Illumination-based Augmentation for Cuneiform Deep Neural Sign Classification

Abstract

Automated content-based search for arbitrary cuneiform signs in photographic reproductions is a challenging task in the analysis of ancient documents, a central component of which is a reliable cuneiform sign classification. We present an illumination-based approach to generate synthetic training data for cuneiform sign classification via deep neural networks to overcome common issues with the transferability of machine learning training results. Starting from an analysis of the negative impact of illumination variations in the processed cuneiform data, we employ an illumination augmentation to two-dimensional (2D) training data generated from annotated 3D datasets. We demonstrate that our method is able to overcome the high visual variance of most digitized 2D cuneiform reproductions and achieve an illumination invariant generalization. The effectiveness of our approach is evaluated by its successful application to several subsets of a cuneiform script dataset with an originally poor transferability of mutual training results. Furthermore, we show that a sufficient sampling of the illumination space mostly removes the necessity to match the training data to specific target illumination conditions. The practical applicability of our approach is validated by applying it to a larger dataset, raising the overall classification accuracy by 4 percentage points to 90%, resulting in a classification error reduction of 28.5% when compared to results without the proposed data augmentation.