A Preliminary Study of Few-shot Learning for Layout Analysis of Music Scores

Abstract

Few-shot techniques offer a promising avenue to reduce the high demand for annotated data required by current machine learning-based applications, such as Optical Music Recognition (OMR). This is a field dedicated to the automatic transcription of music notation from sheet music images. Traditional OMR systems strongly depend on layout analysis, a crucial step involving the identification and segmentation of several components within a music score, such as staff lines, text, or notes. The standard approach requires extensive fully annotated training data, which are resource-intensive and time-consuming to label and curate by domain experts. We present a preliminary study on the use of few-shot learning to alleviate the disadvantages associated with manual annotations. The proposal minimizes the human effort required by employing only partial annotations. For this, we introduce an oversampling technique to train models using a limited set of annotated patches extracted from the score images. Our experimental findings, conducted on four benchmark datasets, underscore the efficacy of the proposed patch extraction. Despite operating with a reduced amount of annotated data, our method achieves performance levels competitive with models trained on the complete dataset. This work points out the potential of few-shot learning in the context of layout analysis for music scores, offering the promise of more efficient and accessible OMR systems.