Segmentation Free Spotting of Cuneiform Using Part Structured Models

Cuneiform scripts constitute an immense source of information about ancient history, dating back almost four thousand years. Documents were written by imprinting wedge-shaped impressions into wet clay tablets, and current scholarly practice typically transcribes the resulting markings by hand with ink on paper. This work develops algorithmic methods for cuneiform script, combining feature extraction for cuneiform wedges with prior work on segmentation-free word spotting using part-structured models. We adapt the inkball model used for word spotting to treat wedge features as individual parts arranged in a tree structure. The geometric relationship between query and target is measured by the energy necessary to deform the tree structure. We also introduce an optimizing method for wedge feature extraction based on optimally assigning tablet structuring elements to hypothesized wedge models. Finally, we evaluate the method on a real-world dataset, and show that it outperforms the state of the art in cuneiform character spotting.