I have an automated visual regression test harness for my web app. It takes baseline and change snapshots using puppeteer and compares them using pixelmatch. If I change the height of an element on the screen however, it pushes everything else on the page down slightly and pixelmatch identifies it as a difference.
I'm wondering if there are any models that can identify visual differences between two very similar images but is robust enough to ignore "slight" layout changes.
The first idea to come to my mind, is to calculate the pixel-wise difference, and then pass it through a median filter of a suitable radius. But if your element's height changes significantly and it pushes everything down, then this will be seen as a significant change even after the median filtering.
If you want to ignore all differences due to constant shifting, you'll need to employ more complex algorithms. Then edit-distance and dynamic programming comes to my mind.
