A Study of the Factors Influencing OCR Stability for Hybrid Security

Abstract

Optical character recognition (OCR) is a critical task in securing hybrid (digital and paper) documents. For this, its key performance criterion is stability. An unstable OCR algorithm will fail to detect two copies of the document as similar thus creating a wrong fraud detection. Having a sufficiently stable algorithm requires a very high level of performance. To improve it, we study a simple disambiguation technique called "alphabet reduction". It is based on the principle that characters that are visually similar should be the same character. It significantly improves the stability of two state of the art OCR algorithms on almost forty three thousand images. Yet the obtained stability is still insufficient. We also study the impact of the document variations on the stability of OCR algorithms.