Generalized form registration using structure-based techniques
Michael D. Garris, Patrick J. Grother · 1996
A new method for registering forms has been developed at the National Institute of Standards and Technol- ogy.This method automatically estimates the amount of rotation and translation in the image without any detailed knowledge of the form.This is accompUshed through the automatic detection of domin ant vertical and horizontal structures (lines) commonly found in forms.A general method for rotation estimation and a robust method for trans- lation estimation are presented.Results demonstrate that this technique is extremely tolerant to spurious annotations on the form and scanner noise in the image, and the computational requirements of the utihty can be tuned by option- ally choosing to process and analyze downsampled versions of the image.All 3,669 Handwriting Sample Forms distributed with NIST Special Database 19 were successfully registered with the new technique, and using the same code, 255 uniformly laid out IRS tax forms and 500 Census minifonns were also tested and registered.Every type of form contained in the numerous NIST (pubhc) form databases can be registered using this technique.These results also demonstrate how easy it is to set up the computer to register new types of forms, introducing a set-up interface that is much more automated and less tedious than what is currently required to specify new forms for the NIST public domain Form-Based Handprint Recognition System.