Character Retrieval of Vectorized Cuneiform Script Bartosz Bogacz Interdisciplinary Center for Scientific Computing (IWR) Heidelberg University, Germany bartosz.bogacz@iwr.uni-heidelberg.de Michael Gertz Institute of Computer Science Heidelberg University, Germany gertz@informatik.uni-heidelberg.de Hubert Mara Interdisciplinary Center for Scientific Computing (IWR) Heidelberg University, Germany hubert.mara@iwr.uni-heidelberg.de Abstract—Motivated by the increasing demand for comput- erized analysis of documents within the Digital Humanities we present an approach to automating handwritten cuneiform character recognition on vectorized cuneiform tablets. Cuneiform is one of the oldest handwritten scripts used for more than three millennia. In previous work we have shown how to extract vector drawings from 3D-models of cuneiform tablets similar to those manually drawn over digital photographs. We approach the problem of recognizing these characters by applying pattern matching against the basic structural features of cuneiform, the wedge-shaped impressions. Then, we find an optimal assignment between the wedge configuration of two characters w.r.t. wedge shape and position. The similarity of two characters is measured by the quality of the assignment. We compare our method against well known methods for handwritten character recognition with favorable results for our method. I. I NTRODUCTION Cuneiform tablets are one of oldest textual artifacts com- parable in extent to texts written in Latin or ancient Greek. Since those tablets were used in all of the ancient Near East for over three thousand years [1], many research questions can be answered regarding the development of religion, politics, science, trade, and climate change [2]. These tablets were formed from clay and written on by impressing a rectangular stylus [3]. This results in wedge-shaped impressions in the clay tablets named after cunei – the Latin word for wedges. There is an increasing demand in the Digital Human- ities domain for handwriting recognition focusing on his- toric documents. Even the recognition of ancient characters sharing shapes with their modern counterparts, e.g., ancient Chinese Sutra [4] is a challenging task. For digitally processing cuneiform script, there exist only a few recent related ap- proaches like proposed by Fisseler et al. in [5] using geometric features of cuneiform tablets acquired with a 3D-scanner [6]. However, with the aim of building a search tool for cuneiform tablets we have to consider the complexity of cuneiform characters in their de-facto standardized 2D- representation. Figure 1 shows a photograph of a cuneiform tablet and its drawing. The representation of cuneiform tablets are sets of wedges, where each wedge can be roughly de- scribed by a Y-shape. These sets of shapes are either traced manually on photographs or computed automatically from 3D- models [7]. In previous work we have experimented with a graph-based representation of cuneiform characters [8]. Most work on character recognition and word-spotting focuses on raster images [10]–[12] and words written with ink (a) (b) Fig. 1. Cuneiform tablet No. TCH92, G127 [9]: (a) Photograph and (b) its drawing. on paper. These approaches make use of the sequential nature of written script. Rath et al. show in [13] that word recognition for handwrit- ten text can be performed by comparing word-profiles encoded as time-signals using Dynamic Time Warping (DTW). The query word (the prototype) and the words in the document (the candidates) are first transformed into three word-profiles. A top-profile, a bottom-profile, and a profile of pixel transitions for each image column. Then, the three word profiles of the candidate are warped to match the prototype. The dissimilarity of two words is measured by the amount of warping necessary to match these words. This approach is extended by Kennard et al. in [14]. An additional step of warping is applied using a regular grid on top of the warping done by the DTW method. Grid nodes are moved horizontally and vertically to match the appearance of the candidate to the prototype. The similarity of the words is measured by the summed difference of their distance transformed medial axis representations. Another approach, which also exploits the wide shape of Latin words, is presented by Rodriguez-Serrano et al. in [15]. Word images are modeled as Hidden Markov Models (HMMs) and split into either character slices [16], sub-character slices [17] or graph slices [18]. Then, a classifier in the form of a probabilistic state-machine is trained to recognize the succession of such slices in a document, a line or a candidate word. Using HMMs to model words can be regarded as a generalization of the DTW method since HMMs are capable of emitting character slices repeatedly and out-of-order [19].