Recovering Segmentation Errors in Handwriting Recognition Systems

Most handwriting recognition systems need a mechanism for handling classification errors. These errors are typically caused by the large shape variability of the handwriting produced by different writers and by the segmentation errors, which occur when the word recognition process is performed by extracting and classifying single characters. In this paper, in order to reduce the segmentation errors, we propose a hierarchical recognition system composed of two classification modules. The first one discriminates isolated characters from cursive fragments using specifically devised features. The second one is an OCR engine that receives as input only those samples classified as isolated characters in the previous module. The whole system works like a highly reliable OCR that rejects most of the cursive fragments avoiding their incorrect classification. The experimental results confirmed the effectiveness of the proposed system.