Wavelets for Dealing with Super-Imposed Objects in Recognition of Music Notation

Abstract

This chapter is concerned with a problem that arises in Optical Music Recognition (OMR) when notes and other music notation symbols are super-imposed upon stavelines in the music image. We investigate a general-purpose, knowledge-free method of image filtering using two-dimensional wavelets to separate the super-imposed objects. Some background is given to the area of wavelets and a demonstration of how stavelines can be located in a wavelet-filtered image. We also explore the separation of foreground objects (notes) from the background (stavelines) over a variety of image resolutions, in binary and greyscale images using a pixel-based truth representation of the image to evaluate the accuracy with which symbols are identified. We find that the Coifmann family of wavelets appear most suitable for vertical image components, and the Daubechies for the horizontal. The motivation for this chapter stems from the desire to (i) make an original wavelet application in image processing, (ii) provide a fresh (theoretical) perspective on the problem of super-imposed objects in music notation, recognizing the duality of the segregation task that exists with staveline removal/symbol extraction and (iii) evaluate how beneficial wavelet image filtering might be to the OMR process.