Automated framework for robust content-based verification of print-scan degraded text documents

Abstract

Fraudulent documents frequently cause severe financial damages and impose security breaches to civil and government organizations. The rapid advances in technology and the widespread availability of personal computers has not reduced the use of printed documents. While digital documents can be verified by many robust and secure methods such as digital signatures and digital watermarks, verification of printed documents still relies on manual inspection of embedded physical security mechanisms.The objective of this thesis is to propose an efficient automated framework for robust content-based verification of printed documents. The principal issue is to achieve robustness with respect to the degradations and increased levels of noise that occur from multiple cycles of printing and scanning. It is shown that classic OCR systems fail under such conditions, moreover OCR systems typically rely heavily on the use of high level linguistic structures to improve recognition rates. However inferring knowledge about the contents of the document image from a-priori statistics is contrary to the nature of document verification. Instead a system is proposed that utilizes specific knowledge of the document to perform highly accurate content verification based on a Print-Scan degradation model and character shape recognition. Such specific knowledge of the document is a reasonable choice for the verification domain since the document contents are already known in order to verify them.The system analyses digital multi font PDF documents to generate a descriptive summary of the document, referred to as \Document Description Map" (DDM). The DDM is later used for verifying the content of printed and scanned copies of the original documents. The system utilizes 2-D Discrete Cosine Transform based features and an adaptive hierarchical classifier trained with synthetic data generated by a Print-Scan degradation model. The system is tested with varying degrees of Print-Scan Channel corruption on a variety of documents with corruption produced by repetitive printing and scanning of the test documents. Results show the approach achieves excellent accuracy and robustness despite the high level of noise.