Automated layout recognition
Lynn Golebiowski · 2004
To develop document image layout classifiers, each document image is represented by a set of labeled polygons corresponding to the pair-wise relationships between objects on the page. "Wanted" and "Unwanted" training sets are used to generate a polygon weight based on frequency of occurrence in both sets (term frequency). Unknown documents are scored by comparing polygons to those occurring in the wanted set. A score, weighted by the term frequency for the matching polygons, is computed. Experiments are performed against the NIST Structured Forms Database based on single and multiple layout collections using a variety of training samples.