Fig. 1. Schematic diagram of MF-RC-BiSRU

Fig. 2. Schematic diagram of residual structure

Fig. 3. Schematic diagram of multi-scale feature fusion

Fig. 4. Structure of SRU

Fig. 5. Schematic diagram of BiSRU

Fig. 6. Difficulties of note recognition in music score

Fig. 7. Three methods of data processing to simulate unsatisfactory music image. (a) Original incipit; (b) incipit of white Gaussian noise added; (c) incipit of Perlin noise added; (d) incipit of elastic transformations added

Fig. 8. Comparison of training loss and accuracy for C-BiLSTM and RC-BiLSTM networks. (a) Comparison of training loss; (b) comparison of symbol error rate

Fig. 9. Comparison of features in different convolution layers. (a) Original incipit; (b) shallow feature map C₁; (c) deeper feature map C₃; (d) deepest feature map C₅; (e) multi-scale feature fusion map F₄

Fig. 10. Comparison of the symbol error rates in the different networks

Fig. 11. Comparison of MF-RC-BiSRU and MF-RC-BiLSTM. (a) Comparison of training loss; (b) comparison of symbol error rates

Fig. 12. Test results of the same incipit in four different networks.(a) Original incipit; (b) C-BiLSTM; (c) RC-BiLSTM; (d) MF-RC-BiLSTM; (e) MF-RC-BiSRU

Fig. 13. Comparison of loss in different methods