We are developing computer software to recognise handwritten music manuscripts.

Optical recognition of handwritten music scores is a major unresolved problem; difficulties such as the compounding of symbols and occlusion often counteract the advantages of the fixed notation and timing information.

Our approach directly reverses the process of music writing. A composer or engraver would normally write a note head followed by a connected stem and beams, as necessary, and lastly other articulation markings, such as ties and slurs; the algorithm would first pick out long and thin features like slurs and ties, followed by beams, and then stems, in effect breaking the complicated features into a lower-graphical-level primitive before recognition.

We are searching for new approaches and methods to segment, classify, represent, and use domain knowledge within optical document analysis.

The project is supported by the Humanities Research Board of the British Academy in the form of an Institutional Fellowship.