Image comparison

[tbreloff]

See https://groups.google.com/forum/?hl=en#!topic/julia-dev/UwZ6L6cTLms for reference. The goal would be a comparison function that would give a measure of difference between two images (for writing tests, and maybe other uses).

Also mentioned was a visual comparison of the images, with a 3rd grayscale image which is the difference (of some to-be-determined measure) between each pixel.

[lobingera]

As mentioned in the thread, pdiff (as in pdiff.sourceforge.org) might be a solution, even if it does not provide a metric but tests against a threshold. pdiff as SW is GPLed, so there are the two options of taking it literally and update the build script, so that it can be accessed by a julia Package and live with the GPL OR re-implement based on the implementation and get the GPL on board as derivative work.

Reading the paper (http://pdiff.sourceforge.net/metric.html) shows the steps of a possible implementation in a clear way, so a clean-room implementation (without GPL as license) could be done. The comparison builds on two color transformations XYZ and Lab which i see avaiable in ColorTypes.jl and a convolution filter (laplace pyramid e.g. here http://academic.research.microsoft.com/Publication/268112/the-laplacian-pyramid-as-a-compact-image-code).

In any case, Images would provide the infrastructure for a comparison framework (testing size and binary equality, first and testing a perceptual metric as option).

[timholy]

Already implemented in #367. #373 is eating up all my time right now, though, and I don't want to merge anything until that gets done.

[lobingera]

Do you have a reference for the algorithm in #367?

[timholy]

No, but it's essentially obvious: blurring reduces the impact of differences due to "same content, small shifts in alignment." (I've noticed different Cairo library versions producing what I think are small shifts, though I didn't look carefully.) Very simple, but very effective for many real-world issues we want to resolve in automated tests.

[timholy]

Closed by #367.