Integrating knowledge components for writer identification in a digital archive of historical music scores

In our work we consider two different approaches to map documents into the feature space. On one hand, we use a semi-automatic, knowledge-based procedure to let musicology experts determine the set of feature values for a document manually. On the other hand, we plan the integration of an automatic approach for feature extraction based on image processing techniques. However, currently we deal only with the results from the manual mapping implementation. We use the features extracted from the collection of music scores, and the information in the distance matrices to cluster the scores according to their handwriting characteristics. In the best case, a cluster represents exactly one writer. For the clustering of the feature sets we use the k-nearest neighbor method. The distance between two feature sets, also referred to as "feature vectors", is derived using a normalized, weighted Hamming distance function. The Hamming distance returned better results than the Euclidean and other higher order distance functions, which were tested.