Historic handwritten text recognition – from research to a useful application

Abstract

Handwritten text recognition is a standard machine learning task - with enough annotated training data and some computing power, a capable sequence-to-sequence model can be trained. However, building a useful application is not that simple. In this talk, I will touch on several practical aspects of building the PERO-OCR application and the models behind it. I will give an overview of our research, focusing on the efficient utilization of available data through active learning, self-supervised learning, self-training, domain adaptation, and transcription style adaptation.

Format

The talk will be held online (Zoom link attached) and the planned format is a 30 min talk followed by a 30 min open discussion session. The talk is open to the public so please distribute the announcement to your teams.