Most of the noise filter programs need a resonable time for filtering
images of more than one megapixel. Although mimp is a fast program to filter
noise it needs about 20-30 sec (P3/800Mhz) for a standard noise reductions level.
Other programs need about 100sec (neat image) to do the same. While it seems that
the time is not really important it is sometimes bothering to wait for
the noise reduction of 200 images which needs about two hours. Thus we have reduced
the noise reduction to the absolut minimum in time and we have written the algorithm
to fit into a FPGA. The result is a filter in mimp (option -fnr) which needs less than 2 secs
to filter a medium noisy 5 MP image. All the following images show a cropped region of a 5 MP
image and you can get the full image by clicking to the images shown in this page.
The original shot from a FZ-20 at ISO 400 shows a very bad noise. The image can not be
used for prints above the size of 10x15cm.
and now the noise filtered one, after 1.8 sec CPU time ..... (click to see the uncropped image)
A nearly noise free picture good for every print up to A4. Also the noise reduction at this
noisy ISO 400 image is not as perfect as the standard method (with color noise removal and
after noise filter image analysis) the result is really surprising! And note with a FPGA
the noise reduction would need 1/10 sec only.
Let us compare it to the image the ISO 100 image (taken at the same time).
The noise filtered ISO 400 image has nearly the same noise but is still
a little bit more blurry than the ISO 100 one.
However the best area of application for the fast noise filter is to apply it to ISO 100
images. Many of the small sensor cameras have still the problem that noise is not completely
removed if making shots at ISO 100. You see this in the above image as a fine noise
in the light grey regions.
So let us apply the fast noise filter. Also we are not using the same strong filtering
the time needed is 2 sec too, because fast noise filtering is completely time independent
from the strength of the filter.
Yes indeed, everthing is perfect and you can compare it to images from SLR's.
All noise is perfectly reduced and no structures are lost.
But let us play a little bit with the ISO 400 image. The one pass filter
(we are looking to each pixel exactly one time) has trouble with color noise which is not perfectly
removed. Thus let us run the color noise reduction
filter first - and after this the fast noise reduction.
If you click to enlarge the image you see that also in very dark regions the noise
is perfectly reduced. However the time needed is about 10 sec because color noise reduction
is not versy fast (but I'm working on it).
