But shouldn't there be software that treats the entire page as an image, and from there reconstruct what parts are important?

Web design practices are far too unpredictable, and you can't rely on them to extract semantics.