Using Physical and Logical Constraints for Invoice Understanding

This work presents a methodology for invoice understanding. The invoices of our domain can be grouped into classes according to their logo. The understanding phase is based on two knowledge levels: a specific knowledge for each class, called a document model; and knowledge on the whole domain of interest, called a domain model. The invoices of a known class are understood by its document model, while the invoices of an unknown class are understood by using the domain model. The main contribution of this work is related to the use of the physical and logical constraints of the domain of interest for document understanding, without using an OCR system. Our approach has been tested by some experiments that are intended to identify some regions within invoices of unknown classes. In most cases, the results have shown the reliability of the approach.