topic Convolution Scaling in Intel® Integrated Performance Primitives

Convolution Scaling

Ryan2 — Thu, 24 Nov 2011 18:21:55 GMT

Hi,

Using ippiConvFull_32f_C3R yeilds very large luminance values. Using a softclip algorithm gives me aesthetic results, but I'd like to inquire about proper post scaling methods for ippiConvFull.

Any thoughts?

Best,

Ryan

Convolution Scaling

Ryan2 — Thu, 24 Nov 2011 22:13:53 GMT

Divided the Kernel by sqrt(Kernel_Width*Kernel_Height)

Convolution Scaling

Chao_Y_Intel — Fri, 25 Nov 2011 01:41:22 GMT

Hi,

If you want to normalize the data, is it to divided the kernel by sum of the all kernel value?

Thanks,
Chao

Convolution Scaling

Ryan2 — Fri, 25 Nov 2011 16:43:47 GMT

Hi,

I actually just worked that out. Thank you for replying.

Here's the code I used for anyone else who's wondering:

[cpp]	//Calculate Sum of Kernel
	ippiSum_32f_C3R(Ipp_Ker, Ker_StepBytes, Ker_Size, KerScaled, ippAlgHintFast);

	// Clamp KerScale to 1.
	for (i=0; i<3; i++)
	{
		if (KerScaled > 1)
			KerScalef = float(1) / KerScaled[0];
		else
			KerScalef = 1;
	}
	// Multipy Ker by Inverse of Sum
	ippiMulC_32f_C3IR(KerScalef, Ipp_Ker, Ker_StepBytes, Ker_Size);[/cpp]