A Multi-level Architecture for Recognition of Polish Handwritten Medical Texts

The paper discusses problems of handwritten text recognition in medical information systems. Aspects specific for medical notes recognition are considered. Limited set of authors, relatively small dictionary of typically used words, characteristic phrases used in particular categories of documents can be utilized in order to increase the accuracy of handwritten text recognition. The concept of five-stage recognition pipeline and its implementation in flexible text recognition system cooperating with hospital information system are described. The recognition pipeline consists of document category identification, writer recognition, isolated character and word classification and finally sentence recognition based on syntax language model. At the stage of character and word recognition, used classifiers are trained individually for each author and for each text category. Preliminary results of medical text recognition using techniques proposed in recognition pipeline are presented. The proposed recognition method can be applied to hand-printed text recognition as well as to cursive unconstrained script.