(Return to Images)
Iterators provide access to an image, pixel by pixel. You often want access to neighbouring pixels (e.g. computing a gradient, or smoothing). Let's consider the problem of smoothing with a
kernel (the code's in image2.cc):
Start by including Image.h defining a namespace for clarity:
Declare an
Image Set the image to a ramp
for (
int y = 0; y != in.getHeight(); ++
y) {
for (ImageT::xy_locator
ptr = in.xy_at(0, y),
end = in.xy_at(in.getWidth(),
y);
ptr !=
end;
}
}
That didn't gain us much, did it? The code's a little messier than using x_iterator
. But now we can add code to calculate the smoothed image. First make an output image, and copy the input pixels:
ImageT out(in.getDimensions());
out.assign(in);
(we didn't need to copy all of them, just the ones around the edge that we won't smooth, but this is an easy way to do it).
Now do the smoothing:
for (
int y = 1; y != in.getHeight() - 1; ++
y) {
for (ImageT::xy_locator
ptr = in.xy_at(1, y),
end = in.xy_at(in.getWidth() - 1,
y),
optr = out.xy_at(1, y);
*optr =
ptr(-1, -1) + 2 *
ptr(0, -1) +
ptr(1, -1) + 2 *
ptr(-1, 0) + 4 *
ptr(0, 0) +
}
}
(N.b. you don't really want to do this; not only is this kernel separable into
1
2
1
in first the
x
then the
y
directions, but
lsst::afw::math
can do convolutions for you).
Here's a faster way to do the same thing (the use of an Image::Ptr
is just for variety)
out2->assign(in);
typedef ImageT::const_xy_locator xy_loc;
for (
int y = 1; y != in.getHeight() - 1; ++
y) {
xy_loc
dot = in.xy_at(1, y),
end = in.xy_at(in.getWidth() - 1,
y);
xy_loc::cached_location_t nw = dot.cache_location(-1, -1);
xy_loc::cached_location_t n = dot.cache_location(0, -1);
xy_loc::cached_location_t ne = dot.cache_location(1, -1);
xy_loc::cached_location_t
w = dot.cache_location(-1, 0);
xy_loc::cached_location_t c = dot.cache_location(0, 0);
xy_loc::cached_location_t e = dot.cache_location(1, 0);
xy_loc::cached_location_t sw = dot.cache_location(-1, 1);
xy_loc::cached_location_t
s = dot.cache_location(0, 1);
xy_loc::cached_location_t se = dot.cache_location(1, 1);
for (ImageT::x_iterator optr = out2->row_begin(y) + 1; dot !=
end; ++dot.x(), ++optr) {
*optr = dot[nw] + 2 * dot[n] + dot[ne] + 2 * dot[
w] + 4 * dot[c] + 2 * dot[e] + dot[sw] +
}
}
The
xy_loc::cached_location_t
variables remember relative positions.
We can rewrite this to move setting nw
, se
etc. out of the loop:
xy_loc pix11 = in.xy_at(1, 1);
xy_loc::cached_location_t nw = pix11.cache_location(-1, -1);
xy_loc::cached_location_t n = pix11.cache_location(0, -1);
xy_loc::cached_location_t ne = pix11.cache_location(1, -1);
xy_loc::cached_location_t w = pix11.cache_location(-1, 0);
xy_loc::cached_location_t c = pix11.cache_location(0, 0);
xy_loc::cached_location_t e = pix11.cache_location(1, 0);
xy_loc::cached_location_t sw = pix11.cache_location(-1, 1);
xy_loc::cached_location_t s = pix11.cache_location(0, 1);
xy_loc::cached_location_t se = pix11.cache_location(1, 1);
for (
int y = 1; y != in.getHeight() - 1; ++
y) {
xy_loc dot = in.xy_at(1, y),
end = in.xy_at(in.getWidth() - 1,
y);
for (ImageT::x_iterator optr = out2->row_begin(y) + 1; dot !=
end; ++dot.x(), ++optr) {
*optr = dot[nw] + 2 * dot[n] + dot[ne] + 2 * dot[
w] + 4 * dot[c] + 2 * dot[e] + dot[sw] +
}
}
You may have noticed that that kernel isn't normalised. We could change the coefficients, but that'd slow things down for integer images (such as the one here); but we can normalise after the fact by making an Image that shares pixels with the central part of out2
and manipulating it via overloaded operator/=
{
ImageT center = ImageT(
*out2,
center /= 16;
}
N.b. you can use the iterator embedded in the locator directly if you really want to, e.g.
for (
int y = 0; y != in.getHeight(); ++
y) {
for (ImageT::xy_x_iterator
ptr = in.xy_at(0, y).x(),
end = in.xy_at(in.getWidth(),
y).
x();
ptr !=
end;
}
}
we called the iterator
xy_x_iterator
, not
x_iterator
, for consistency with
MaskedImage.
Finally write some output files and close out main()
:
out.writeFits("foo.fits");
out2->writeFits("foo2.fits");
return 0;
}