Agreed. What I do for tables is use the OCR as a step 1, where I just extract the coordinates of each textbox. Then, use one of several methods to recognize co

Agreed. What I do for tables is use the OCR as a step 1, where I just extract the coordinates of each textbox. Then, use one of several methods to recognize columns (Hough transform, or we search for empty areas). Then, do some clustering to recognize lines. The problem is that there's a million of edge cases that you need to handle (split boxes that span multiple columns, in the case of tables recognize cells that span multiple columns or rows, etc.). There should be a solution for all these edge cases, but it gets tricky quickly.