This is the authors' abstract. We don't add key points for this paper.
Optical music recognition requires reliable visual symbol discrimination and sequence-level musical consistency, yet single-model pipelines often trade local accuracy for structural coherence. A fusion framework is developed in which staff-normalized score images are converted into visual-semantic tokens, decoded by a quantized and low-rank-adapted Large Language Model Meta AI (LLaMA) backbone, and refined by adaptive Particle Swarm Optimization (PSO) under rhythmic, staff-position, and harmonic constraints. The interface represents each particle as continuous corrections to token-attribute logits, enabling PSO to optimize a discrete symbolic sequence through a deterministic projection step. On the retained 6:4 split of 7,065 annotated samples, the fusion model achieved 95.7% accuracy for simple scores and 88.3% for complex scores, averaging 92.0%. With 100 PSO iterations and a learning rate of 0.0005, accuracy reached 93.2% with a runtime of 12.5 min. For 200 scores, processing time decreased from 36.2 to 20.1 min, corresponding to a 44.5% efficiency improvement.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Shao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: