topic Yesterday I checked that in Intel® Integrated Performance Primitives

Is there an inverse to AlphaPremul?

David_N_ — Thu, 06 Jul 2017 19:44:16 GMT

Hello,

Is there an inverse to the AlphaPremul method?

Based on the thread here, it seems that AlphaComp always produces pre-multiplied results. If I want to save those results to PNG, I have to un-premultiply them. I dont see an inverse to the AlphaPremul method. Is there some other way of un-premultiplying an image, or so I have to write it by hand?

Hi David,

Zhen_Z_Intel — Fri, 07 Jul 2017 08:38:31 GMT

Hi David,

I am afraid we do not have, you probably could consider implemented by self. If you use AlphaPremul that means you already get C_r|g|b= A_r|g|b * α_A , (α_A=α_C) and you would like to get RGB value of A, you could calculate by C_r|g|b*1/α_C

Hello,

David_N_ — Fri, 07 Jul 2017 17:39:20 GMT

Hello,

Maybe you can consider this an addition for the future. Since you already have optimized code for doing the premul, it should be really easy do 1/alpha at the same time. Of course I can do it myself, but it likely will not be as optimized as the code you have for premul.

Have you tried ippsDiv?

BMart1 — Fri, 28 Jul 2017 15:15:17 GMT

Have you tried ippsDiv?

We did not try that. Do you

David_N_ — Fri, 28 Jul 2017 18:53:14 GMT

We did not try that. Do you think that one of the div functions will help in this case?

Yesterday I checked that

BMart1 — Fri, 28 Jul 2017 18:55:55 GMT

Yesterday I checked that ippiMul and ippiAlphaPremul are equivalent and remembered this thread, so ippiDiv should work with the caveat that it won't support the alpha packed with the RGB. You would need separate planes.

ippiDiv has the problem of

Adriaan_van_Os — Tue, 10 Apr 2018 13:21:24 GMT

ippiDiv has the problem of division by zero (returning an warning error code for u and s integer types and producing NAN_32F, INF_32F or INF_NEG_32F values for floating-point types). So, I do feel that AlphaUnPremul functions are missing. For example, Apple's vImage library has vImageUnpremultiplyData functions for the purpose.

Regards,

Adriaan van Os