Adaptive combination of commercial OCR systems

Abstract

Combining multiple classifiers to achieve improved recognition results has become a popular technique in recent years. As for OCR systems, most investigations focus on fusion strategies on the character level. This paper describes a flexible framework for the combination of result strings which are the common output of commercial OCR systems. By synchronizing strings according to geometrical criteria, incorrect character segmentations can be avoided, while character recognition is improved by classical combination rules like Borda Count or Plurality Vote. To reduce computing time, further expert calls are stopped as soon as the quality of a temporary combination result exceeds a given threshold. The system allows easy integration of arbitrary new OCR systems and simplifies the determination of optimal system parameters by analyzing the input data at hand. Quantitative results are shown for a two-recognizer system, while the framework allows an arbitrary number of experts. © Springer-Verlag Berlin Heidelberg 2004.