Automated page layout simplification of Patrologia Graeca
Bruce Robertson, Christoph Dalitz, Fabian Schmitt · 2014
This paper illustrates how automated image preprocessing can improve OCR results of an important 19th century series of Greek authors. The roughly 75,000 ancient Greek pages from the volumes published by J.-P. Migne known today as Patrologia Graeca have a wholly regular layout, but one that is not easily reduced to a reading order by standard layout-analysis algorithms. We use a Hough transform and k-nearest-neighbor distance rejection to locate and then remove the features that confuse modern layout analysis. This allows us to separate the valuable Greek text from the less useful Latin translation, greatly simplifying later stages of processing, including de-hyphenation, automatic spell-checking and editing. Our implementation is based upon the Gamera Python package. We make the source code of our solution freely available.