Document categorisation: projects

Design and implementation of image processing modules for data entry applications
February 2007 - September 2009

Funded by STI S.p.A.
Partners: Signal Processing & Telecommunications Group - Numerical Image Processing (Dept. of Biomedical and Electronic Eng., University of Genoa, Italy)
Fundings:

• Phase I (Feb. - Dec. 2007): Euro 58,150.00 
• Phase II (June 2008 - Sept. 2009): Euro 70,650.00

The whole project in which our group is involved is named "Extended recognition of digitized documents" (REDD). Its aim is to develop a prototype of a system for automatic extraction and recognition of user text from scanned images of complex, real-world document forms like invoices and tax payment receipts. An example is shown in the figure below. The goal is to localize all the form fields, extract the text entered by the user and recognize it. Our task is to develop the first four processing steps: skew angle estimation and correction, form layout recognition, localization of the individual fields, removal of pre-printed form graphics and text and reconstruction of user text. The subsequent steps (text segmentation, character recognition and contextual post-processing) are in charge of the project partner DIBE.

This real-world task is particularly difficult because of image low quality due to noise introduced during form scanning, complex and non-fixed form layout and pre-printed form components, and the frequent superposition between user text and pre-printed form components.

In order to summarize the current status of our prototype, developed in collaboration with our Ambient Intelligence Laboratory at the Sardinia DisctICT, we make available a video and on-line demo tool. 

Example of a tax payment form considered in this project (original size: 165 mm (W) x 102 mm (H); user information has been removed).