DEMUN: DEtector of MUsic Notation

DEMUN stands simply for DEtector of MUsic Notation. This describes exactly what DEMUN does: it detects whether there is or isn’t music notation on a given input page. Crucially, DEMUN is capable of doing this at very large scales: digital libraries with tens of millions of digitized pages (or more!), with very diverse documents. The users of DEMUN are digital libraries and their technical staff who are tasked with providing and maintaining metadata for their digital library. There is no GUI: just components intended to run servers. DEMUN is not intended for library end-users. “Using DEMUN” means setting it up and running it as part of a library collection processing pipeline. In order for DEMUN to be usable in practice, there are two requirements: Accuracy with extremely imbalanced data, and speed. In order to achieve both the speed and accuracy requirements, DEMUN is decomposed into two stages: a pre-filter (Stage 1), and a main stage with a “clean” filter that includes a notation type classifier as well (Stage 2). The pre-filter (Stage 1) is intended to be run on premises, with low hardware requirements and massive parallelization possible without having to take into account network speed and bandwidth. The “clean” filter (Stage 2) has significantly higher accuracy, but requires a GPU to run quickly. Typically, such computing infrastructure is not easily available for libraries (though if it is, then of course it is possible to take advantage of this and run the second stage locally as well). DEMUN achieves a total false positive rate (FPR) below 0.02 %. On experiments with 4.000.000 images from the collections of the Moravian Library, it outputs roughly 3 sheet music images for every 1 false positive, while the input collection contains some 2.500 non-music pages for every page that contains sheet music. DEMUN thus achieves a nearly 7,500x increase in sheet music concentration, and is capable of discovering musical material at a useful rate. The combined false negative rate (FNR) is estimated at 5 %, which means DEMUN misses only approx. one page in twenty. Its two-stage architecture cuts down processing times to approx. a day per 10.000.000 images. This capability to accurately and quickly (and relatively cheaply!) discover music in large-scale library collections independently on pre-existing metadata is currently unique.