Fink, Gernot A.Hammerla, Nils Y.Plötz, ThomasVajda, Szilárd2011-01-122011-01-122011-01-12http://hdl.handle.net/2003/27556http://dx.doi.org/10.17877/DE290R-8202Statistical modelling techniques for automatic reading systems substantially rely on the availability of compact and meaningful feature representations. State-of-the-art feature extraction for offline handwriting recognition is usually based on heuristic approaches that describe either basic geometric properties or statistical distributions of raw pixel values. Working well on average, still fundamental insights into the nature of handwriting are desired. In this paper we present a novel approach for the automatic extraction of appearance-based representations of offline handwriting data. Given the framework of deep belief networks -- Restricted Boltzmann Machines -- a two-stage method for feature learning and optimization is developed. Given two standard corpora of both Arabic and Roman handwriting data it is demonstrated across script boundaries, that automatically learned features achieve recognition results comparable to state-of-the-art handcrafted features. Given these promising results the potential of feature learning for future reading systems is discussed.enArabic/Roman handwriting recognitionautomatic feature extractionHidden Markov Modelsnon-linear regularized NCARestricted Boltzmann Machines004Text