An Experimental Comparison between Deep Learning and Classical Machine Learning Approaches for Writer Identification in Medieval Documents

Abstract

In the framework of palaeography, the availability of both effective image analysis algorithms, and high-quality digital images has favored the development of new applications for the study of ancient manuscripts and has provided new tools for decision-making support systems. The quality of the results provided by such applications, however, is strongly influenced by the selection of effective features, which should be able to capture the distinctive aspects to which the paleography expert is interested in. This process is very difficult to generalize due to the enormous variability in the type of ancient documents, produced in different historical periods with different languages and styles. The effect is that it is very difficult to define standard techniques that are general enough to be effectively used in any case, and this is the reason why ad-hoc systems, generally designed according to paleographers’ suggestions, have been designed for the analysis of ancient manuscripts. In recent years, there has been a growing scientific interest in the use of techniques based on deep learning (DL) for the automatic processing of ancient documents. This interest is not only due to their capability of designing high-performance pattern recognition systems, but also to their ability of automatically extracting features from raw data, without using any a priori knowledge. Moving from these considerations, the aim of this study is to verify if DL-based approaches may actually represent a general methodology for automatically designing machine learning systems for palaeography applications. To this purpose, we compared the performance of a DL-based approach with that of a “classical” machine learning one, in a particularly unfavorable case for DL, namely that of highly standardized schools. The rationale of this choice is to compare the obtainable results even when context information is present and discriminating: this information is ignored by DL approaches, while it is used by machine learning methods, making the comparison more significant. The experimental results refer to the use of a large sets of digital images extracted from an entire 12th-century Bibles, the “Avila Bible”. This manuscript, produced by several scribes who worked in different periods and in different places, represents a severe test bed to evaluate the efficiency of scribe identification systems.