[Image-SIG] another out by 0.5 patch

[Douglas Bagnall]

Fredrik Lundh wrote:

>>In my tests, all the resize and stretch filters appear to shift the
>>image down and to the right by about half an output pixel. 
>  
> I've attached a (preliminary) patch.
> 

It certainly works for me.  Area-averaged and stretched images are 
mostly visually indistinguishable -- the exceptions being high contrast 
drawings.

Stretch-resized images are consistently 1/256th darker than the same 
images resized by other means.  My understanding of this is that when
a region of input pixels is (say) all 255s, except for one pixel of 254, 
then the float average is 254.9xx, so the output pixel becomes 254.  On 
average this should darken by 0.5, but over the two passes, it darkens 
by a unit.  I've attached a patch.


Here's an example, resizing a 450x325 image to 16x12 (as seen on 
http://halo.gen.nz/pil/examples.html), showing mean, RMS, and extreme 
differences between pixels.

before the patch, comparing stretched(NEAREST) to gimp:

R:  mean: 1.020833  RMS: 1.190238, extrema: (-2.0, 4.0)
G:  mean: 0.973958  RMS: 1.120361, extrema: (-1.0, 4.0)
B:  mean: 0.911458  RMS: 1.058202, extrema: (-1.0, 3.0)

with the patch:

R:  mean: 0.010417  RMS: 0.595119, extrema: (-3.0, 3.0)
G:  mean: 0.000000  RMS: 0.595119, extrema: (-2.0, 3.0)
B:  mean: -0.08854  RMS: 0.535218, extrema: (-2.0, 2.0)


Contrary to my earlier assertion, I have discovered the area averaged 
images differ from the gimp on the bottom row of the image.

comparing area-averaged to gimp:

R:  mean: 0.041667  RMS: 0.204124, extrema: (0.0, 1.0)
G:  mean: 0.052083  RMS: 0.228218, extrema: (0.0, 1.0)
B:  mean: 0.020833  RMS: 0.144338, extrema: (0.0, 1.0)




douglas

-------------- next part --------------
A non-text attachment was scrubbed...
Name: pil-antialias.diff
Type: text/x-patch
Size: 2276 bytes
Desc: not available
Url : http://mail.python.org/pipermail/image-sig/attachments/20050513/38fefd50/pil-antialias.bin