Categorizing Document Images into Script and Language Classes

In order to properly archive and index large numbers of international documents, several challenging processing steps must be completed even before optical character recognition (OCR) can be applied. We present a system that preclassiies documents for further processing and OCR. The system operates in four phases: preprocessing (includ-We present a set of statistical techniques, based fundamentally on connected component analysis and horizontal projections, for the rst two phases. Even with little training, the system predicts the correct script category in 91% of the cases, when tested on real-life documents of varying kinds, diverse formats and qualities from many sources. The third and fourth phase are based on expert systems approaches. Language identii-cation combines several heuristics based on a statistical analysis of our training corpus. It currently has a 95% success rate on real-life documents of moderate quality. We will discuss the techniques and their combination and the process of improving performance.