An evolutive OCR system based on continuous learning

Abstract

The paper presents an evolutive OCR system based on a cooperation between the recognition stage and the contextual stage which makes possible continuous training. The authors use the contextual correction in order to modify the behavior of the recognition stage by adjusting the internal representation of character models. They also introduce a specific classifier suitable for continuous training. The proposed classifier is based on the k-nearest neighbor rule modified by the introduction of weights. During the continuous training, the system selects models of pattern which contribute actively to a correct recognition.