End-to-End Optical Music Recognition Using Neural Networks

This work addresses the Optical Music Recognition (OMR) task in an end-to-end fashion using neural networks. The proposed architecture is based on a Recurrent Convolutional Neural Network topology that takes as input an image of a monophonic score and retrieves a sequence of music symbols as output. In the first stage, a series of convolutional filters are trained to extract meaningful features of the input image, and then a recurrent block models the sequential nature of music. The system is trained using a Connectionist Temporal Classification loss function, which avoids the need for a frame-by-frame alignment between the image and the ground-truth music symbols. Experimentation has been carried on a set of 90,000 synthetic monophonic music scores with more than 50 different possible labels. Results obtained depict classification error rates around 2 % at symbol level, thus proving the potential of the proposed end-to-end architecture for OMR. The source code, dataset, and trained models are publicly released for reproducible research and future comparison purposes.