Gaussian Filter: A Gaussian Filter Smoothes An Image by Calculating Weighted Averages in A Filter Box

Gaussian Filter
A Gaussian filter smoothes an image

by calculating weighted averages in a filter box.
Coordinates xo, yo are arbitrary pixel positions in a bitmap image.

x,y is a local coordinate system, centered in xo, yo, as shown.
The gray area is a filter box with m·m knots.
x and y reach from -n to +n.
The box width m = 2·n +1 is assumed odd.
Weight factors are calculated for a Gaussian bell by w(x,y) = e-a
with a = (x2 + y2) /(2·r2 ) .
The filter radius r is in statistics the standard deviation sigma.
Select n = (2...3)·r for a reasonable reproduction without clipping.
E.g. for x = r, y = 0 we find w = e-0.5 = 0.6065.
The image shows the function relative to the filter box vertically shifted.
The radius is r = 2 in the image.
The algorithm on page 2 is not optimized.
The algorithm can be made much faster by binary weight factors, because
then the whole calculation is in Integer and the multiplications are merely
shifts (page 4).
Note:
The image quality is optimal only for direct view by Acrobat Reader.
Browsers are sometimes not accurate. Please use Zoom=100%
1
Gaussian Filter
General Weight Factors General Image Filtering

S=0 For yo=n to ymax-n Do
r2=2·Sqr(r) For xo=n to xmax-n Do
For y=-n to +n Do Begin
For x=-n to +n Do newred=0
Begin newgrn=0
a=(Sqr(x)+Sqr(y))/r2 newblu=0
w(x,y)=exp(-a) For y=-n to n Do
S=S+w(x,y) For x=-n to n Do
End Begin
newred=newred+w(x,y)·red(x+xo,y+yo)
newgrn=newgrn+w(x,y)·grn(x+xo,y+yo)
newblu=newblu+w(x,y)·blu(x+xo,y+yo)
End
newred=newred/S
newgrn=newgrn/S
newblu=newblu/S
End
Blur Control
The blurring is controlled by two parameters:
1) The box size, described by (2·n+1) pixels in one direction

2) The radius r
The Gaussian bell in one direction delivers:
x/r 0 1 2 3
w(x) 1.0 0.6065 0.1353 0.0111
We can choose r=0.465·n. This results in a weight factor 0.1 at the outermost
pixel, at x=n , which seems to be reasonable. Less than 0.1 doesn´t make much
sense. For pixels on the diagonal corners of the xy-box the value is smaller.
Weak blur: small box n=1 r=0.465

Weight factors 0.1 1.0 0.1
Strong blur: large box n=3 r=1.398

Weight factors 0.1 0.358 0.773 1.0 0.773 0.358 0.1
22
Improvements
The Gauss formula can be separated. This will make the calculations faster.
w(x,y) = e-(x·x+y·y) = w(x)·w(y) = e-x·x ·e-y·y
Another methods uses one source image, one array of the same size for the
accumulation and a sequence of shifted images. This shifted image is made once
for each position x,y for all pixels in the source image.
The author prefers the standard structure, because this is valid for any linear filter,
like softening, sharpening and contour finding filters, also for some effects which
use oszillations in the box.
A 5 x 5 binary Gaussian Filter, programmed mainly in Intel Assembly Language,
needs about one second for 1000 x 1000 pixels.
3
Binary Weighting
This filter shows a crude approximation

of the Gauss bell function. 0 1 2 1 0
The weight factors are powers of 2, thus
multiplications by weight factors can be 1 4 8 4 1
replaced by binary shifting. The sum of
2 8 16 8 2
the weight factors is S = 80. If all colors
values are 255, then the weighted sum is
1 4 8 4 1
20400, which does not exceed the positive
number space for Integer 215-1 = 32767. 0 1 2 1 0
Radius r= 0.85, approximately.
For more general applications binary weighting by LongInt (32bit) could be used.
This is still much faster than floating point operations.
Part of digital photo. Left: not filtered. Right: softened by Binary Weight Filter.
Then both images scaled down to 50% and placed into PDF by pixel synchronization.
The synchronization is perfect only for direct view by Acrobat Reader, not by
browser.
4
Binary Weighting Transfer Function
Gain transfer function for Binary Gauss 2, 8, 16, 8, 2
5
Arbitrary Filter Scaling
If the weight factors belong to a consis-

tent set of data, like in the previously men- -0 -1 -2 -1 -0
tioned softening filters, then we have to
divide all raw weight factors by the sum -1 -4 -8 -4 -1
the raw weight factors.
Not so for a general sharpening filter. -2 -8 +64 -8 -2
We start by a positive peak in the center
and negative values are taken from a -1 -4 -8 -4 -1
Gauss bell. The binary weighting is not
essential in this example. -0 -1 -2 -1 -0
Generally spoken, we have raw positive weight factors Pi and raw negative weight
factors N i .
The actual filtering is done by scaled weight factors pi = Pi / S p and ni = N i / S n .
S ip = Sum of positive weight factors Pi

S in = Sum of negative weight factors |N i | (absolute values)
S p = 0.5·S ip
S n = S in
This scaling is based on the demand, that a uniform color area should deliver the
same color after filtering.
Example, as above :
S ip = 64
S in = 4·8 + 4·4 + 4·2 + 8·1 = 64
The center peak in the scaled matrix is +2 and the other negative values are
divided by 64 .
6
A General Sharpening Filter
1 2 3 4 5 -0.0035 -0.0159 -0.0262 -0.0159 -0.0035
6 7 8 9 10 -0.0159 -0.0712 -0.1173 -0.0712 -0.0159
11 12 13 14 15 -0.0262 -0.1173 2.0 -0.1173 -0.0262
16 17 18 19 20 -0.0159 -0.0712 -0.1173 -0.0712 -0.0159
21 22 23 24 25 -0.0035 -0.0159 -0.0262 -0.0159 -0.0035
The drawing shows the numbering and the weight factors for the kernel for a
sharpening filter, here with n=2. The algorithm works for any n ≥1. The total
number of elements is m= (2·n+1)2. The center weight factor fs[13]=2.0 is a
positive peak. The other weight factors fs[k] are calculated by a negative
Gauss Bell, according to the code below.
The sum of negative weight factors is -1.0 and the sum of all weight factors is
+1, therefore a uniformly colored area remains unfiltered, as required.
Tutorial code, not optimized

sm:=0;
k :=1;
For j:=-n to n Do
For i:=-n to n Do
Begin
ra:=Sqrt(Sqr(i)+Sqr(j))/n;
ra:=exp(-2*Sqr(ra));
fs[k]:=-ra;
If (i<>0) Or (j<>0) Then sm:=sm+ra;
Inc(k);
End;
k :=1;
For j:=-n to n Do
For i:=-n to n Do
Begin
fs[k]:=fs[k]/sm;
If (i=0) And (j=0) Then fs[k]:=2.0;
Inc(k);
End;
Gernot Hoffmann
December 20, 2002
Website: please load browser and click here
7

Gaussian Filter: A Gaussian Filter Smoothes An Image by Calculating Weighted Averages in A Filter Box

Uploaded by

Copyright:

Available Formats

Gaussian Filter: A Gaussian Filter Smoothes An Image by Calculating Weighted Averages in A Filter Box

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gaussian Filter: A Gaussian Filter Smoothes An Image by Calculating Weighted Averages in A Filter Box

Uploaded by

Copyright:

Available Formats

Gaussian Filter

A Gaussian filter smoothes an image

Coordinates xo, yo are arbitrary pixel positions in a bitmap image.

General Weight Factors General Image Filtering

The blurring is controlled by two parameters:

1) The box size, described by (2·n+1) pixels in one direction

The Gaussian bell in one direction delivers:

Weak blur: small box n=1 r=0.465

Strong blur: large box n=3 r=1.398

This filter shows a crude approximation

Gain transfer function for Binary Gauss 2, 8, 16, 8, 2

If the weight factors belong to a consis-

S ip = Sum of positive weight factors Pi

1 2 3 4 5 -0.0035 -0.0159 -0.0262 -0.0159 -0.0035

6 7 8 9 10 -0.0159 -0.0712 -0.1173 -0.0712 -0.0159

11 12 13 14 15 -0.0262 -0.1173 2.0 -0.1173 -0.0262

16 17 18 19 20 -0.0159 -0.0712 -0.1173 -0.0712 -0.0159

21 22 23 24 25 -0.0035 -0.0159 -0.0262 -0.0159 -0.0035

Tutorial code, not optimized

You might also like