Composer Style Classification of Piano Sheet Music Images Using Language Model Pretraining

• DOI: 10.48550/arxiv.2007.14587
• Publication Date: 2020-01-01
• Authors: T. J. Tsai; Kevin Ji
• Keywords: Sheet music

This paper studies composer style classification of piano sheet music images. Previous approaches to the task have been limited by a scarcity data. We address this issue in two ways: (1) we recast problem be based on raw images rather than symbolic format, and (2) propose an approach that can trained unlabeled Our first converts image into sequence musical "words" bootleg feature representation, then feeds text classifier. show it is possible significantly improve classifier performance training language model set data, initializing with pretrained weights, finetuning small amount labeled train AWD-LSTM, GPT-2, RoBERTa models all IMSLP. find transformer-based architectures outperform CNN LSTM models, pretraining boosts accuracy for GPT-2 from 46\% 70\% 9-way task. The also used as extractor projects space characterizes compositional style.