Sharpening in ACR 7.1

[grampus45]

I'm reading the part of Martin Evening's "Adobe Photoshop CS6 for Photographers" dealing with sharpening in ACR.  As in his previous edition for CS5, he seems to be saying that putting the Detail slider at 100 is akin to USM in PS, while placing it at 0 minimizes halos.  I would gather this means the 0 setting is deconvolution sharpening. 

As I recall, Eric Chan indicated that the settings were the other way round: 0 was akin to USM and 100 deconvolution. Which is which?

thanks,

grampus45

[grampus45]

I thank everyone for their various comments.  The one that indicated that Detail = zero was USM-like seems to be old news that Jeff's earlier comment negates.  He now contends that USM is not part of the ACR sharpening model.

It is because of the complexity of the interaction of the various sliders that knowing just what they are doing would be of immense help.  For the most part, each image needs its own sharpening, and just knowing a set of presets is of little use.  One needs to know what's happening under the hood so that one can properly understand what he/she is doing with each image.

I am also confused by the various statements I have read here and elsewhere about the use of the ACR sharpening for doing capture sharpening.  These statements seem to indicate one should achieve a good degree of sharpening, but not too much, particularly if one plans on doing further output sharpening later on, say, in PS.

Rather generally, capture sharpening should be used to counter the various elements of the capture (lens blur, diffusion, AA filtering, and demosaicing) that can cause the image to lose some of its inherent sharpness.  These effects are, in total, reasonably subtle, and thus so should be the capture sharpening -- indeed noticeable, but barely so.  But the various comments I am reading indicate using capture sharpening that is far from subtle, and, indeed, partway toward what one might consider output sharpening -- although not necessarily at output size/resolution.

The effects of capture error are also far less significant in larger areas of solid color, where none of the issues mentioned above comes into play in a meaningful manner.  This would suggest that a significant use of masking would be appropriate in applying capture sharpening.  But many things I read, and many suggested 'presets,' tend to downplay or even ignore masking.  Once again, these settings would appear to be more appropriate to output sharpening.

So it would be very nice if someone could provide meaningful information about what's really happening as you move the Detail slider from zero to 100.  I know very well what each slider is supposed to be doing (I can read the books as well as anyone -- and have), but none of us really knows what's actually going on with that Detail slider.

We know from Bruce Fraser what the general principles of sharpening are.  What we need now is someone like Bruce to write a treatise on how to enact those principles, such as they are possible, with the sharpening facility in ACR. We can attempt to infer what's happening by moving the sliders themselves, but it's really going to take someone with "inside knowledge" to do the job properly. So far we only have vague hand waving.

[areohbee]

grampus45 wrote:

The one that indicated that Detail = zero was "USM-like" seems to be old news that Jeff's earlier comment negates.  He now contends that USM is not part of the ACR sharpening model.

Not at all negated. Whatever the algorithm is (and I have *no* idea), it's definitely more like a traditional USM at detail=0, than at detail=100. I said "USM-like" with the quotes as a way to distinguish it from "deconvolution-like", which is at the other end of the spectrum. I really don't think finding the proper "name" for the algorithm, or even reading about how it is "enacted" will help you to determine the settings you like best.

I am a person who craves, and thrives upon understanding what's going on under the hood, but when it comes to sharpening, I think it's entirely academic.

I think Noel nailed this one when he said "play with them until your feel develops" (paraphrased).

I don't think this necessarily applies to all things, but sharpening in ACR: yes.

I think it's best to make your photos look "nicely sharpened" in ACR, but *not* oversharpened. In other words, I don't think it's best practice to stop shy of looking sharp enough in ACR, but I definitely recommend using output sharpening based on output medium, so you are less inclined to oversharpen in ACR (but the proof is in the pudding). Obviously if you are also going to be sharpening using some other software too, or instead, then what I just said does not apply.

I know *exactly* what's going on with that detail slider as it's moved from 0 to 100, from experience, although I'm not qualified to give technical details...

Good luck,

Rob

[Jeff Schewe]

grampus45 wrote:

This would suggest that a significant use of masking would be appropriate in applying capture sharpening.  But many things I read, and many suggested 'presets,' tend to downplay or even ignore masking.

I would argue against the term "significant" and more likely call it a beneficial use of masking. You want to be sharpening edges and generally don't want to sharpen surfaces (which are broad areas of tone/color). The default for masking is zero because, well, Thomas decided that no masking should be applied at default but almost any image will benefit from having some edge masking used.

How much really depends on the edge frequency of the image…if you are shooting portraits you want a fairly high edge masking (40 or above). If you are shooting low ISO landscape images with a high frequency of image texture, you probably want less; 10-25 or so. But…that depends on the amount, radius, detail and noise reduction settings…you really can't give a range of settings for any single parameter because they all depend on each other...

grampus45 wrote:

We know from Bruce Fraser what the general principles of sharpening are.  What we need now is someone like Bruce to write a treatise on how to enact those principles, such as they are possible, with the sharpening facility in ACR. We can attempt to infer what's happening by moving the sliders themselves, but it's really going to take someone with "inside knowledge" to do the job properly. So far we only have vague hand waving.

I did...when I revised Bruce's Real World Image Sharpening book...which came out just before the PV 2010 noise reduction functionality (unfortunately). The aim is to get the image to look good at 1:1. That is the intent and design of the Detail panel in ACR/LR. And yes, it's tough because there are tons of cameras and tons of image types to deal with...yes, it's tough to evaluate just enough sharpening but not too much. You can't accomplish that with a few presets...you need to educate your eyes...no way around that.

And the other 2 phases of image sharpening, creative and output sharpening are further complicating factors. The ACR/LR creative sharpening is primitive but useful. The output sharpening in LR is actually very, very good. It's less good in ACR because of the limitations of the size functionality in ACR. Output sharpening MUST be done at the final output size...and sizing in ACR other than native are problematic.

The bottom line is to make the image look good at 1:1. Don't try to do over/under sharpening in the Detail panel, don't try to sharpen for effect or do creative sharpening and don't worry at all about output sharpening. If you are shooting low ISO on high rez cameras, you can sharpen more aggressively, the smaller the capture size and higher the ISO the more you need to be careful of setting the sharpening and noise reduction correctly.

The only generalizations I can make os that you do want to adjust all of the following; Amount, Radius, Detail, Edge Masking and Luminance Noise Reduction to get an optimal capture sharpening result. The numbers will vary by camera size, lens type, Exposure, ISO and shooting techniques. YMWV...

[grampus45]

Thank you, both Jeff and Rob, for replies with some substance.  It appears that the real answer to my question is not possible, at least at this point in time.  Jeff, here's hoping you are able to provide an update to Bruce's and your book that takes the new ACR sharpening facilities into account. You've both given me some good things to think about as I continue my (now several years long) effort to figure out just what's happening as those sliders move. 

[areohbee]

You're welcome Grampus.

I know my replies did not answer your true question - I can't answer it.

But maybe with some practice you'll find your way to optimal sharpening, even without the understanding you seek.

Or, maybe a satisfying treatise will come along...

Until then,

Rob

[Noel Carboni]

grampus45 wrote:

Thank you, both Jeff and Rob, for replies with some substance. 

I'm not sure whether you meant that as a mild put-down or what, but I find it humorous that Rob has said exactly what I said - basically you have to fool with the settings until you like what you see. 

-Noel

[grampus45]

Noel Carboni wrote:

I'm not sure whether you meant that as a mild put-down or what, but I find it humorous that Rob has said exactly what I said - basically you have to fool with the settings until you like what you see. 

-Noel

No, Noel, no mild put-down, just straight goods.  My question is apparently, at the moment, unanswered and unanswerable, but your and Rob's comments added to the discussion in a substantive manner.  So I thank you both.  Now let's hope for all of us that more information from the top level is coming forth. I was hoping MadManChan would chime in here, but he seems to be lying low.

take care,

grampus45

[MadManChan2000]

Hi grampus45 and folks,

Yes, moving Detail towards 100 will use a deconvolution-based method to sharpen the image.  I recommend this only for clean (e.g., high ISO), finely detailed scenes.  It will perform the most detail and texture extraction that it can, with no effort to suppress noise or other anomalies.  Use on portraits at your own peril!   

(The earlier high-detail sharpening method in PV 2003 in Lr 1 and Lr 2 was indeed based on USM.  But that was then and this is now.  )

[Jeff Schewe]

MadManChan2000 wrote:

Yes, moving Detail towards 100 will use a deconvolution-based method to sharpen the image.  I recommend this only for clean (e.g., high ISO), finely detailed scenes.

I'm pretty sure Eric meant to say (e.g., low ISO)

:~)

[Andrew_Hart]

MadManChan2000 wrote:

Yes, moving Detail towards 100 will use a deconvolution-based method to sharpen the image.

Could someone please explain to me what this "deconvolution-based method" means, and how it differs from other, and what other, methods.

I've been trying to follow this thread and haven't got a clue what you are all talking about,sorry.

[Jeff Schewe]

Read: Deconvolution on Wikipedia and in particular Richardson-Lucy deconvolution

In basic terms it uses a point spread function (PSF) to "un-blur" a blurred image. ACR's deconvolution processing is very similar to Photoshop's Smart Sharpen when set to the Lens Blur option. It is also the basis for Adobe's tech demo on Image Deblurring that was shown late last year that can be seen on Adobe TV. Note, that demo caused a bit of a controversy because the demo showed a deblurring based on a know synthetic blur that was easier to deblur.

[Andrew_Hart]

Thanks, Jeff.

Still not completely clear but I'll do your suggested further reading/viewing and see if it then makes more sense.

[grampus45]

Andrew: Let me try a basic description of deconvolution sharpening.  I'm taking rather significant liberties here in order to keep it reasonably simple.

Most of the light coming from a given point in the scene (the source) you are shooting ends up in the pixel corresponding to that point in the image being formed on the camera's sensor, which is where that light belongs.  However, because of lens imperfections, refractions in the anti-aliasing filter, and other issues, some of that light gets dispersed to neighboring pixels, where it doesn't belong, thus creating blur.

But, if one has a good model of how that spread occurs, one can attempt to estimate from the light values (actually electronic charges) that actually occur in the sensor's pixels what part of them were inappropriately diverted from the neighboring pixels -- and then put those parts back where they "belong," thus unblurring or sharpening.

The fraction of the light coming from a given point of the scene that is inappropriately diverted to a neighboring point is known as the point-spread value, and the set of all these together comprises the point-spread function. 

Solving the statistical problem of inferring from what is actually observed on the sensor's pixels to what should have been there if there were no inappropriate diversions is the deconvolution method of sharpening, and it is rather generally a maximum-likelihood problem.  These problems are typically solved iteratively until a consistent set of "proper" values is obtained, and this solution is typically computationally very costly.  Thus, most of the methods used in practice do not carry the solution to a full convergence, but only use the first one or two iterations.  I believe this is the case for PS and ACR/LR.

[areohbee]

Great description Grampus, and if I may add:

* It's what FocusMagic uses too.

* It's used in Astronomy to get clearer pictures...

* GoogleEarth?

Other sharpening methods are more about separating and/or exagerating micro-tonal variations (the art of illusion...)..., where-as deconvolution is about constructing a more accurate image from a less accurate (convoluted) collection of data.

R

[grampus45]

Rob Cole wrote:

Other sharpening methods are more about separating and/or exagerating micro-tonal variations (the art of illusion...)..., where-as deconvolution is about constructing a more accurate image from a less accurate (convoluted) collection of data.

Yes, Rob.  Let me add a summary to my above description:

Deconvolution Summary:

What happens at each pixel on a camera's sensor when you snap the shot is a convolution (a rolling together) of light that belongs on that pixel and bits from neighboring pixels that don't actually belong there.  Deconvolution sharpening is an attempt to unroll (deconvolve) those elements in each pixel that don't belong there and put them back into the pixels to which they properly belong.

In short, deconvolution sharpening is an attempt, under reasonable assumptions, to infer from what's actually there what ought to be there.

grampus45

[Noel Carboni]

In practice it makes slightly blurry things look slightly less blurry.  That's about as good as it gets, except on cop shows on TV. 

-Noel

[Noel Carboni]

FYI, one deconvolution tool I regularly use on astroimages runs 10 minutes on a powerful workstation on a single 6 megapixel astroimage.  It doesn't seem to be multithreaded, but even one of my cores is no slouch.

-Noel

[Andrew_Hart]

Thank you very much grampus45.

Your explanation of deconvolution sharpening is beautifully put and it is much more helpful and concise than anything I could find on Wikepedia.

With this new understanding of what is supposed to be happening with deconvolution sharpening, I am utterly amazed, and just a little incredulous, that Adobe is able to say that they can achieve this with the simple positioning of the Detail slider. 

[Jeff Schewe]

Andrew_Hart wrote:

With this new understanding of what is supposed to be happening with deconvolution sharpening, I am utterly amazed, and just a little incredulous, that Adobe is able to say that they can achieve this with the simple positioning of the Detail slider. 

Yeah, the elves at Adobe (in particular the Camera Raw team) do a pretty darn good job of producing magical results...par for the course (the boys are brilliant).

The Detail slider is pretty special...course, so is the Masking (and all of the Noise Reduction sliders).

[grampus45]

Andrew_Hart wrote:

Thank you very much grampus45.

Your explanation of deconvolution sharpening is beautifully put and it is much more helpful and concise than anything I could find on Wikepedia.

You're most welcome Andrew.  It's alway fun for me to try to make a messy subject unmessy -- and it's always possible.

With this new understanding of what is supposed to be happening with deconvolution sharpening, I am utterly amazed, and just a little incredulous, that Adobe is able to say that they can achieve this with the simple positioning of the Detail slider. 

I quite agree, and I second Jeff's comments below: the ACR/LR teams are super and have provided us with excellent tools.  I thank them with all my heart.

I would, however, still love a good answers to my original questions (there are two of them) about that Detail slider.  What shaprening method is being employed when the Detail slider is at 0?  and in what sense is there maximal halo suppression at that setting?  MadManChan (Eric, ma'love): are you there?

[Isseman]

It´s quite simple, at 100 the detail slider is deconvolution-based and at 0 it´s convolution-based. I.e.it´s the same algorithm run backwards, hence the halo suppression

[Noel Carboni]

wurre wrote:

It´s quite simple, at 100 the detail slider is deconvolution-based and at 0 it´s convolution-based. I.e.it´s the same algorithm run backwards, hence the halo suppression

You're a politician by day, right? 

-Noel

[Jeff Schewe]

wurre wrote:

It´s quite simple, at 100 the detail slider is deconvolution-based and at 0 it´s convolution-based. I.e.it´s the same algorithm run backwards, hence the halo suppression

Actually, it's not...it's far more complicated than that and the way the slider works is to interpolate between the two different algorithms. Halos suppression is not the opposite of deconvolution sharpening.

[grampus45]

MadManChan2000 wrote:

Hi grampus45 and folks,

Yes, moving Detail towards 100 will use a deconvolution-based method to sharpen the image.  I recommend this only for clean (e.g., high ISO), finely detailed scenes.  It will perform the most detail and texture extraction that it can, with no effort to suppress noise or other anomalies.  Use on portraits at your own peril!   

(The earlier high-detail sharpening method in PV 2003 in Lr 1 and Lr 2 was indeed based on USM.  But that was then and this is now.  )

Hi Eric; thanks for tuning in.  But my original question was: what sharpening method is being used as detail goes to 0?  And just what does it mean to say that it has maximum halo suppression? 

thanks,

grampus45

[areohbee]

Regarding masking:

----------------------------

I know one person who sets default masking at 30, and another who sets it at 70. Optimal setting definitely depends on the image (I use non-zero masking values only occasionally).

I like the priniciple of masking, but not the reality: Unless masking gets 100% in areas of say sky, I think it does more harm than good: there is a dancing (call it: alternation) of unmasked and masked regions that I find more offensive than the noise itself. And in areas where edges meet smooth, sharpening extends just enough beyond the edges into the smooth to produce a twinkling around things that drives me crazy. Some people say they have no idea what I'm talking about, others say "Amen - brother".

Personally, if an image would need masking, I prefer a combination of desharpening (often at exactly -50: details available upon request) and/or now noise reduction in the regions where the noise is offensive, and/or enhanced sharpening in areas of interest (using locals, usually brush).

I'm not disagreeing with Jeff about masking - he's expressed his opinion and his taste - can't be disagreed with... Just saying: I really think each individual needs to discover for themselves how to sharpen things according to their taste.

More regarding detail:

-------------------------------

I know some people who always keep detail at zero, never ever above zero. Why? they find it's effect "distasteful". I know some people who often crank it way up. Why? - because they like it's effect. I tend to keep it on the low side usually - that's my taste: I usually like a little (ISO permitting), but not too much... - no right answer me-thinks...

And radius

---------------

To make a long story short, some people tend to prefer smaller ones, others bigger. But also it needs to be balanced with other settings: People who like large amounts and high detail settings, tend to use lower radius... I usually like a larger than average radius, but I often use smaller than average amount and detail.

Cheers,

Rob