Automatic metadata extraction via image processing using Migne's Patrologia Graeca

Abstract

A wealth of knowledge is kept behind libraries and cultural institutions in various digital forms without however the possibility of a simple term search, let alone of a substantial semantic search. One such important collection that contains knowledge, accumulated in the passage of the ages and remain inaccessible for the greater part, is Patrologia Graeca. So far, little research has been conducted to make this digital collection searchable to a certain degree, in order to retrieve and reveal its gathered knowledge in an efficient way. In this study, a novel approach is proposed which strives towards recognizing words from large printed corpora such as Patrologia Graeca. The proposed framework firstly applies an efficient segmentation process at word level and transforms the word-images of Greek polytonic script of the Patrologia Graeca into special compact shapes. Afterwards the contours of these shapes are extracted and compared with the contour of a similarly transformed query wordimage in order to locate the specific word in the digitized documents. For the comparison, we use a series of three descriptors, Hu's invariant moments for discarding unlikely similar matches, Shape Context for the contour similarity and the Pearson's correlation coefficient for final pruning of the dissimilar words and additional verification. Comparative results are presented by using instead of Pearson's correlation coefficient the Long-Short Term Memory Neural Network engine of Tesseract Optical Character Recognition system. The described framework due to the simplicity and efficiency that provides, can be applied for massive creation of search indexes and consequently semantic enrichment of Patrologia Graeca. The framework has the potential to be applicable for other printed collections with proper configuration of the parameters. An additional and very significant consequence of our method's effectiveness and simplicity is that it can be used as a pre-stage to provide a large number of word-image and label pairs, These pairs can be used for training neural networks or common classifiers such as k-nearest neighbor or state vector machine.