One-Step Detection of Background, Staff Lines, and Symbols in Medieval Music Manuscripts with Convolutional Neural Networks

One of the most complex stages of optical music recognition workflows is the detection and isolation of musical symbols. Traditionally, this goal is achieved by performing preprocesses of binarization and staff-line removal. However, these are commonly performed using heuristics that do not generalize widely when applied to different types of documents such as medieval scores. In this paper we propose an effective and generalizable approach to address this problem in one step. Our proposal classifies each pixel of the image among background, staff lines, and symbols using supervised learning techniques, namely convolutional neural networks. Experiments on a set of medieval music pages proved that the proposed approach is very accurate, achieving a performance upwards of 90% and outperforming common ensembles of binarization and staffline removal algorithms.