Intel® Integrated Performance Primitives
Deliberate problems developing high-performance vision, signal, security, and storage applications.

ippiResize: interpolation quality

pld
Beginner
272 Views
Hi folks,
taking a closer look at the results of scaling using ippiResize_8u_C3Rit looks as though linear,cubic or lanczos interpolation all produce extremely similar (and unexpectedly 'blocky') results (this with IPP 5.1).
Using super even when downscaling produces results closer to what I'd expect linear to return.
This strikes me as a little bit odd, so any hints as to what I might be missing would be greatly appreciated.
Thanks,
--
Patrick
0 Kudos
3 Replies
Vladimir_Dudnik
Employee
272 Views

Hi Patrick

the difference is not big at downscaling, but will be noteiciable at upsampling. (super does not work at up-sampling). BTW, what resize parameters do you use? What are resize factors in X and Y axis?

Regards,
Vladimir

0 Kudos
pld
Beginner
272 Views
Hi,
it was most apparent when downscaling by a large factor (e.g. 3000x2000 -> 384x288), where there was all kinds of aliasing issues in the resulting images.
I guess there was just a missing step... the last libraries I'd used obviously pre- (or post-?) filtered images internally before doing the actual scaling, whereas ippiResize is just a "naked" resize.
Even a simple box filter before the resize produces better results, so I guess now it's just a matter of mixing & matching the right filter & interpolation :)
Cheers,
--
Patrick
0 Kudos
Vladimir_Dudnik
Employee
272 Views

Patric, yes, thanks for your update. Remember that IPP is set of primitive functions, just building blocks, which (due to its primitivism) can easely be combined into any complex processing algorithm.

Vladimir

0 Kudos
Reply