You can still do that.

The only thing I'm proposing is to do the filtering on graphical data (what the user actually sees) rather than on the DOM (which can be obfuscated; and which may change every week).

Are you saying you would get the image, detect the ad in the image, then regenerate the HTML without the ad based on where it was located in the image?

I don’t see how presenting an image would allow you to interact with the page, especially if all you are doing is modify the image you’re presenting, not the page itself.