Boosting Offline Handwritten Text Recognition in Historical Documents With Few Labeled Lines

In this paper, we face the problem of offline handwritten text recognition (HTR) in historical documents when few labeled samples are available and some them contain errors train set. Three main contributions developed. First analyze how to perform transfer learning (TL) from a massive database smaller database, analyzing which layers model need fine-tuning process. Second, methods efficiently combine TL data augmentation (DA). Finally, an algorithm mitigate effects incorrect labelings training set is proposed. The analyzed over ICFHR 2018 competition Washington Parzival. Combining all these techniques, demonstrate remarkable reduction CER (up 6% cases) test with little complexity overhead.