[kasparloog]

You can write a whole PhD thesis about it. Basically it is a result of a 2+ year research project how to detect differences that are relevant to people. In fact the computers find a lot more differences when you compare pixel-by-pixel. The keywords are machine vision, artificial intelligence, machine learning.

The hardest part of working out the right balance of what to detect was to make the algorithm less picky without losing an eye for detail.

We've benchmarked the algorithm against professional layout testers and the fault detection quality is on the same level with them.