Alpha Compositing

This demonstration shows how alpha compositing can be done in 10 lines of code using OffsetArrays and PaddedViews

using Images
using OffsetArrays # provide `OffsetArray`

1. basic offset and pad operation

Assume that we have two objects with transparent color and we want to place them on a canvas with partially overlapping.

red_patch = fill(RGBA(1., 0., 0., 1), 24, 24)
green_patch = fill(RGBA(0., 1., 0., 1), 32, 32)
mosaicview(red_patch, green_patch; npad=20, nrow=1, fillvalue=colorant"white")

Assume the canvas axes starts from (1, 1), here we keep the red patch unshifted, and shift the green patch 6 pixels downward and 16 pixels rightward, and then pad them to a common axes

green_o = OffsetArray(green_patch, 6, 16)
r, g = paddedviews(Gray(0.2), red_patch, green_o)
mosaicview(r, g; npad=20, nrow=1, fillvalue=colorant"white")

Note that their axes and sizes are changed after shifting and padding:

# Regardless of implementaion details, `Base.OneTo(32)` is mostly equivalent to `1:32`
println("before shifting -- size: ", size(green_patch), " axes: ", axes(green_patch))
println("after shifting  -- size: ", size(green_o), " axes: ", axes(green_o))
println("after padding   -- size: ", size(g), " axes: ", axes(g))

before shifting -- size: (32, 32) axes: (Base.OneTo(32), Base.OneTo(32))
after shifting  -- size: (32, 32) axes: (OffsetArrays.IdOffsetRange(7:38), OffsetArrays.IdOffsetRange(17:48))
after padding   -- size: (38, 48) axes: (1:38, 1:48)

Axes are preserved after padding, which means you can easily get original image from padded results using

r[axes(red_patch)...]
g[axes(green_o)...]

As described here, there are several compositing methods:

# add operation
out_add = r .+ g

# clear operation
out_clear = copy(r)
out_clear[axes(green_o)...] .= colorant"black"

# multiply operation
out_mul = copy(r)
# channel-wise multiplication
channelview(out_mul)[:, axes(green_o)...] .*= channelview(green_o)

# overlap operation
out_over = copy(r)
out_over[axes(green_o)...] .= green_o

# display the results of these operation
mosaicview(out_add, out_clear, out_mul, out_over;
           npad=20, nrow=1, fillvalue=colorant"white")

2. build the three-primary color panel

Now, let's use the same trick to build something more meaningful. First we create three circles with colors red, green and blue

using ImageDraw
function make_circle(sz, c::T) where T
    # fill with transparent color to avoid black region
    fillvalue = ARGB(c)
    img = fill(ARGB{eltype(T)}(0., 0., 0., 0.), sz...)
    origin = sz .÷ 2
    r = sz .÷ 4
    draw!(img, Ellipse(origin..., r...), fillvalue)
    img
end

# create three circles with color red, green and blue
red_c   = make_circle((256, 256), ARGB(1., 0., 0., 1.))
green_c = make_circle((256, 256), ARGB(0., 1., 0., 1.))
blue_c  = make_circle((256, 256), ARGB(0., 0., 1., 1.))

mosaicview(red_c, green_c, blue_c; nrow=1)

Then, shift these circles to appropriate positions, pad them to common axes, and finally composite using the add operation:

r = size(red_c, 1) ÷ 8
red_o   = OffsetArray(red_c,    r,  r)
green_o = OffsetArray(green_c, -r,  0)
blue_o  = OffsetArray(blue_c,   r, -r)

color_panel = sum(paddedviews(zero(eltype(red_o)), red_o, green_o, blue_o))
color_panel = color_panel[axes(red_c)...] # crop empty region

This page was generated using DemoCards.jl and Literate.jl.