(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

2 1
4 2
2 1

kernel (the code's in image2.cc):

Start by including Image.h defining a namespace for clarity:

#include "lsst/geom.h"
#include "lsst/afw/image/Image.h"
namespace image = lsst::afw::image;
typedef image::Image<int> ImageT;
int main() {

Declare an Image

ImageT in(lsst::geom::Extent2I(10, 6));

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;
             ++ptr.x()) {
            *ptr = y;
        }
    }

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:

    //
    // Convolve with a pseudo-Gaussian kernel ((1, 2, 1), (2, 4, 2), (1, 2, 1))
    //
    ImageT out(in.getDimensions());  // Make an output image the same size as the input image
    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);
             ptr != end; ++ptr.x(), ++optr.x()) {
            *optr = ptr(-1, -1) + 2 * ptr(0, -1) + ptr(1, -1) + 2 * ptr(-1, 0) + 4 * ptr(0, 0) +
                    2 * ptr(1, 0) + ptr(-1, 1) + 2 * ptr(0, 1) + ptr(1, 1);
        }
    }

(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)

    //
    // Do the same thing a faster way, using cached_location_t
    //
    std::shared_ptr<ImageT> out2(new ImageT(in.getDimensions()));
    out2->assign(in);
    typedef ImageT::const_xy_locator xy_loc;
    for (int y = 1; y != in.getHeight() - 1; ++y) {
        // "dot" means "cursor location" in emacs
        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] +
                    2 * dot[s] + dot[se];
        }
    }

The xy_loc::cached_location_t variables remember relative positions.

We can rewrite this to move setting nw, se etc. out of the loop:

    //
    // Do the same calculation, but set nw etc. outside 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) {
        // "dot" means "cursor location" in emacs
        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] +
                    2 * dot[s] + dot[se];
        }
    }

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/=

    //
    // Normalise the kernel.  I.e. divide the smoothed parts of image2 by 16
    //
    {
        ImageT center = ImageT(
                *out2,
                lsst::geom::Box2I(lsst::geom::Point2I(1, 1), in.getDimensions() - lsst::geom::Extent2I(2)),
                image::LOCAL);
        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;
             ++ptr) {
            *ptr = 0;
        }
    }

we called the iterator xy_x_iterator, not x_iterator, for consistency with MaskedImage.

Finally write some output files and close out main():

    //
    // Save those images to disk
    //
    out.writeFits("foo.fits");
    out2->writeFits("foo2.fits");
    return 0;
}