Handwritten and Printed Text Identification in Historical Archival Documents

Abstract

Historical archival records present many challenges for OCR systems to correctly encode their content, due to visual complexity, e.g. mixed printed text and handwritten annotations, paper degradation and faded ink. This paper addresses the problem of automatic identification and separation of handwritten and printed text in historical archival documents, including the creation of an artificial pixel-level annotated dataset and the presentation of a new FCN-based model trained on historical data. Initial test results indicate 18% IoU performance improvement on recognition of printed pixels and 10%IoU performance improvement on recognition of handwritten pixels in synthesised data when compared to the state-of-the-art trained on modern documents. Furthermore, an extrinsic OCR-based evaluation on the printed layer extracted from real historical documents shows 26% performance increase.

Improvements in Handwritten and Printed Text Separation in Historical Archival Documents

“…When microfilming, original records are scaled down and decolourized, which causes information loss and may influence the performance of the text separation system. In [1], we addressed the problem of automatic identification of handwritten and printed text in historical archival documents, in the context of the "Pilotprojekt zur Transformation der Wiedergutmachung" 11 [Pilot Project for Transformation of Reparations]. This project is focused on accessing knowledge hidden in historical records related to the claims of compensation and compensation proceedings that were submitted after 1945, in the state of Baden-Württemberg, Germany.…”

“…The records hold crucial information pertaining to one of the darkest periods in German history, offering first-hand accounts of atrocities committed against individuals who were persecuted and discriminated against on the basis of their ethnicity, religion, political beliefs, and sexual orientation during the National Socialist regime. Our contributions in [1] included creation of WGM-SYN, an artificial pixel-level ground-truth dataset made from historical data, and WGM-MOD, a model trained on this dataset. The data synthesis pipeline in [1] is initiated by denoising and binarising homogenous text regions extracted from original archival documents.…”

“…Our contributions in [1] included creation of WGM-SYN, an artificial pixel-level ground-truth dataset made from historical data, and WGM-MOD, a model trained on this dataset. The data synthesis pipeline in [1] is initiated by denoising and binarising homogenous text regions extracted from original archival documents. This initial step generates an artificial dataset that has been cleansed of noise and document background texture, resulting in information loss when compared to original archival documents.…”

“…In order to make artificial datasets that look more similar to original historical documents, new data synthesis pipelines are proposed and implemented. In this paper, which is an extension of [1], three different data synthesis pipelines for creation of artificial pixel-level annotated datasets are presented. These datasets are created from historical archival documents.…”

“…Furthermore, different models for identification of printed and handwritten text trained on these datasets are introduced. These models are tested and compared with the baselines model from [1], in two different evaluation scenarios. All the artificial datasets and models trained on these datasets are publicly available on GitHub 2 .…”