Delphi blur Effect using VCL - image

Question

I'm implementing a custom panel blur effect. I would like to simulate a glass effect.
I found the following code in my searches:
function blur(source:TBitmap):TBitmap;
var
x, y, x3:integer;
tline, mline, bline:PByteArray;
begin
for y := 1 to source.Height - 2 do
begin
tline := source.ScanLine[y-1];
mline := source.ScanLine[y];
bline := source.ScanLine[y+1];
for x := 1 to source.Width - 2 do
begin
x3 := x*3;
mline^[x3] :=
(mline^[x3 - 3] + mline^[x3 + 3] +
tline^[x3 - 3] + tline^[x3 + 3] + tline^[x3] +
mline^[x3 - 3] + mline^[x3 + 3] + mline^[x3]) div 8;
mline^[x3 + 1] :=
(mline^[x3 - 2] + mline^[x*3 + 4] +
tline^[x3 - 2] + tline^[x*3 + 4] + tline^[x3+1] +
mline^[x3 - 2] + mline^[x*3 + 4] + mline^[x3+1]) div 8;
mline^[x3 + 2] :=
(mline^[x3 - 1] + mline^[x3 + 5] +
tline^[x3 - 1] + tline^[x3 + 5] + tline^[x3+2] +
mline^[x3 - 1] + mline^[x3 + 5] + mline^[x3+2]) div 8;
end;
end;
result := source;
end;
The above function works fine, but since I am not the author and do not understand anything about image manipulation, I would add to this function the possibility of setting the blur level.
I also found the following code, which works and is configurable, but, it is quite slow to apply.
A temporary solution I found to "configure" the blur was to adjust function to run itself recursively according to the level of blur you want. See the code snippet below:
function blur(source:TBitmap; nTimes: Integer = 1): TBitmap;
var
x, y, x3, I : integer;
tline, mline, bline: PByteArray;
begin
if nTimes = 0 then Exit(source);
for I := 0 to nTimes do...
The question is: how to make a configurable and efficient blur effect or could anyone help me to improve the above function?

Answer 1

This function does not work correctly, perhaps due to mistakes in copy-paste.
Blur is kind of convolution. To get value for pixel in blurred image, you have to calculate weighted sum of source pixels in some vicinity.
Dest[y,x] = Sum(dy=-size/2..size/2, dx=-size/2..size/2) (Src[y+dy, x+dx] * Weight[dy, dx])
In your case coefficients (weights) are
1/8 1/8 1/8
1/8 0 1/8 or 1/8 * (1,1,1,1,0,1,1,1,1)
1/8 1/8 1/8
This is not Gaussian blur - its coefficients for 3x3 kernel are 1/16 * (1, 2, 1, 2, 4, 2, 1, 2, 1). Note that central pixel is involved with the largest coefficient.
Moreover, code does not exploit lower scanline - the third line in every sum should use bline.
In general, you can calculate and use Gaussian kernel of any size (formula depending on sigma is described everywhere), but direct use of large-sized filter is slow. Multiple applying of small Gaussian filter to the same image is equivalent to applying of some larger filter.
For large kernel size there is fast (but more complicated) approach - fast convolution through FFT (fast Fourier transform).