Case study - sharpening and noise

rutt

After reading this article by Dan Margulis I got very excited and thought I'd apply to a problem image of mine that's been vexing me recently. The image was shot for the current challenge and captures the sin of envy (there is some controversy about how well, but that's a different story.)

The image has three problems:

1. It's noisy. It was shot at ISO 1000 in low light.
2. The subject's face is not sharp. The shutter speed was slower than I would have liked, but there really wasn't an option under the circumstances.
3. Ferrari red as shot is out of gamut and so won't render properly on every monitor (especially with windows standard gamma.)

Here is the original just after RAW conversion. I used very conservative parameters in this conversion. The main customization was to white balance since museum lights are an odd mixture of tungsten and halogen.



and here is a more than 100% crop on the subject's face:



So I tried Dan's suggested trick of using the dust & scratch filter on the A and B channels. A strike out. No matter how I set the parameters, it made no difference that I could see. I won't show you the results. They are essentially unchanged from the originals.

Why? The problem is that noise on this image is in the L channel, not in the A or B channels. Here is a crop of the L channel only:



Why not try Neat Image? It's been working pretty well for me. I set it for the proper camera, ISO, and told it to reduce noise and sharpen a slightly out of focus image. Here is the result:



and here is a crop:



Not too bad, but still not winning any prizes. Noise is definately reduced, but sharpness as well. We've lost the creases in his forehead which were key to the message of this particular image. Something bad has happened around his ear. And there is a general "plastic" aspect about it. It's important that this is a middle aged guy and not a Barbie doll.

What about using the dust & scratch filter on the L channel where the noise really is? This isn't quite as sound theory as using it on the A and B channels where it really can't harm anything, but necessity is the mother of invention. Here is the result of applying the dust & scratch filter to the L channel with radius 2 and threshold 3:



and here is the crop:



This seems to be on the right track. Noise is about as low as in the Neat Image version but without the negative side effects just listed. What about sharpening this?

Here is the result of USM with amount 298, radius 1.6, threshold 19:



and the crop:



An aside about choosing values for the USM and the dust & scratch filter. Dan taught me a technique that seems to work well for choosing USM values:

1. Work in LAB. Trust me.
2. Select the LAB channel, then make all the channels visible by checking the eye box to the right of the composite row on the top.
3. Zoom to about 100% on an area of interest.
4. Bring up the USM dialog. Set amount to 500 and radius to 4
5. Adjust threshold until you aren't sharpening noise and other fine detail you don't want to. I find values between 10 and 30 often work well for me.
6. Adjust radius until the halos don't obscure detail in the image. I find values between .5 and 3 work for me. It depends a lot on the image, camera, lens, ISO, etc.
7. Adjust the amount until it actually looks good. Use the preview on/off button to compare to the unsharpened image. You want the image to look better and sharper without being able to see any obvious sharpening artifacts, halos in particular, even at 100% magnification.

My approach with the dust & scratch filter was similar. I turned up the radius a lot and played with the threshold to target the filter. Then I adjusted the radius to get the effect I actually wanted.

OK, noise is reduced and we do have some sharpening. The image is still too soft to win a photo contest, but fixing that is probably going to take some real skill with the wacom tablet, pencil and brush tools. But the image is much improved and we've learned something.

There is still the issue of fixing the out-of-gamut Italian racing red on the car. I fixed this by converting to CMYK (bringing the entire image into gamut automatically). Unfortunately, this also dulled the wonderful Ferrari red into a sort of used Buick orange. I restored the color by selecting the car and jacking the magenta channel quite a bit, but being careful to observe the ink limits. (Hint: view->gamut warning). Selecting the right area of the car was really really easy. There is absolutely no cyan there. So I used the magic wand in the C channel and one touch was all it took.

Here is the final image. In addition to the steps outlined above, I use L curves (before the noise reduction step, actually) to lighten the background and bring up the subjects face detail by steeping contrast across his face and made a final very moderate crop:

rutt

Pathfinder wrote (in a different thread):

pathfinder wrote:

I will give your method a try and see what I can find out. ALtho my experience with the dust and scartch filter when I was scanning slides was that it blurred the images significantly. In LAB, maybe it won't though too.

Jim, notice the very small values I've used in the scratch filter when I used in on the L channel. Dan used huge values on the A+B channels and that makes sense if that's where the noise is. It should be strictly better than blurring them.