Objekterkennung in historischen Bilderbüchern

The automatic recognition of objects in images has great potential for the Humanities. The appearance, frequency and dissemination of motifs can be studied without demanding a huge effort from the researchers. Modern computer vision performs amazing tasks, but for historical prints the quality of object recognition does not reach the same level as for realistic photographs. This chapter presents an evaluation of modern methods for books of the 19th century. The results show the heterogeneity of recognition rates for diverse books and objects. Transfer learning for pre-trained systems enhances the performance, but still it reaches only some 60 %. The resulting challenges are discussed in this chapter.