profile and converted to Adobe. I have verified several times that it does not matter which way you take it, they will be almost identical. (But versions of SilverFast before summer 2003, however, behaved much more erratic in this respect.)
Solutions?
The above images were presented to LaserSoft in October 2003, in a request for support. The response I got was that they use the preview for the IT8 calibration! Yes, that's right, they use the 8 BIT PREVIEW!!! Although that is a serious flaw by itself, I think there are also other factors involved, most notably interpolation/extrapolation.
The first request for support in this issue of bad profiles, however, was taken already in April 2003, so it's been quite a while since then, and it still cannot produce fully professional results! In the months before that, the application (or the driver) did not work at all, since it mostly either just crashed, or completely froze at the end of 4000 ppi scans — it would only work (sometimes) if scanning small files (not my full 130 MB-files). Now it does at least store files, but not yet with professional color quality. There is some hope, however: I was told that the developers at LaserSoft are discussing if they should use 16 bit data for the calibration. Also, later in this article, I show tests of profiles made by a free program called Scarse.
The issue about using only 8 bits per channel for calibration had me investigating what happens when I scan at other gammas (described below). The 8 bit limitation, however, is not the only cause of the loss of fidelity. The shadow problem occurs to a large extent because the IT8 target hardly has any dark colored patches! E.g, there is really not a single really dark green color! (Except one in the custom part of the Fuji target.) A good profiling software must therefore do an intelligent interpolation/extrapolation from the colored patches in the IT8 target down to the really dark values; this is an area where SilverFast fails.
Gamma 1.0
It turns out that shadow results are better at gamma 1.0! I have tried gammas from 1.0 up to 2.5. First, we can check the accuracy differences between them by first making profiles for these gammas, then scanning the IT8 target itself at different gammas, and then read the Lab values of each patch comparing them to the values in the characterization file.
Assigning/converting profiles can be done either inside SilverFast or scanning them raw (HDR) and do it in your image application — the results with either path should be just about identical, but as it turns out, at gamma 1.0, the values will be way off if we let SilverFast do the color management! (They will be too dark and the grays will not be neutral.) This difference is not a matter of choosing different rendering intents! At gamma 2.0, the two paths give roughly the same results, but at gamma 1.0 we must do everything in our image application!
In theory, all gamma versions should give exactly the same end result, but scans at high gammas are much more inconsistent than those at low gammas. E.g, with gamma 2.5, patches in column 3 to 4, rows A through G are just way off! The accuracy of the scans at gamma 1.0 and 2.0 are, however, fairly equal — some patches are better with 1.0, some are better with 2.0 — but overall I think there is a slight advantage to the 2.0 version, except for the patches near black.
Generally speaking, the highlights are fairly okay, but darker muted colors are a bit off, and the worst are dark reds and greens in columns 3 through 8. All versions seem to be pretty much off near the blacks; GS22 for instance (the black just to the left of the white line) should be [1.64, 0.66, -0.05] in Lab coordinates, but the gamma 2.0 version shows about [5, 2, 4], so not only is it too bright, it is not neutral. the difference in 'L' between patch GS22 and GS21 should be 2.0, and another 2 units to patch GS20. SilverFast gamma 2.0 version, however, shows only a difference of about 1 between these patches, so the shadows are heavily compressed (and not neutral!) This is also precisely my experience when scanning slides! The gamma 1.0 version is slightly better: GS22 is [4, 1, 1], and the difference to the next patch is about 1, and another 2 units to the next one.
Keep in mind that we need to assign/convert profiles in our image application — if you let SilverFast do it, then the gamma 1.0 version will be way off; e.g patch GS18 should be L=14, but it becomes only 10 using SilverFast, whereas we really get 14 if we let our image application do the conversion. Also, Gray patches GS16 and 17 are not neutral at all using SilverFast.
This way of checking the quality only concerns the patches in the IT8, not the behavior outside/between the patches. The shadow performance can be a whole different issue, as we will see in the next paragraph:
The following is the troublesome image, now scanned at gamma 1.0 HDR "clean" and color managed by our image application, i.e assigned the right profile and converted to 'Adobe RGB (1998)'. It has then been "tweaked" by increasing the curves from 26 to 51, exactly the same way as was shown under the heading "THE FINAL COMPARISON" above. To be "web-friendly", it has been converted to sRGB just as the previous one, but in this case it seems, for some reason, necessary to actually look at the image in a program that is profile-aware, otherwise there is a slight cyan color cast to the image (and if you are using Safari, it will also be a bit too dark, whereas Explorer and Netscape get the contrast/brightness right). Click here to see the image in your browser (100 kB). Even though it is not exactly accurate in your browser (unless it understands sRGB) this looks much better! But still not good enough...
The following is the same thing but color managed by SilverFast, and "tweaked" the same way: Click here to see the image in a separate window (84 kB). The image is in sRGB, but not tagged at all, and it should look okay in a normal browser at gamma 2.2. In this image, there is a color cast and much shadow detail is gone. As you can see, this version is much darker than the one color-managed by our image application.
Remember, both of these images are "tweaked" by raising the curve! Still, the SilverFast-managed one is very dark! Both of the above images should be compared with the two images shown under the heading "THE FINAL COMPARISON" — they are directly comparable! As you can see, there are enormous differences, but none of them are good! So far, the "best" is probably the totally non-managed scan, although it is a bit grayish
Scarse
Scarse is a free color calibration and management software package under the hat of SourceForge (http://sourceforge.net) aimed at producing ICC profiles, see http://www.scarse.org. It is still only at an alpha 0.3 stage, so it is not very easy to use, and still has some problems. To install the package, you need to first download it plus a package called libtiff, and then compile it yourself, provided you have the right compiler etcetera. Usage is command-line oriented. The package can read 48 bit TIFF images, and the internal precision is 56 bits per channel!
The author of the package, Andrei Frolov, was very kind to produce a set of profiles for me, based on scans I had done with SilverFast in "clean" mode, i.e with no color management, and saved as 48 bit HDR images with various gammas (although not at full resolution). I have done the same experiments as was done above with SilverFast, for gammas 1.0, 2.0 and 2.2.
The profiles generated by Scarse use absolute colorimetry, so the first thing one has to do to get an image right, after having assigned the right profile to the image, is to convert it to a working D50 space (such as a D50 version of the Adobe space) using the intent Absolute Colorimetric; then you can convert it to your usual space, such as 'Adobe RGB (1998)' (which is D65) using the intent Relative Colorimetric. If you don't use this procedure, the white point will be way off.
The first thing I did was opening the non-managed IT8 images and assigning them profiles and converting them to the working space 'Adobe RGB (1998)'. This was a delight to see, since the accuracy of the patches were way much better than the SilverFast profiles! Interestingly, both the image at gamma 1.0 and the one at 2.0 produced almost identical results (within one unit in the Lab coordinates)! The image scanned at 2.2, however, was significantly (but not drastically) off in several patches. In total, very good quality, especially if you stay away from gamma 2.2!
That said, I think the GS22 patch may perhaps be slightly too dark (should be 1.64, but I see only 1 here) and for the image scanned at gamma 2.0, the step between GS22 and GS21 is too large (shows a difference of 3 here, should be 2) and this is also seen visually. Also, the GS22 patch under large magnification looks artificially blotchy, as if the grain had been strongly amplified. It is somewhat less blotchy at gamma 1.0. So These last differences make the gamma 1.0 version a tad better.
Actually, if scrutinized under high magnification, it is revealed that this blotchiness is strongly amplified at the stage of profile conversion from the scanner profile to a working profile — it happens at the stage of conversion to the D50 version of the Adobe space. Before the conversion, it looks very similar to the SilverFast profile at gamma 1.0. But whereas the blotchiness with SilverFast is reduced when converting it to 'Adobe RGB (1998)', it gets strongly amplified (but made neutral and more black) with Scarse! So I suspect there is something wrong here with Scarse, and as we soon shall see, it strongly affects our image:
The following image is treated just as we did with the SilverFast version, but in this case we must do all color management in our image application, since we need to get the white point right by first using absolute colorimetric and then relative colorimetric to get to the working space 'Adobe RGB (1998)'. So we take the scanned gamma 1.0 "clean" HDR image, assign the right profile and convert it in our image application, then "tweaked" by increasing the curves from 26 to 51, exactly the same way as was shown under the heading "THE FINAL COMPARISON" above. It has been converted to sRGB, just as the previous tests, but for some reason, it is actually necessary to look at the image in a program that is profile-aware, otherwise there is a slight cyan color cast to the image (and if you are using Safari, it will also be a bit too dark). Click here to see the Scarse image in your browser (212 kB).
The above image was not really a nice sight, but the blotchiness seems to be a problem that might be corrected in some future version of Scarse(?). The image should be directly compared with those images under the heading "THE FINAL COMPARISON".
CONCLUSION
Unfortunately, there is currently no method to get SilverFast Ai 6 to produce good color managed results. The least destructive methods so far is to use profiles generated for gamma 1.0 and scan at 1.0, but then we must do all color management in our image application rather than in SilverFast. A promising, but currently somewhat awkward option, is to use the free software Scarse to generate profiles, which have high accuracy except near black. For the future, to make Scarse profiles easy to use, it is necessary that the program is changed so that it uses relative colorimetry — otherwise we simply can't just enter the generated profile in SilverFast CMS settings as our scanner profile to use. But also SilverFast must be changed so that it does profile assignment/conversion the same way as e.g Adobe Photoshop does.
Epilog
I have also experimented with the program VueScan, and used profiles generated by VueScan from files that have been created by scanning the raw IT8 target by SilverFast. That may very well be a questionable approach, and there were white point issues I could not get around, so I don't want to show any results here. Overall, VueScan profiles seem to have some merits, but very clearly also some problems. I have also had a profile made for me by Ian Lyons in his HDR version of SilverFast, based on a low resolution image. The result shows that the HDR version is better than the one generated by SilverFast Ai. However, also here, some white-point problems crept up so that I cannot assign the profile to an image and have the white point correct.
In any case, users should definitely require that their scanner application (e.g the Ai version of SilverFast, not just the HDR version) should be able to produce fully color managed images with device profiles that work in the whole dynamic range of modern slides!
First published 2003-12-29
Harald E Brandt Hägersten, Stockholm Sweden |
Photo: http://photo.bragit.com BragIt: http://bragit.com |
Email:
|