Cost issues aside yeah image embedding tech is really good and can already do tasks like this, LLMs with native image understanding are even better at it.

Reasons this wouldn't work are you'd really need to get it to a good enough and small enough model to run low energy use and very fast to not impact browsing. The other issue is it'll be a fixed target for a period of time so the ads could be tweaked to find how to get them through it.