What the heck is this, II ?

[Skipper]

Check this out! This looks like the same color spillover I complained about in a recent thread but it's not. In that thread, Paul explained that the color spillover was likely (or definitely) related to jpeg compression and said not to shoot in jpeg to avoid in future.

This time, these photos were shot in RAW. When I converted to jpeg the teeth were still white. But, when I applied a noise reduction filter -- the one that came with PS CS3 -- at a level of 6, look what happens to the teeth! Pink! Luckily I rarely use noise reduction and only just started using it. Won't use again.

Do the separately-purchased noise reduction softwares do this? Like.... Neat Image and Noise Ninja?

The first is without noise reduction. The second is with.

Edit: Wow, notice the color difference in the face too. That's weird.

Message Edited by Skipper on 07-22-2008 12:32 PM

[Swimmerjoe2]

I can't speak for other programs, but for Photoshop's noise reduction filter this is what happens when you go overkill on the noise reduction. Basically you have set the luminance and color noise reduction parts so high that not only are you telling it you want the small bits of noise reduced, but also to use such a large radius for it's analysis that it is deciding that detail is noise as well and that it should reduce it. Don't use such a strong setting, especially for the color noise reduction.

Conversely, using Photoshop's color noise reduction can actually cure chromatic aberrations and color moire as well!

[AstroImager]

Yep, joe's got it right.  The color part of Photoshop's noise reduction (and almost everybody else's) reduces color noise by averaging color over a number of pixels.  The stronger you apply it, the larger the area over which color is averaged...and the more colors will "smear" over small details and bleed into one another.  It will also reduce saturation.

 

Stick to luminance noise reduction as much as possible, no matter what noise reduction software you use.  That's usually where most of the noise is anyway.  If you are still left with blotchy color noise, you can either deal with it manually (using the clone tool in color mode to deal with particularly noisy areas), or go ahead and apply the color noise reduction then use the same technique (clone tool in color mode) to touch up places where the color spread out too far.

 

Or shoot so that you don't get much noise in the first place

(that doesn't mean, by the way, that you have to shoot at super low ISOs all the time -- it *does* mean that no matter what ISO you shoot at, you get enough signal!  A well-exposed ISO 1600 image with plenty of signal from most DSLRs has very little noise, it's under-exposed images that show the most)

 

Paul

 

[BobF]

Skipper, since you're using CS3, try the colour noise reduction in ACR. I find it works really well. I've used it a fair bit on JPEGs and there's no colour bleed. The luminance noise reduction softens the shot but the colour NR doesn't. What ends up happening is that the colour noise ends up looking like film grain.

I've attached a copy of your first JPEG that I ran through the colour NR in ACR. Have a look at this thread as well where I attached samples of ACR's noise reduction. The shot is an in camera JPEG at ISO 1600 from a 40D

[peirceman]

Skipper,

The only thing I would add is to use noise reduction selectively. Unless you just have a horribly noisy image to start with, noise on your subject's teeth should not be that noticable. I like using a layer mask to only apply noise reduction to those areas that really need it.

[Skipper]

Thanks a lot everyone. That is super helpful. I thought I had set the NR setting on a medium level of 6 but obviously it was too high and apparently other settings also. Glad to know what was going on. Thanks for all the tips on selective NR and color NR.

To tell you the truth I was not applying NR here to address noise per se. I was experimenting with doing it to give a photo a polished smooth look as I know some photographers routinely run all their portrait photos through an NR program. I was admonished a few months ago here on forum for suggesting this was the right way to treat photos (I forget by whom -- someone who was probably right). But I was still experimenting with it anyway.

Paul, What is a signal?

Thanks again everyone!

[BobF]

Skipper, in this case, signal can be equated to light or exposure. Light hitting the sensor gets translated into an electronic signal that is used to determine exposure. More light = more signal = more exposure. It affects the signal to noise ratio. Noise is a constant. So if the signal is lower (lower exposure) the signal to noise ratio is negatively impacted and the noise is more evident.

If you're looking for that smooth, creamy portrait look, I'd suggest that using noise reduction isn't the best way to do it. Using selective blurring with layer masks and blend modes is a far better way to do it.

Message Edited by BobF on 07-23-2008 09:31 AM

[AstroImager]

Yep. 

This is where the "expose to the right" idea has merit -- no matter what ISO you're using, expose so that the brightest part of your histogram is as far right as possible without clipping the whites...that gives you the maximum amount of signal, which in turn means noise will be the least visible.

 

Skipper, here's an exercise for you to try: 

Open an image in PS.  Duplicate the background layer twice, so that you have three identical layers.  Set the layer blend mode of the top layer to "Color."  For the middle layer, change it to black and white (by using the "Black and White" tool at the default settings), and set its layer blend mode to "Luminosity."  Your image will look the same, but it's now actually an LRGB image -- monochrome luminance with an overlaid RGB color layer (the bottom layer in this case is irrelevant, but has to be there since PS won't let you change the blend mode of the Background layer).

 

Now select the top (color) layer, and do a Gaussian Blur -- with the Gaussian Blur dialog box open and "Preview" checked, try different radius settings for the blur, and watch how it affects the image.  You'll find you can actually do a pretty fair amount of blur (usually around 4-6 pixel radius) without seeing a whole lot of difference in the image, except that it'll look a little smoother...our eyes notice changes in luminance (which controls spatial resolution and brightness) much more than color.  Beyond 4-6 pixels of blur, however, you'll see that the colors start to bleed on edge lines, and start to desaturate -- if you push the radius up to 100 or higher, the image will get *really* desaturated, but will still be sharp because you're not affecting the luminance layer at all.  It's actually a good tool to use when you want a softly desaturated look.  Now close the GB dialog without applying the blur.

 

Select the luminance layer, and do the same thing -- notice how fast the image starts to look soft and how fast you lose detail, even at low settings of 1-2 pixels.  The luminance part of an image is where your brain gets all of  the sharpness information it processes, so it doesn't take much to make it look very, very soft.  Again, however, making an LRGB image and doing things just to the luminance layer is a good tool to have around -- some very gentle blur (or noise reduction) on the luminance followed by a low-radius but high percentage unsharp mask can give you a "smoother" image that retains good edge sharpness, and may just be the kind of "creamy" image you're talking about. 

 

Give it a try

 

Paul

 

[Skipper]

Thanks very much guys for all that really great information. I had not heard the term 'signal' before. 'Expose to the right' means expose for the highlights right? That's so true, as anyone who has ever tried to retrieve information from blown out highlights can attest.

Not sure if I want smooth and creamy. Maybe. I was going for polished. Hmmm.... Thanks Bob, I'll play around with that.

Paul, Thanks for that exercise. I'll try it. It does involve a lot of things I've never done before but I'll give it a go.

Let me ask you this -- what do you think the histogram should look like of a photo of properly exposed 30 white people wearing white tops and they are in a yard with a dark wooded background? A, B or C?

[BobF]

None of the above. It'll likely have a peak on the left representing the dark, wooded area background; a peak on the right representing the people and their white clothing with a trough in the middle representing the lower number of middle toned values.