
332 IV Mobile Devices
7.4.2 Implementation Details
In our implementation, the input image I is an H × W array of unsigned 8-bit
integers (
uchar’s) and the output images O
dx
and O
dy
are H × W arrays of
signed 8-bit integers (
char’s). The results are computed as signed 16-bit integers
(
short’s) and are normalized by dividing by 8 (i.e., shifting right by 3). The
results are only computed for (H − 2) × (W − 2) inner pixels of the output
images, leaving pixels on the one-pixel wide border intact.
10
7.4.3 Performance Characteristics
Table 7.1 shows the results for all the kernel versions discussed in this section
on a 512 × 512 input image. We present the