Abstract

Rectangular objects frequently appear in certain classes of images. Scanned images of photographs, receipts or other small objects are obvious examples. It is often desirable to detect the existence, number and location of such rectangular objects to single them out for separate treatment. While image segmentation is generally a hard problem, we find that by considering one dimensional projections of the image it is possible to recursively simplify the segmentation problem efficiently and robustly. We outline the algorithm and its components and discuss its performance on real data from user scanned images.