Classification based character segmentation guided by Fast-Hessian-Affine regions

This paper presents a method of fast and accurate character localization for OCR (Optical Character Reader). We already proposed an acceleration framework of arbitrary classifiers, classifier molding, for real-time verification of characters printed by Industrial Ink Jet Printer (IIJP). In this framework, the behavior of accurate but slow character classifier is learnt by linear regression tree. The resulted classifier is up to 1,500 times faster than the original one but is not fast enough for real-time pyramidal scan of VGA images, which is necessary for scale-free character recognition. For solving this problem, we also proposed CCS (Classification based Character Segmentation). This method finds character arrangement that maximizes the sum of the likelihood of character regions assuming that all characters are horizontally aligned with almost regular intervals. This assumption is not always true even for the characters printed by IIJP. For solving this problem, we extended the idea of CCS to arbitrary located characters. Our method first generates character-region candidates based on local elliptical regions, named Fast-Hessian-Affine regions, and finds most likely character arrangement. Through experiments, we confirmed that our method quickly and accurately recognizes non-uniformly arranged characters.