A Sequence-to-Sequence Framework Based on Transformer With Masked Language Model for Optical Music Recognition

Optical music recognition technology is of great significance in the development digital music. In recent years, convolutional recurrent neural network framework with connectionist temporal classification has been used recognition. However, its loss function calculated serial mode, which leads to low efficiency training and difficulty convergence. Additionally, because gradient disappearance excessive long sequences, existing models are hard learn relationships between musical symbols, resulting high sequence error rate. Therefore, we propose a sequence-to-sequence based on transformer masked language model deal these problems. The context representation symbols can be captured further by self-attention module transformer, will reduce addition, refer design mask matrix predict each symbol parallel way, so as speed up process. Our experiments carried out printed images stave dataset, results show that our proposed method training-efficient improvement accuracy