Improving Deep-Learning-based Optical Music Recognition for Camera-based Inputs

Recently, convolutional neural networks (CNNs) have demonstrated promising performance in optical music recognition (OMR). However, the recognition accuracy of an OMR is significantly degraded for a noisy image input given by a camera. To address the problem, this study investigates the use of well-known preprocessed techniques and a state-of-the-art deep learning OMR method. The proposed framework is, then, trained with the Camera-PrIMuS dataset. The experimental results demonstrate that the proposed approach improves the accuracy by 11.1% on average. We conclude that developing OMR algorithms to read camera-based input music scores is plausible and should be further explored.