Improved Feature Pyramid Convolutional Neural Network for Effective Recognition of Music Scores.

Area of Science:

• Computer Science
• Digital Signal Processing
• Music Information Retrieval

Background:

• Sheet music recording is crucial for music communication and cultural inheritance.
• Digital technology aids music storage and distribution but image-based scores present extraction challenges.
• Existing convolutional neural networks struggle with accurate musical score recognition from image data.

Purpose of the Study:

• To develop an improved convolutional neural network (CNN) for accurate musical score recognition.
• To address the limitations of traditional CNNs, such as pixel misclassification and lack of multiscale semantic information.
• To enhance the generalization performance of musical score recognition models for broader applicability.

Main Methods:

• An improved CNN model incorporating a feature pyramid structure was proposed.
• Additional branch paths were utilized to fuse shallow image details, texture features, and high-level global information.
• The model was designed to enrich multiscale semantic information, mitigating recognition performance issues.

Main Results:

• The proposed model demonstrated higher recognition accuracy compared to existing methods.
• Experimental results indicated stronger generalization performance of the improved musical score recognition model.
• The enhanced model effectively alleviated the problem of insufficient multiscale semantic information.

Conclusions:

• The improved CNN with a feature pyramid structure offers superior musical score recognition accuracy.
• The model's enhanced generalization performance allows for application in diverse musical score recognition scenarios.
• This approach provides a more practical and valuable solution for extracting editable music score information from images.