|
|
micromysore wrote:
> I dont think the padding is an issue here.
>
> The difference image has higher values (>>0) where there is a texture
> in the image or edges.
>
> -mic
Just to make sure... you said that the max error was ~30 and that the
worst errors were on the edges. What's the max error excluding the edge
region [for a 10 tap filter, max(max(abs(J(5:end-5)-J1(5:end-5))))]?
In any case, the two filtering operations probably generate different
results innately because one does the operations sequentially and the
other does them in one fell swoop. Consequently, it could be that in
conv2, some error is introduced in the first convolution, and then it
is propogated in the input to the second, which makes it a little
worse. The filter2 operation does both operations at once and doesn't
suffer from this same problem. So without being able to see your data,
I'd still guess that the problem is in precision differences.
Cheers! --M
|
|