QImage versus Lightroom for Printing

[mchaney]

Well, not really a fair example above, since you applied double the sharpening strength with DFS (3/200) than you did with USM (3/100).  DFS, like USM, is actually only sharpening a 3 radius but as you can see, it sharpens the concentric rings that make up the gradient.  In fact, that is "real" detail that is being sharpened too: it may only be single steps, but it is detail.  And when you go to 200% strength, those are going to show.  As you can see by my example (attached), the USM version does show some accentuating of the gradient where it darkens/brightens the gradient steps inside the circle, but just not as much as DFS.  And DFS has no dark halo on the outside and bright halo on the inside.  In fact, you could probably go with a lower setting here like 3/50 with DFS and get similar results to USM... without the halos.

BTW, original in the center, USM on the left, DFS on the right: zoomed to show individual pixels.

Mike

[Robert Ardill]

In the Fraser/Schewe three steps of sharpening, I appliy the first two steps to the native image until it looks the way I want it on the screen.

The final step, output sharpening, is applied by Qimage based on output size, and any adjustments for specific paper.  I do not have to remember the differences by paper.  Once I have tested and arrived at what I consider the correct settings, Qimage remembers and will use the same settings whenever I select that paper.

Hi John ... yes, a lot of people seem to use that sequence of sharpening. Personally, I prefer to do all of the tonal and color adjustments with no sharpening (I'll temporarily add sharpening in Lightroom or in the ACR Smart Object as it's difficult to work on an image with no capture sharpening); then I resize the image (with capture sharpening off); then I do the capture sharpening with an edge mask (or using ACR with the Masking feature as that effectively gives an edge mask); then if appropriate I may do some 'creative' sharpening; and finally I will sharpen for output, again usually with an edge mask or with high-pass sharpening, or both.

Of course the disadvantage is that for every print size I will have to go through the sharpening saga, but once the workflow is set up it's quick to do.  If I had to do many prints every day I wouldn't do it this way, needless to say!

But I think this is a better way of sharpening as I believe that it does the least possible damage to the image.

Robert

[Robert Ardill]

Well, not really a fair example above, since you applied double the sharpening strength with DFS (3/200) than you did with USM (3/100).  DFS, like USM, is actually only sharpening a 3 radius but as you can see, it sharpens the concentric rings that make up the gradient.  In fact, that is "real" detail that is being sharpened too: it may only be single steps, but it is detail.  And when you go to 200% strength, those are going to show.  As you can see by my example (attached), the USM version does show some accentuating of the gradient where it darkens/brightens the gradient steps inside the circle, but just not as much as DFS.  And DFS has no dark halo on the outside and bright halo on the inside.  In fact, you could probably go with a lower setting here like 3/50 with DFS and get similar results to USM... without the halos.

BTW, original in the center, USM on the left, DFS on the right: zoomed to show individual pixels.

Mike

Hi Mike,

Yes, I applied a lower amount to the USM sharpening because it's much more aggressive than DFS.  Photoshop Smart Sharpen is closer to DFS in that sense, so with Smart Sharpen a value of 200 would be close to the same value with DFS.

At any rate, I agree that a small radius and small amount is good - but sometimes a higher amount is needed and IMO DFS is potentially problematic then.

The thing is, that sharpening relies on accentuating edges ... which means lightening and darkening at either side of the edge, so you could achieve something similar to your DFS example simply by applying a lower USM sharpening (or better still, by using Smart Sharpen as that allows fading the sharpening at the highlights and shadows).  Other than that we can give an effect of greater sharpness by using local contrast or by blurring the areas we don't want to be sharp, like the background.  So DFS has to work on the edges (which it clearly does) ... the question that I have is how much away from the edges does it go and what does it do there? 

It seems to me that it does go quite a long way, presumably attempting to add local contrast (which is great ... I have no issue with that!).  The only issue I have is that it seems to me that DFS really needs to be throttled back quite a lot, otherwise it does appear (to me) that it causes, as I put it, a sort of posterisation.  Of course that is easily avoided by keeping the radius and amount low.

Robert

[mchaney]

I had to take another look at the DFS code to see what it was doing in low contrast areas.  What I found was that it wasn't taking the radius into account when measuring local contrast so the piece of code that decides what is statistically significant (to sharpen) needed that factor.  About a half dozen lines of code later and we have the improvement in 2014.250.  See attached for the comparison.

Thanks for bringing this to my attention in a way that I could easily see/test it.  The problem doesn't easily show up on "organic" images: more on mathematical gradients.  But the result (in 2014.250) produces a much smoother result with less noise in areas like sky and evenly shaded areas, without affecting sharpness of notable edges.

Regards,
Mike

[Bart_van_der_Wolf]

I had to take another look at the DFS code to see what it was doing in low contrast areas.  What I found was that it wasn't taking the radius into account when measuring local contrast so the piece of code that decides what is statistically significant (to sharpen) needed that factor.  About a half dozen lines of code later and we have the improvement in 2014.250.  See attached for the comparison.

Thanks for bringing this to my attention in a way that I could easily see/test it.  The problem doesn't easily show up on "organic" images: more on mathematical gradients.  But the result (in 2014.250) produces a much smoother result with less noise in areas like sky and evenly shaded areas, without affecting sharpness of notable edges.

Thanks for having another look Mike,

It does improve the natural look of more organic image content enough to make a difference. Users might even want to review their amount settings for the Smart output sharpening. Since shadows are usually lower contrast areas, as are sky gradients, they are now both smoother, and can tolerate higher amount settings.

As an illustration I've attached crops of a 300% upsampled print-to-file example that was posted earlier in this thread (I've updated the temporary link to a 2014.250 version upsample in that post). First attachment is the pre-2014.250 version, then the new 2014.250 version of the same source. The smoother sky and less 'posterized', more organic looking low-contrast shadows are clearly visible at this 100% zoom version. Of course at the final print size the difference is more subtle due to the smaller pixel size / higher PPI of the printed output versus the display.

Cheers,
Bart

[Robert Ardill]

I had to take another look at the DFS code to see what it was doing in low contrast areas.  What I found was that it wasn't taking the radius into account when measuring local contrast so the piece of code that decides what is statistically significant (to sharpen) needed that factor.  About a half dozen lines of code later and we have the improvement in 2014.250.  See attached for the comparison.

Thanks for bringing this to my attention in a way that I could easily see/test it.  The problem doesn't easily show up on "organic" images: more on mathematical gradients.  But the result (in 2014.250) produces a much smoother result with less noise in areas like sky and evenly shaded areas, without affecting sharpness of notable edges.

Regards,
Mike

Mike,

I have to say that I am really impressed at your customer service!  Small is beautiful when it comes to software development, and many hands do not always make light work (I speak from bitter experience).

The difference this has made is huge.  The circle/gradient test is now fine; and this test below is even more interesting:



The orginal is just a square to which I applied a small amount of Gaussian blur to soften the edges.  As you can see, QU with DFS 3/100 has done a remarkable job – it has almost entirely restored the original, removing the Gaussian blur without introducing any haloes. Amazing!  As a comparison, Smart Sharpen, also with 3/100 but without any backing off on Lightness and Shadows shows the expected lightening and darkening around the edges.  On this test DFS blows Smart Sharpen out of the water.

I redid the test with the crop of the photograph. I upsized by 3x in Photoshop (to keep things equal), sharpened in QU with DFS 3/200 (to keep the test the same as before) and sharpened in Photoshop with Smart Sharpen at 3/200 (but with back-off on the Highlights and Shadows … not entirely fair, but this is a feature of Smart Sharpen). 

First of all, the ‘posterization’ type artifacts are gone from the QU image – great!  Secondly, in terms of sharpening, I see very little difference between the two: considering the upsizing, both images are very sharp (this was a hand-held shot, so I wouldn’t expect incredible detail).

BUT … on closer examination, Smart Sharpen sharpens areas that should not be sharpened (the water in the bottom corner of the image is a good example), resulting in sharpening artefacts.  DFS is perfect in those areas.  It is possible to partly correct the Smart Sharpen problem by dialling in Reduce Noise … but this is getting complicated and the effect is not perfect.

You can see the images here: http://www.irelandupclose.com/customer/LL/T47221---QU250vSS.jpg

I have to say that I’m really impressed.  You should think of offering DFS as a plug-in for Photoshop … I think you would get a lot of buyers!

Robert

[Denis de Gannes]

Good work Robert in raising this issue with Mike. A great update for all Q users, some more magic from Mike.

[mchaney]

Robert,

Thanks again for raising the issue and retesting 2014.250!  After 15 years doing just the digital imaging part of this, I consider myself a "supreme pixel peeper" which means artifacts are my enemy and when I see something that can obviously be improved, I can't sleep until it is done.    I try to never forget that image quality is a complex dynamic and I can't possibly see or test every aspect on my own.  I continue to be impressed by others who dig deep like I do and can eloquently illustrate and recommend improvements.  I appreciate the feedback because being a small company, being able to tap into the "hive mind" that electronically connects us is the only way I can compete with the big boys.

Regards,
Mike

[Robert Ardill]

Robert,

Thanks again for raising the issue and retesting 2014.250!  After 15 years doing just the digital imaging part of this, I consider myself a "supreme pixel peeper" which means artifacts are my enemy and when I see something that can obviously be improved, I can't sleep until it is done.    I try to never forget that image quality is a complex dynamic and I can't possibly see or test every aspect on my own.  I continue to be impressed by others who dig deep like I do and can eloquently illustrate and recommend improvements.  I appreciate the feedback because being a small company, being able to tap into the "hive mind" that electronically connects us is the only way I can compete with the big boys.

Regards,
Mike

Well Mike, you're certainly very welcome and my thanks to you too!  I have to say that my testing wasn't entirely altruistic - I really would like to use QImage, and DFS was (but no longer is, I'm glad to say) a show-stopper.

Rather than doing even more testing, could you give me a quick rundown of what happens during resizing? 

As far as I can see, I have two options: resize with 0 sharpening and then sharpen with DFS in the image editor, or resize with sharpening and then forget about sharpening in the image editor (except perhaps for some creative-type sharpening).  It would clearly make more sense to sharpen with resize so I would appreciate a bit more information on how that works.

What I would like to know is this: assuming I have selected DFS as the resampling sharpening algorithm, how do you determine what a sharpening level of 1, 3,  10, or whatever corresponds to in terms of DFS radius and strength?  I assume you base this on various factors like the print size, the image resolution and the paper type (matte, gloss etc)? If so, it would be very useful to know what formula you use as it would make the initial dialing in of the correct sharpening amount much easier.

Many thanks!

Robert

[Bart_van_der_Wolf]

What I would like to know is this: assuming I have selected DFS as the resampling sharpening algorithm, how do you determine what a sharpening level of 1, 3,  10, or whatever corresponds to in terms of DFS radius and strength?  I assume you base this on various factors like the print size, the image resolution and the paper type (matte, gloss etc)? If so, it would be very useful to know what formula you use as it would make the initial dialing in of the correct sharpening amount much easier.

Hi Robert,

Maybe the Mike's Tech corner forum is a more appropriate place for such specific questions, but I'll leave that up to Mike.

My take on it is that Qimage tries to take as much hassle out of the print process as possible, yet do so with very high quality. That's where the very good resampling algorithms and Smart output sharpening come in, and utilizing the maximum resolution that the printer has to offer. That means that the Smart output sharpening tries to retain the original/input image's sharpness/look, by modifying the 'Smart' sharpening radius with output size, both up and down (although down-sampling may require additional user adjustable anti-aliasing filtering to prevent trouble with repetitive patterns).

The sharpening amount to apply will depend on several factors, such as image content, output medium, and user preference. Default amount is 5, and is usually a good point to start, but 10 or more may be usable in certain situations as mentioned before, or when the input image was not or hardly sharpened to begin with. Whatever the settings you like best turn out to be, they can be saved in a recipe that can be recalled for future use, otherwise Qimage will simply remember most of the last used settings for the type of paper and printer and profile used.

Cheers,
Bart

[Bart_van_der_Wolf]

As far as I can see, I have two options: resize with 0 sharpening and then sharpen with DFS in the image editor, or resize with sharpening and then forget about sharpening in the image editor (except perhaps for some creative-type sharpening).  It would clearly make more sense to sharpen with resize so I would appreciate a bit more information on how that works.

Hi Robert,

The latter is the normal MO. Automatic Smart sharpening after resizing is one of the things that makes Qimage so powerful. It will still allow optional full control by the user to resize without sharpening and save that, and then add sharpening in its editor when subsequently printing, but that defies the set-and-forget benefits of Qimage. In fact, one can do such a convoluted operation with a global Print filter, so even that can be automated, but is more commonly used to adjust for certain media that print a bit darker than others, or correct underexposed or low contrast images from someone else, and things like that. The original file will remain unaltered, only a separate filter description will be saved, so you have total freedom to create and adjust such a procedure.

Cheers,
Bart

[Bart_van_der_Wolf]

Here are a couple of examples how Qimage handles a very organic (sinusoidal) range of spatial frequencies. The target used for upsampling to 300 percent of its original size, and automatic DFS Smart sharpening, shows how smooth Qimage transitions from no sharpening to very high sharpening of the finest detail levels (the finest printed details if one uses 600 or 720 PPI resampling and printer resolution). An original star target was given a very mild 0.7 sigma Gaussian type of blur to produce the resampling target, comparable to what a very sharp lens would produce with excellent technique (heavy tripod, and perfect focus, at the best aperture for the sensor).

The automatic DFS Smart sharpening was applied after Fusion interpolation, with amounts 5, 10, and 15, as indicated in the filenames.

This also shows the benefits of not over-sharpening the source image file, it allows to apply more sharpening during print/output sharpening, without wreaking havoc.

Cheers,
Bart

[Robert Ardill]

Bart,

Many thanks for your replies.  You're probably right and my question probably should be on Mike's tech corner - however Mike has been very helpful on this forum and what makes it interesting to keep the discussion open here for a while longer is that we are not all QImage aficionados on this forum (so that there are quiet a lot of potential customers for QImage around . 

Do you have a test image that could be used to check out the best resizing/sharpening settings for different papers?  I'm thinking of a chart to test print MTF, like the lens MTF charts, either for visual or scanned evaluation.

As you've no doubt guessed from the other post, I'm moving in the direction of less and later sharpening, which fits in nicely with the QImage modus operandi.

Robert

[Bart_van_der_Wolf]

Do you have a test image that could be used to check out the best resizing/sharpening settings for different papers?  I'm thinking of a chart to test print MTF, like the lens MTF charts, either for visual or scanned evaluation.

I have designed a test image/target that allows to do a lot of those things, like measuring the central blur diameter of a 'star' to determine limiting resolution and give a quick visual comparison, but also a slanted edge to determine the MTF (or the SFR as the ISO standards committee calls it), and a grayscale to calibrate the response curve. There are a few other line patterns on that target that directly challenge the printer, as a check to verify the printed quality of the target. It's pretty merciless for printers but also allows to test cameras by shooting an image of it.

Another cruel test is a so-called zone-plate target, but that's most often used as a down-sampling test target, although some upsampling algorithms also make a mess of upsampling that.

As you've no doubt guessed from the other post, I'm moving in the direction of less and later sharpening, which fits in nicely with the QImage modus operandi.

I agree, although if our tools would allow a better (deconvolution) Capture sharpening, without halos and without going crazy on noise, there would be little harm in doing that from the start. Better signal to noise of the image data, allows more detailed post-processing. Given the current state of affairs, there is a lot to be said for postponing it until we have more pixels that allow 'sub-pixel' (versus original pixels) accurate sharpening, when upsampling for e.g. print is our destination. I see little benefit in upsampling of artifacts and making them easier to see, so postponing sharpening can help with that.

Cheers,
Bart

[Robert Ardill]

Thanks Bart,

I'll have a go with your test images ... after I've caught up on some work for customers.

Yes, of course, if we could correct things like blurring due to the optics, for example, without introducing additional artifacts then that would be absolutely great ... and the sooner in the process the better (in the same way that using a better lens is a good thing!).  But to introduce an artifact, and then enlarge it, only to add more artifacts ... seems a step or two too many.

Do programs like DXO Optics use deconvolution techniques, do you know?  What's the state of play?  Seems like something to keep a close eye on!

Just had a look at your link ... have you used Imatest (Norman Koren)?  It's seriously good for testing lenses (and adjusting microfocus etc) ... but it's one of these programs that sort of sucks you in like a black hole (one needs to tread carefully!).

Robert

[mchaney]

Hi guys,

I see sharpening as a two stage process.  For the majority of work which tends to start with a photo from a digital camera, the demosaicing of the sensor data combined with any AA filter that might be used over the sensor tends to soften the final image a bit so that you don't get single-pixel-level sharpness.  So the first step is to get the image itself to have the intended sharpness.  In Qimage Ultimate, you can set your sharpening level in raw options so that when the image is developed, enough sharpening is added to get the image back to "baseline" sharpness: bringing the sharpness of the image to the intended level in the image itself.  That is typically done as mentioned, in the raw format options, or in the image editor if you are not shooting raw (using DFS to a level you are satisfied with).  It's easier if you are shooting raw because you can dial in the sharpening (and adaptive noise reduction) levels needed for your particular camera and then leave it alone.  Some cameras render sharper raw images than others, some have AA filters removed, and so on, so the level you set there is based on a particular camera model and can be customized globally or by specific model if you have more than one camera.

Once the image itself looks as sharp as it "should" to your eye (I typically judge that at 1:1 zoom), the final phase is to match the sharpness you see on the screen with what comes out of the printer.  The final print sharpening in QU is designed to allow you to do that.  For your printer, driver settings, and paper, you can print a few samples and as soon as you set your final (smart) sharpening level to something that matches well with your screen, you can save that as a printer setup and never worry about it again: because now your visual sharpness of the print matches what you see on screen.  QU is designed to give you the same level of visual sharpness on each print, regardless of what size you print.

Yes, it does use an algorithm that takes into account the original image size, how much it has to be upsampled (or downsampled) for the current print size you are using, modified of course by where you have your final print sharpening set (5 being the default).  I don't disclose the actual algorithm for (probably obvious) reasons but it is based on years of R&D of what it takes to replicate visible sharpness at various input and output PPI.

I can tell you that at output/print time, the sequence is:

- Interpolate to final print resolution
- Sharpen at that final resolution
- Apply color management

Note that the application of color management must come last because the first two steps (interpolation and sharpening) should never be done in the printer's color space (profile) as those printer profiles are often jaggy and have out-of-gamut colors which could introduce artifacts if processing is done in a printer color space.

Regards,
Mike

[Bart_van_der_Wolf]

Do programs like DXO Optics use deconvolution techniques, do you know?

Yes, they do, but a drawback is that their program is limited to an AdobeRGB gamut workingspace, even when exporting as ProPhoto RGB.

What's the state of play?  Seems like something to keep a close eye on!

Lot's of things going on, but the industry giants are (as usual) slow at picking up the pace and implement real solutions for e.g. Photographers.

Just had a look at your link ... have you used Imatest (Norman Koren)?  It's seriously good for testing lenses (and adjusting microfocus etc) ... but it's one of these programs that sort of sucks you in like a black hole (one needs to tread carefully!).

Yes, I've used it and it is very good, but my main issue with it is that is only analyses, without making the connection to implementing solutions based on those analysis. I do understand Norman's business model, so that explains for me why they stop at the point they do.

Cheers,
Bart

[Ernst Dinkla]

A deconvolution sharpening choice added tot Qimage's RAW development? Based on the slanted edge tool Bart provides so anyone with an odd camera/lens combination can still use it while more common combinations can be found in a user's pool of sharpening profiles.

Ernst Dinkla

[Robert Ardill]

A deconvolution sharpening choice added tot Qimage's RAW development? Based on the slanted edge tool Bart provides so anyone with an odd camera/lens combination can still use it while more common combinations can be found in a user's pool of sharpening profiles.

Ernst Dinkla

That would be a good reason to do raw development in QImage! (if I understand you correctly).  But why stop at the raw development? Surely the deconvolution would work just as well on a tif? (providing no sharpening during the raw development, that is).  Still, easier said than done I would guess.

Robert

[Robert Ardill]

Just to give you a bit more feedback, I've just done 3 prints of the same image for three different customers (30x20, 20x16 and 16x12), and with slight trepidation did this with QImage (on a lovely Canson Platine paper).  All three images had minimal capture sharpening in Lightroom (low radius, low strength, masking) and I then printed with QImage with the DFS sharpening set to 5 (and held my breath!).

The results were perfect - sharp, no haloes, no artifacts.  I couldn't have been more pleased.

Robert