Form identifiction and skew detection from projections

In this paper we describe a system we have built to solve the preprinted forms identification and field extraction problem for Optical Character Recognition (OCR) applications. The strength of this system is that unlike other approaches it solves the problem in the most general and unrestricted sense. It works equally well for any type of preprinted form because it does not rely on any special features like patterns of line crossings or other symbols found only in a particular type of form. We have used the power spectrum as a shift invariant feature vector of the form's horizontal projection from which we identify the type of form and detect rotation. The horizontal and vertical projections themselves are also used to detect the shift of the form. Unlike the expected loss in response time to the benefit of generality, the proposed system is fast, highly accurate, even at reduced resolutions and with minimal user intervention it can be trained to recognize new types of forms.