2018/02/27 by Christian Reul, Reul, Christian, Uwe Springmann +5 · 1 citation
Computer Science · Arts and Humanities · #Handwritten Text Recognition Techniques #Mathematics, Computing, and Information Processing #Digital Humanities and Scholarship
paper · pdf · doi:10.48550/arxiv.1802.10038
We combine three methods which significantly improve the OCR accuracy of OCR\nmodels trained on early printed books: (1) The pretraining method utilizes the\ninformation stored in already existing models trained on a variety of typesets\n(mixed models) instead of starting the training from scratch. (2) Performing\ncross fold training on a single set of ground truth data (line images and their\ntranscriptions) with a single OCR engine (OCRopus) produces a committee whose\nmembers then vote for the best outcome by also taking the top-N alternatives\nand their intrinsic confidence values into account. (3) Following the principle\nof maximal disagreement we select additional training lines which the voters\ndisagree most on, expecting them to offer the highest information gain for a\nsubsequent training (active learning). Evaluations on six early printed books\nyielded the following results: On average the combination of pretraining and\nvoting improved the character accuracy by 46% when training five folds starting\nfrom the same mixed model. This number rose to 53% when using different models\nfor pretraining, underlining the importance of diverse voters. Incorporating\nactive learning improved the obtained results by another 16% on average\n(evaluated on three of the six books). Overall, the proposed methods lead to an\naverage error rate of 2.5% when training on only 60 lines. Using a substantial\nground truth pool of 1,000 lines brought the error rate down even further to\nless than 1% on average.\n