Self-training and Active Learning with Pseudo-relevance Feedback for Handwriting Detection in Historical Print

Handwritten text recognition research largely focuses on entirely handwritten documents, yet many bibliographic researchers are interested in handwriting left by readers in historical print. Due to the sparse and inconsistent appearance of handwritten annotations, compiling sufficient datasets of handwriting in print can be difficult. We propose a method for utilizing visual similarities among text exemplars for improving handwriting detection in historical print. We investigate the effect of pseudo-labeled page images on improving object detection model performance for handwriting localization across multiple exemplars of Shakespeare’s First Folio. We compare differences in self-training and active learning with pseudo-labels for positive and negative-sample images, using pseudo-relevance and relevance feedback as selection methods. We find that pseudo-labels from positive and negative-sample images improve detection task performance on individual exemplars with an average precision increase of 15%. Tests on collections of multiple exemplars are less conclusive. We discuss how variations in historical print’s materiality may explain these results and outline further research to investigate this matter.