Synthetic Data for the Analysis of Archival Documents: Handwriting Determination

Abstract

Archives contain a wealth of information and are invaluable for historical research. Thanks to digitization, many archives are preserved in a digital format, making it easier to share and access documents from an archive. Handwriting and handwritten notes are commonly found in archives and contain a lot of information that can not be extracted by analyzing documents with Optical Character Recognition (OCR) for printed text. In this paper, we present an approach for determining whether a scan of a document contains handwriting. As a preprocessing step, this approach can help to identify documents that need further analysis with a full recognition pipeline. Our method consists of a deep neural network that classifies whether a document contains handwriting. Our method is designed in such a way that we overcome the most significant challenge when working with archival data, which is the scarcity of annotated training data. To overcome this problem, we introduce a data generation method to successfully train our proposed deep neural network. Our experiments show that our model, trained on synthetic data, can achieve promising results on a real-world dataset from an art-historical archive.