Kai Brandenbusch, Eugen Rusakov and Gernot A. Fink
Proc. Int. Conf. on Document Analysis and Recognition, 2021.
Lausanne, Switzerland
With the advent of deep learning in many research areas, the need for very large datasets is emerging. Especially in more specific domains like cuneiform script, annotated data is scarce and costly to obtain by experts only. Therefore, approaches for automatically generating labeled training data are of high interest. In this paper, we present an approach for generating cuneiform signs in larger images of cuneiform tablets. The proposed method allows to control the class of the generated sample as well as the visual appearance by considering context information from surrounding pixels. We evaluate our method on different numbers of cuneiform tablets for training and examine methods for determining the number of training iterations. Besides generating images of promising visual quality, we are able to improve classification performance by augmenting original data with generated samples. Additionally, we demonstrate that our approach is applicable to other domains as well, like digit generation in house number signs.