Eliminating Dark

Current Noise in
Long-Exposure
Digital Images

ne of the best things about digital photography conditions. For instance, infrared and night photog-
is the instant feedback. Snap a picture and raphy each require some practice.
you’ve got a 1x2-inch preview of the image you The downside to night photography with digital
just took, and minutes later you can view that cameras is dark current noise. In addition to higher grain,
image on a monitor or make a print. Compare long exposures turn certain CCD elements on, giving
this turnaround to the hours or days required to get your picture of the night sky a few extra stars! Fortunately,
prints or slides from film. This rapid feedback is terrific with a little care, this problem can be greatly reduced.
for learning how to take great pictures in unusual Below you see the original noisy image on the left and

Left: The original image (16 second manual exposure, f/2.8, 100 ISO equivalent) shows dark current noise.
Right: the final image after adjustment.

30 • PEI • JANUARY 2001

the same image after correction on the right. What’s the Difference?
Dark current noise is a kind of noise in long- Fortunately, the
exposure images taken on CCDs. Camera imagers sensitivity of CCD
consist of an array of CCD elements. Normally these elements to dark
elements are triggered by light entering the camera, but current noise is
they can also be triggered spontaneously by stray relatively stable.
electrons. Since electrons from noise are as good as The same pixels
electrons generated from light, this noise shows up as that are noisy in
bright pixels in the image where they shouldn’t be. one picture will be
How much of this spontaneous current occurs depends noisy in the next
on the exposure duration (longer exposures get more picture, if it’s taken
noise) and the camera temperature. At room temperature, soon and at about A portion of an image taken soon
today’s digital cameras begin to exhibit noticeable dark the same tempera- after the original, under very similar
conditions but with the lens cap on,
current noise for exposures longer than a few seconds. ture and conditions. reveals dark current noise.
Astronomers typically avoid dark current noise by Because of this level
cooling the CCD. The noise is typically halved for of stability, you can measure the noise and remove it.
every 5 degree Celsius drop in temperature, so To record what dark current noise is infecting your
cooling it to -30 to -40 degrees Celsius (-20 to -40 image, take another image (the “dark” image) immedi-
degrees Fahrenheit) avoids much of the problem. ately after taking your “light” image, using the same
Unfortunately, few photographers carry around liquid settings, but with the lens cap on. At the same temperature
nitrogen, and fewer still would want to apply it to and exposure, this dark image will record which pixels are
their cameras (something the camera makers probably generating noise. We’ll call this image the DCN correction.
appreciate!). Fortunately, we can remove most dark
current noise by post-processing the image.

Averaging Away the Problem

The simplest way to eliminate noise is to blur the
image slightly.
Since noisy pixels
are randomly
distributed around
the image, rather
than clustered
together, Filter>
Blur>Gaussian Blur
will quickly blend
them into the back-
ground. Blurring
will also reduce the
grain in the image,
not a bad thing for
long-exposure,
Noisy pixels are eliminated in this
image, blurred with Radius=1.5, but high-grain images.
the high-contrast edges are also lost. Of course, the
downside of
blurring is that it gets rid of the high-contrast edges in
the image in addition to the noise. For example, the
details in the refinery lights are gone. You can
partially correct for this by resharpening the image
(Filter>Sharpen>Unsharp Mask), but we’d really like
to change just the noise, not the whole image.
More info? PEInfo No. 39

PEI • JANUARY 2001 • 31

 CCD elements fail by becoming oversaturated. A duplicate it into another layer, and heavily blur that
natural way to fix this is to simply subtract these pixels. layer. (The exact blur radius depends on image
Take the base image, hold down the shift key and drag resolution, but don’t be shy. I used Radius=5.0.) No
the DCN correction image onto it to create a new layer, more noise here. Name this layer “DCN replacement.”
and select the Difference mode in the Layers palette.
Now the noise in the dark sky and trees is gone.
But there’s a downside to differencing. Subtracting
the noise works well for astronomers looking at stars at
night. Unfortunately for photographers, subtraction
creates the opposite problem of lit-up pixels: dark spots
now show up in bright parts of the image (look at the
refinery lights). Dark current noise occurred even in
these bright areas of the image. Although we don’t
want noise in the dark areas of the picture, we need to
keep it in the bright parts!

Turn up the gain in your DCN correction image to use it as a mask.

Now open the DCN correction image in another

window. When you open it, do not allow any color
management; the raw data is needed in the next steps.
We’ll use it to construct a mask to choose which pixels to
replace. Convert it to grayscale (Image>Mode>Grayscale).
Typically only a few elements are completely blown,
but many will be partly corrupted. We need to replace
all the pixels, so turn up the gain (Image>Adjust>
Levels; click “Auto”). Make sure the output levels are
Paste the dark image over the original image and select scaled 0 to 255 and that most of the image remains
Difference in the Layers palette to eliminate noise in the dark areas. black. This image will serve as the mask for the DCN
replacement layer, so select it all and copy it (this will
Adding Things Up be the “DCN mask”). Go back to the base image and
What we really want to do is not subtract off the select the “DCN replacement” layer in the Layers
noise, but to replace noisey pixels with the better ones. dialog, add a mask (Layer>Add Layer Mask>Reveal
This leads to our All). Then paste the DCN layer into the mask (opt/alt-
final approach, click on the mask and paste it).
combining these As a final step, you may wish to apply a light
two methods. We Gaussian blur (Radius=0.5 pixel) to the image, which
use blurring to will remove some of the remaining grain.
generate good data You’re now set to take long-exposure pictures at
for the replacement night. §
pixels, and the
noise-only image to
control which pixels
to replace.
First we need to
get good replace- John Heidemann is a research assistant professor of computer
science at the University of Southern California. Before
Use a very blurred image layer for
ment pixels. Take joining USC, he developed image processing software at the
your replacement pixels. the base image, University of Nebraska-Lincoln Remote Sensing Center.

32 • PEI • JANUARY 2001