Profiling leads to the next step, removing the noise, or filtering. We can tweak the software to remove more or less noise (via various parameters and sliders), or let the plug-in do it automatically. We can also set up user profiles, based on a sample image, although that often fails to consider the vagaries inherent in variable shooting conditions or scanner settings. (If you haven't guessed, I'm not a fan of user profiling.)

While color and luminosity noise can be seen in this Canon EOS 5D image, the problem is not pronounced, even at ISO 1600 (top). Photoshop ACR (center) and Noise Ninja (in Auto mode) (above) improved the shot equally well, although ACR was much more labor-intensive. Dfine left noticeable color noise, while Noiseware and Neat Image produced plastic images that needed some tweaking.

We often try to filter out color (chroma/chrominance) noise (evident as red/green/blue pixels or clusters) as much as possible, since the net effect on image quality is often simply to improve the image without noticeably degrading it. The same cannot be said for filtering out luminosity (gray scale/contrast) noise. If we attempt to remove too much of this mosaic of darker/lighter pixels, image quality becomes degraded and the photograph begins to look plastic and artificial. That's the point where you know you've gone too far, so be less aggressive, ensuring that just enough luminosity noise remains to give the image some texture.

Another important point, as my colleague Jon Canfield pointed out, is that when printing on textured paper, you may also be visually filtering out some of that luminosity noise (as well as low-level color noise). All noise becomes more pronounced as we enlarge the image. When making 4x6 prints, I wouldn't even bother with luminosity noise, although a modest correction for more pronounced color noise may be useful to avoid a flat-looking print (for film veterans, one that looks like it was made from an underexposed negative).

Plugging Away At The Problem
You can use the various noise filters provided in Photoshop (or other image editor), but they only go so far. Photoshop's noise filters tackle the problem from a different perspective. Despeckle is subtle and may be useful for removing faint specks. The stronger Median filter may blur the image when set too high, with a Radius of 1 pixel often enough to do the trick. However, Median appears to deal well with hot pixels (overexposure) when set to a Radius value of 2. Reduce Noise, I found, is best for removing color noise (when used with hot pixels, it leaves faint traces of those pixels). After using an editor's noise filters, you may need to employ a modest amount of sharpening (Smart Sharpen or Unsharp Mask), being sure not to reintroduce noise (or add artifacts) in the process. ACR also offers noise reduction, and I've found this step to be quite useful. Still, ACR NR may not always go as far as necessary. (Note: When converting from ACR for these tests, I used the default settings, except that I set NR at zero.)

#1

#2

#3

#4

A friendly interface is always welcome, and these screen shots show what greets you in each plug-in. The preview screen is equally important. Dfine (#1) and Noiseware (#2) give you a split screen that shows before and after images, but Dfine also gives you a magnified before/after view adjacent to the main preview window. Noise Ninja (#3) provides two windows, the smaller one for previews. Neat Image (#4) offers only one small window with alternating views (original/preview).

What an imaging application fails to do, however, is diagnose the problem and tailor a profile to the present set of circumstances. If you use a third-party NR plug-in, bypass ACR noise reduction and go straight to the plug-in. In fact, the documentation with every NR plug-in tells you to work with minimally processed images, at least where sharpening and noise reduction are concerned. The plug-in may include its own unsharp mask filter with an added dose of artifact (halo) prevention from sharpening, so it might be prudent to use this feature where available, instead of sharpening in the host application. Detail protection is another useful feature. This parameter prevents NR from blurring key details (but can also be tricky and work against you if used incorrectly).

Put To The Test
Noise Ninja, Noiseware, and Dfine let you sample multiple areas, whereas Neat Image limits you to one selection. In theory, one good sample is all you need, but I haven't found that to always be the case. However, you can manually override any selected sample in any of these plug-ins. Noiseware lets you further define frequency and tonal/color ranges to be addressed--which is good and bad, depending on how involved you want to get (you could find yourself spending too much time with these parameters). In the end, I found one or two sliders that needed tweaking and largely let the program do its thing. Noise Ninja has fewer sliders (parameters), which largely simplified the process. I found Neat Image's controls to be a bit cryptic and the preview window too small. Dfine gives you fewer parameters to tweak, which again can be good and bad, but lets you sample color ranges, or you can set up control points (using Nik Software's U Point technology). Control points target discrete areas in the frame.