What is the problem:
Sometimes (WPF, but I think it might be valid for other platforms as well) you might have very small difference in shadows caused by aliasing mechanisms.
In this case one of R,G,B is a bit different (+- 1 or 2).
Image comparison will count it as different pixel, but it is not visible for human beings.

See attached screenshots.
If you open expected and actual you will see no difference, but actually there is ~2-3% difference (colored in green on diff picture) due to shadowing and aliasing of WPF texts.

How we can improve image verification:
Implement something like “small difference” and “big difference”.
That will allow testers to be more flexible in image verification. 

I'm attaching a simple C# method that implements my idea.