An interesting thesis from 2010 showing PDF document reconstruction via machine learning in order to rebuild a PDF with an object model.

Can anyone shed light on what the state of this art is currently?