My first thought was why not use the simple edge detection with connected components? Assume the document of interest is the prominent feature of high connectio

My first thought was why not use the simple edge detection with connected components? Assume the document of interest is the prominent feature of high connection. Discard high frequency (low length) line segments that are not connected to form the largest quadrilateral. Further segmentation could be done by having the user "tap" to select the document.