Tutorial: Histogram

Original implementation
GIL implementation

This is a short tutorial presenting an example of a very simple sample of code from an existing code base that calculates histogram of an image. Next, the program is rewritten using GIL featres.

Original implementation

Actual code from a commercial software product that computes the luminosity histogram (variable names have been changed and unrelated parts removed):

void luminosity_hist(
    std::uint8_t const* r, std::uint8_t const* g, std::uint8_t const* b,
    int rows, int cols, int sRowBytes, Histogram* hist)
{
    for (int r = 0; r < rows; r++)
    {
        for (int c = 0; c < cols; c++)
        {
            int v = RGBToGray(r[c], g[c], b[c]); // call internal function or macro
            (*hist)[v]++;
        }
        r += sRowBytes;
        g += sRowBytes;
        b += sRowBytes;
    }
}

Let’s consider the following issues of the implementation above:

Works only for RGB (duplicate versions exist for other color spaces)
Works only for 8-bit images (duplicate versions exist)
Works only for planar images

GIL implementation

template <typename GrayView, typename R>
void grayimage_histogram(GrayView& img, R& hist)
{
    for (typename GrayView::iterator it=img.begin(); it!=img.end(); ++it)
        ++hist[*it];
}

template <typename View, typename R>
void luminosity8bit_hist(View& img, R& hist)
{
    grayimage_histogram(color_converted_view<gray8_pixel_t>(img),hist);
}

Using the Boost.Lambda library (or C++11 lambda) features it can written even simpler:

using boost::lambda;

template <typename GrayView, typename R>
void grayimage_histogram(GrayView& img, R& hist)
{
    for_each_pixel(img, ++var(hist)[_1]);
}

Let’s consider the following advantages of the GIL version:

Works with any supported channel depth, color space, channel ordering (RGB vs BGR), and row alignment policy.
Works for both planar and interleaved images.
Works with new color spaces, channel depths and image types that can be provided in future extensions of GIL
The second version is as efficient as the hand-coded version

Shortly, it is also very flexible.

For example, to compute the histogram of the second channel of the top left quadrant of the image, taking every other row and column, call:

grayimage_histogram(
    nth_channel_view(
        subsampled_view(
            subimage_view(img,
                0,0, img.width() / 2, img.height() / 2), // upper left quadrant
                2, 2                                     // skip every other row and column
            ),
        1   // index of the second channel (for example, green for RGB)
    ),
    hist
);

Since GIL operates on the source pixels of img object directly, no extra memory is allocated and no images are copied.