Illumination-based Augmentation for Cuneiform Deep Neural Sign Classification

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