Best Paper Award for PLANET AI at ICDAR 2021

ICDAR – International Conference on Document Analysis and Recognition – is the most important conference for academia and industry in the field of document analysis and character recognition. The PLANET AI research team participated with three papers and won one Best Paper Award.

In the 6th International Workshop on Historical Document Imaging and Processing (HIP ’21), “Mixed Model OCR Training on Historical Latin Script for Out-of-the-Box Recognition and Fine-tuning” was awarded as best scientific paper.

In cooperation with researchers at the University of Würzburg and LMU Munich, a large corpus of OCR training data was first collected and processed, covering several languages and almost 450 years of printing history. Then, using existing best practices combined with new data-side optimizations, an OCR model was trained. It significantly outperformed existing methods when applied out-of-the-box to unseen material, and also proved to be a valuable starting point for further, more targeted training processes.

ICDAR, which took place in Lausanne (Switzerland) from September 5 to 10, has been a successful venue for the PLANET team several times in the past. In 2019, 2017, and 2015, PLANET AI won six awards in the field of handwriting recognition. Organizations that want to digitize historical documents can benefit from the award-winning technology used in the IDA Suite, PLANET AI’s Intelligent Document Analysis product.

At ICDAR, the paper “Transformer for Handwritten Text Recognition using Bidirectional Post-Decoding” (cooperation with University of Rostock) was presented with a poster. On the one hand, it showed how handwritten texts can be read using complex transformer networks. On the other hand, an improvement method was introduced that could significantly increase accuracy by reading text lines twice – once from left to right and once from right to left.

Likewise, PLANET AI launched the paper “One-Model Ensemble-Learning for Text Recognition of Historical Printings” with an oral talk. In cooperation with the Center for Philology and Digitality “Kallimachos” (ZPD) at the University of Würzburg, a novel method was developed that allows to efficiently train models for unknown fonts with only a few lines of manually generated ground truth.

Basic research is a vital part of PLANET AI’s innovation process. Latest research results always are integrated into the further development of PlanetBrain, the artificial intelligence that can be found in all our products.