Clustering On-Line Dynamically Constructed Handwritten Music Notation with the Self-organising Feature Map

Abstract

In this paper we consider the problem of recognising handwritten music notation in the context of a pen-based interface. The motivation for the paper stems from current pen-based input technologies that do not achieve true recognition of unconstrained handwritten music. The practical applications of music notation recognition in education, composing, music search tasks and other are obvious, warranting investigation of the problem. This paper explores the self-organising feature map (SOM) as a coarse classifier to categorise pendown movements used by people when writing music notation, so creating a set of person specific ‘primitives’ based on pen strokes. Three different preprocessing methods are used to scale pendown movements and a 5 by 5 SOM is used to cluster the strokes. The stroke clusters form the basis of categories with which a multi-layer perceptron (MLP) could be trained for stroke recognition of pen-movements that comprise handwritten music notation.