<p>Automatic music score recognition, also known as Optical Music Recognition (OMR), is required for transforming printed or handwritten sheet music to a digital, editable version. Traditional rule-based OMR systems often fail under complex staff layouts, noise, and stylistic variations, resulting in reduced transcription reliability. This research proposes a deep learning framework named Transformer-based Convo Memory Network optimized with Locust Swarm Algorithm (TCMN-LSA) to enhance recognition accuracy, convergence stability, and robustness across diverse music manuscripts. The model is trained and evaluated using the Sheet Music Transformer dataset, which includes the GrandStaff and Camera-GrandStaff subsets containing real-world distortions. Preprocessing improves image consistency using Z-score normalization and Wiener filtering. The TCMN-LSA architecture integrates a CNN feature extractor for spatial symbol detection, a Transformer encoder to model contextual dependencies, and an LSTM module for temporal sequence learning. A Transformer-based decoder produces structured symbolic notation. The Locust Swarm Algorithm (LSA) performs adaptive optimization, improving convergence efficiency and reducing parameter instability. Connectionist Temporal Classification (CTC) enables alignment-free transcription for variable-length polyphonic inputs. Experiments are performed using Python, TensorFlow, and PyTorch, ensuring high-performance training and large-scale batch processing. Evaluation metrics include precision, recall, accuracy, CER, SER, and LER. The proposed model achieves 95.84% of precision, 82.67% of recall, and 88.41% of accuracy, significantly outperforming classical approaches. On the GrandStaff dataset, TCMN-LSA attains 1.62% CER, 2.10% SER, and 6.45% LER, and 1.21% CER_bug, demonstrating improved symbol recognition and transcription stability. These findings confirm TCMN-LSA as a scalable and reliable solution for music archiving, score editing, and educational applications.</p>