Photoshop LAB Color: Chapter 5

rutt

There are three major topics of this chapter:

1. How to deal with color noise by blurring the A and B channels
2. How to sharpen the L channel
3. The theory of why sharpening the L channel is sometimes much better than the RGB alternatives, sharpening a Luminosity blended layer or lumonsity fade after sharpening.

[size=+1]Blurring as a cure for color noise[/size]
Dan has some examples of images with a lot of noise. Converted to LAB, the noise migrates mostly to the A and B channels. By blurring them, he decreases the noise a lot without reducing sharpness. Using gaussian blur with a radius of 3 gives acceptable results for this, but CS2 has Surface Blur and this gives even better results, even at higher radii.

This is neat stuff. Dan has been teaching it at least since I took his course a couple of years ago. Unfortunately, I've never actually been able to make it work with my own shots. Modern Canon cameras do produce noise at high ISOs, but not the kind of colored noise which can be blurred away in the LAB A and B channels. I don't know about other manufacturers and would like to hear.

Here is a test image which Andy posted a few weeks ago. (It's a very close crop of ISO 3200 images.)



Converted to LAB, here are the L, A, and B channels, respectively:



Well, there is noise, certainly, but not the same kind that Dan shows how to fix.

I've emailed this point to Dan's ColorTheory yahoo group a couple of times. Dan has never actually answered (though I intend corner him when I take his advanced class in San Diego next month.) I did get some good info last time, though. Basically:

1. Wow, look at that great low noise 3200 ISO image!
2. Jeez, Canon has gotten really really good at reducing the color noise in the camera.
3. What noise there is looks nice, doesn't it?
4. Stop whining and enjoy the great film-like texture. Andy also said almost exactly the same thing.
5. If it really bothers you, you need to use some big gun like Noise Ninja or Neat Image.

On the other hand, I did find plenty of colored noise when working on the Cecil Stoner pictures. These were scanned from old photographs, mostly pretty badly damaged. But even for these images, colored noise that could be addressed by A+B blurring not the main issue. I found that the Dust and Scratch filter applied to the L channel was most effective, but needed to be applied locally and with great care since it definitly reduces sharpness. This filter worked, well, because the images had dust and scratches after all. They also had noise, but again, Neat Image worked much better than blurring the A+B channels. And again, it needed a very light hand because by the time I turned it up enough that it actually worked, it did bad stuff to faces and other critical parts of the images. Great for dark backgrounds, though.

I suspect the colored noise thing really happens, but I haven't encountered it. I'd like to hear from people how have.

[size=+1]Sharpening[/size]
I've already summarized a lot of what Dan teaches about L channel sharpening in these two threads:
[list=*]
[*]Basic L channel sharpening
[*]Separate control over light and dark halos
[/list]
I highly reccomend these posts to anyone who isn't really sure about how USM really works and how to use it. Even better is to read the chapter, "Sharpening with a Stiletto" in Professional Photoshop. Dan does explain the fundemental justification for sharpening a lot better in the LAB book than in PP, though. I'll quote him since he says it so well (and has given me permission to quote him at length.) Discussing an image of an American flag, he says:

Dan Margulis wrote:

If this were a perfect world, the transition between the white and red stripes would always be razor-sharp. If this were a vector graphic created in Adobe Illustrator, and the stripes were absolutely horizontal or absolutely vertical, we'd have a chance. But here, they're diagonal and curved, most awkward for people who work with rectangular pixels. To make a diagonal line, we have to arrange those pixels in a stairstep pattern. Magnify the file enough, and the stairstepping becomes painfully obvious. But even at normal sizing, the stairstep pixels must be neither red not white,as otherwise the effect would be too noticeable. Instead, these edge-defining pixels are pink, a compromise known as anti-aliasing.

USM hides the stairstep pattern with the halos it draws around edge transitions. So the take away from this is that sharpening is required not because our images are out of focus, but because of the fundementals of digital imaging. Your image will actually benefit more from sharpening if it's in better focus.

[size=+1]Why Sharpening in LAB is Better than Sharpening in RGB in Luminosity Mode[/size]
This is the nerdly part of the chapter. If you are only interested in technique you can use, just sharpen on the L channel in LAB instead of trying to sharpen in RGB and forget about it. The results will be at least as good in all cases and quite a bit better in some cases. (Sharpening in CMYK is another story, but not our subject here.) That's all you need to know.

For the nerds:The reason L channel sharpening works better than luminosity mode RGB sharpening is that in some cases it's light halos can be slightly colored instead of white and ditto for the dark halos not always being black. This is because LAB can represent those impossible colors such as yellow which is also as light or dark as possible. When the image is displayed or converted to RGB, these colors split the difference and become very light but slightly colored. The result is a lot more subtle and realistic looking than the pure white halos that result for RGB luminosity USM.

rutt

HIRALOAM Sharpening, an important alternative
OK, pretty boring chapter summary. The noise reduction technique isn't applicable to your camera; you already know the sharpening techinque which is explained better elsewhere, and the theory part is either just for nerds or not you were already bought into sharpening in L anyway; so what's the big deal?

There is a nugget contained in this chapter that I didn't cover, though. It's a new technique to me, though I've seen references on dgrin to a similar idea. It's just a hint in a sidebar in this chapter, but I've been using it and loving it.

Dan calls it HIgh RAdius LOw AMount (HIRALOAM) sharpening. In contrast to classic USM sharpening, it works well even when there is a lot of noise (as in high ISO shots) and even for shots that are very out of focus. Here is an example of a shot with both issues. Before:



After:



Look what this has done for the policeman's face, the asphalt in back, and even the totally OOF foreground people. To get this I use USM with the Amount-44, Radius=27, Threshold=4. Very different parameters than you get if you follow my sharpening tutorials or read Professional Photoshop. Dan says that "conventional USM emphasizes edges but HIRALAM sharpening emphasizes gives shape."

I'm still learning to use USM, but here is what Dan says to do and what I do:

1. LAB mode, L channel only active
2. Amount=500
3. Threshold=0
4. Now play with radius from about 10 to about 40. It's going to look terrible at this point but what you are looking for is a value that empasizes shape. For portraits, Dan says look for a value that "emphasizes eye and cheekbone structure. Too high and high and the entire face will lighten. Too low and it will just look silly."
5. Now turn the Amount down to 50. Experiment with values between 40 and 50. It should look good at this point.
6. Turn up the threshold a bit, say to 10. Did you lose something? Turn it back down until you regain it.

Try with a few portraits at first. You can use both conventional USM first and then HIRALOAM after. Sometimes this will both sharpen the details like eyeglasses and eyes themselves and also give lots of shape.

dandill

rutt wrote:

Dan calls it HIgh RAdius LOw AMount (HIRALOAM) sharpening.

See also Thomas Knoll's Local Contrast Enhancement scheme, described at http://luminous-landscape.com/tutorials/contrast-enhancement.shtml

dandill

rutt wrote:

Dan has never actually answered (though I intend corner him when I take his advanced class in San Diego next month.)

If you or others here have not already commented elsewhere, I am very interested to learn about the experience of being in his course(s).



Thanks
Dan

rutt

dandill wrote:

If you or others here have not already commented elsewhere, I am very interested to learn about the experience of being in his course(s).



Thanks
Dan

Dan's courses are very intense 3 day courses. At least 12 hours/day. It totally changed the way I looked at things. Made some friends. It took me months afterwards before I stopped absorbing what I learned. I'm taking my second one next month.

rutt

Another HIRALOAM example. This time I conventional USM'ed and the rest first, so HIRALOAM was the last thing. Right before:



After:



I used these parameters:



The difference is very dramatic, even in the OOF faces in front. If you have trouble seeing that by scrolling back and forth between the two images, try and set it up so you can "A/B" the images in place easily. If your browser supports tabbed browsing (firefox, for example), open each image location in a separate tab and switch back and forth between the tabs. Or download the images and put them into different layers in PS. You'll see.

I'm getting a handle on tuning these HIRALOAM parameters.

1. As Dan says, start with Amount=500 and Threshold=0.
2. Tune the radius until the features you want to emphasize are outlined. Faces should not turn all white or all black; if that happens the radius is too high. When Dan says that too low a radius just makes it look "silly", I think he means that it isn't really doing anything but following edges, no empahasis of shape.
3. Turn down Amount to 50. You may want to play with lower values if you see obvious halos at 50.
4. If the image looks too harsh at this point, for example loss of subte gradations in tone in faces, turn up the Threshold until it looks better. This kind of sharpening is very sensitive to the difference between thresholds in the 0-20 range.

Viking

Why is this topic so Dead? I checked my pictures taken with a Konica KD400 (Not 100% sure of themodel number). A great point and shot camera for like 4 years ago. Some pictures are Good and some a very, very noisy. Color noise. So I tryd Dans tip and it worked very well. Its a 100% Crop image. No color correction or other things have been applyd to it. Just a Surface Blur of the AB channel I think its a great tip to know. Even if todays more high tech SLR cameras dont produce such evil color noise.

dandill

rutt wrote:

... Now turn the radius down to 50. Experiment with values between 40 and 50. It should look good at this point.

... Now turn the *amount* down to 50?

Dan

rutt

dandill wrote:

... Now turn the *amount* down to 50?

Dan

Exactly. I'll correct the original. Thanks!

gwilson

rutt wrote:

This is neat stuff. Dan has been teaching it at least since I took his course a couple of years ago. Unfortunately, I've never actually been able to make it work with my own shots. Modern Canon cameras do produce noise at high ISOs, but not the kind of colored noise which can be blurred away in the LAB A and B channels. I don't know about other manufacturers and would like to hear...

I've emailed this point to Dan's ColorTheory yahoo group a couple of times. Dan has never actually answered (though I intend corner him when I take his advanced class in San Diego next month.)

May I ask if you ever received any information that cleared up your question about noise reduction?

Gloria

Duffy Pratt

In general, there are two kinds of noise. One variety is luminosity based, and the other is color based. I used to see the color type noise all the time, when shooting JPGs at high ISO with my 20D. This noise typically showed up in the Blue channel in RGB, and then got converted mostly to the B channel in LAB. Since I've started shooting in RAW, I've noticed that I don't do the noise blurring trick as much as I used to. I still get noise, but its more luminosity based, and if I can't get rid of it with ACR, then it's basically here til I get Noise Ninja or Neat Image.

On sharpening superiority in the L: it's superior when you have some very lightly colored edges. The sharpening in RGB can boost the white halos to where they blow out, and that's the main problem that can happen. You get nasty white lines where there should be a light color (especially with yellows).

In Photoshop Professional, Dan now recommends just doing the luminosity blend in RGB for sharpening if there is no other reason to go to LAB. But you might want to go to LAB if you know you have some bright, lightly colored edges.

Duffy

rutt

These days, I get really good results from ISO 1600 shots (essentially all the ballet work) by using a skeleton K channel as the basis for a mask for both Surface Blur and USM. The key insight is that shadows have much more noise than highlights and so we want to blur the shadows not the highlights and we want to sharpen the highlights not the shadows. The K channel is a good highlight mask.

Early on, I make a duplicate image and convert to CMYK. Then after any contrast blending, I do an L channel surface blur as Dan suggests. But often I find this blurs stuff I don't want blurred. Invert the K channel into a layer mask and the blur will be limited to the shadows.

Sharpening is the opposite. Use the K channel as is for a layer mask and the sharpening avoids the shadows.

Blur and/or curve-steepen those k-channel based layer masks to fine tune. I find especially that blurring the mask is important for the sharpening step.

Also steepening the shadow end of the L channel curve through an inverted K channel based layer mask is a fine way to turn any shadow noise into solid shadows.

I've been asked umteen times what noise reduction program I use for the ballet shots. This is the answer. For this particular application, I think I can outperform the noise reduction programs. Dan thinks judicious use of surface-blur is all the noise reduction you need to know.

Duffy Pratt

Thanks Rutt. That's a very interesting and clear approach. It sounds like its worth building an action for retrieving a K channel. There are other areas where I've started using it as a mask, to isolate shadow and/or highlight casts (as per the fiddler). The key concept that makes this work, I suppose, is that people are pretty forgiving of not seeing detail in shadow areas, so a blur there is more acceptable.

Could you get acceptable results using Gaussian Blur? I have PS CS, not CS2, so I don't have Surface Blur available.

Duffy

rutt

Duffy Pratt wrote:

Could you get acceptable results using Gaussian Blur? I have PS CS, not CS2, so I don't have Surface Blur available.

IMHO, surface blur alone is worth the upgrade cost. But with CS3 around the corner, jeez, it's a little hard.

With surface blur, you don't lose much detail. Really, if you get the parameters right it only blurs where there is no detail. But noise/detail are very hard to tell apart.

The shadow noise at high ISOs is really an essential property of light. As exposure time gets shorter and the sensor size smaller, how many photons hit the sensor and are measured becomes a statistical issue and the variance becomes statistically significant. So you really don't get good detail in the shadows in this case, at least not as good as in the highlights.

At least in the ballet shots, the main detail I lose is stuff the lighting director is really trying to lose anyway: the curtain and the floor.

gwilson

rutt wrote:

IMHO, surface blur alone is worth the upgrade cost. But with CS3 around the corner, jeez, it's a little hard.

With surface blur, you don't lose much detail. Really, if you get the parameters right it only blurs where there is no detail. But noise/detail are very hard to tell apart.

The shadow noise at high ISOs is really an essential property of light. As exposure time gets shorter and the sensor size smaller, how many photons hit the sensor and are measured becomes a statistical issue and the variance becomes statistically significant. So you really don't get good detail in the shadows in this case, at least not as good as in the highlights.

At least in the ballet shots, the main detail I lose is stuff the lighting director is really trying to lose anyway: the curtain and the floor.

These last few posts have contained excellent information. Thank you both. This site is incredibly helpful--

Gloria

Gloria

BinaryFx

Duffy Pratt wrote:

Thanks Rutt. That's a very interesting and clear approach. It sounds like its worth building an action for retrieving a K channel. There are other areas where I've started using it as a mask, to isolate shadow and/or highlight casts (as per the fiddler).

Hi Duffy, there is a faster simpler method for isolating tonal ranges such as deep shadows or highlights. Blend If sliders. A work in progress can be found here:

http://members.ozemail.com.au/~binaryfx/blendif.html

Blend if sliders can be a faster/easier alternative to layer masks, or they can be used in addition to layer masking. One may be able to perform 80% of a difficult masking operation via the blend if sliders and then finish off the rough areas with a layer mask.

Could you get acceptable results using Gaussian Blur? I have PS CS, not CS2, so I don't have Surface Blur available.

Yes, this provides the strongest averaging of pixels, but can also introduce haloing on sharp edges and desaturate fine colour detail. I like combinations of filters at smaller settings, such as median, smart blur and smaller gaussian blur to smooth things over but not introduce edge blur/halos on sharp edges. The 'blockier' median and smart blur help a lot. More on the subject of LAB mode AB or RGB/CMYK color blend noise reduction here:

http://members.ozemail.com.au/~binaryfx/howto_colourblur.html

For images that are in CMYK mode, I recommend fading the blur to color mode or working in a duped layer set to color blend. Due to the K channel and GCR, unique channel structure can be lost when converting out of CMYK and it will not be the same when reseparated back to CMYK. The most common example of this is black only elements such as text, keylines and drop shadows becoming CMYK rather than K only when converted back to CMYK from LAB.

Two points for discussion in this chapter review:

1. Has anybody tested and compared noise reduction in the AB channels vs. doing it in RGB in color blend mode? What did you find?

2. Camera Raw converters offer colour noise reduction, by default Adobe Camera Raw applies 25%. Has anybody tested and compared colour noise reduction in the camera raw converter vs. doing it in Photoshop in LAB or RGB color blend?


Regards,

Stephen Marsh.
http://members.ozemail.com.au/~binaryfx/

arodney

Sharpening techniques that don't take specific's of the input device or output device into the equation are pretty useless. Here's a much more logical workflow:

http://www.creativepro.com/story/feature/20357.html

As for Sharpening the L channel, you can skip the move into LAB, do it in RGB and fade using Luminosity. Not 100% identical but you fix the so called 'issue' with color noise and you have fade control. Faster, no data loss from hopping around from RGB to LAB and back.

There's no reason to make all this stuff so complicated!

Oh, you want to really reduce noise and you have the ability like the book image to shoot more than one image (cause it's not moving), use CS3 with the Align Layer command, make a smart object and use the Median blend mode option. Nothing will match this for clean noise removal!

http://photoshopnews.com/2007/03/27/image-stacks-in-photoshop-cs3-extended/

Simple, quick, no color space jumping around. You get the point <g>.

edgework

BinaryFx wrote:

Yes, this provides the strongest averaging of pixels, but can also introduce haloing on sharp edges and desaturate fine colour detail. I like combinations of filters at smaller settings, such as median, smart blur and smaller gaussian blur to smooth things over but not introduce edge blur/halos on sharp edges. The 'blockier' median and smart blur help a lot. More on the subject of LAB mode AB or RGB/CMYK color blend noise reduction here:

Lens blur is the best way to avoid haloing, since that is its precise raisone d'etre. Well, not specifically; it's meant to mimic depth of field blurring, but it does so by allowing you to mask out areas that you want to exclude. It is very precise, far more so than the later Surface Blur. It requires a mask, which brings its own host of problems, but, once created, the blurring is totally under your control.