Altera Stratix - Page 333

Altera Corporation 7–47
September 2004 Stratix Device Handbook, Volume 2
Implementing High Performance DSP Functions in Stratix & Stratix GX Devices
Figure 7–26. Convolution Using a 3 × 3 Kernel
The output pixel value, y(m,n) depends on the surrounding pixel values
in the input image, as well as the filter weights:
To complete the transformation, the kernel slides across the entire image.
For pixels on the edge of the image, the convolution operation does not
have a complete set of input data. To work around this problem, the pixels
on the edge can be left unchanged. In some cases, it is acceptable to have
an output image of reduced size. Alternatively, the matrix effect can be
applied to edge pixels as if they are surrounded on the “empty” side by
ym n,()w
1
xm 1 n 1–,–()w
2
xm 1 n,–()w
3
xm 1 n 1+,–()++=
+ w
4
xm n 1–,()w
5
xm n,()w
6
xm n 1+,()++
+ w
7
xm 1+ n1–,()w
8
xm 1 n,+()w
9
xm 1 n 1+,+()++