Generating Synthetic Data to Allow Learning from a Single Exemplar per Class

Abstract

Recent years have seen an explosion in the volume of historical documents placed online. The individuality of fonts combined with the degradation suffered by century old manuscripts means that Optical Character Recognition Systems do not work well here. As human transcription is prohibitively expensive, recent efforts focused on human/computer cooperative transcription: a human annotates a small fraction of a text to provide labeled data for recognition algorithms. Such a system naturally begs the question of how much data must the human label? In this work we show that we can do well even if the human labels only a single instance from each class. We achieve this good result using two novel observations: we can leverage off a recently introduced parameter-free distance measure, improving it by taking into account the “complexity” of the glyphs being compared; we can estimate this complexity using synthetic but plausible instances made from the single training instance. We demonstrate the utility of our observations on diverse historical manuscripts.