Music Sheet Classification using Squeeze-and-Excitation ResNet50 Network

Music score classification is an important task in music information retrieval. ResNet networks have achieved great success in image classification and have been applied to music score classification. However, directly using a ResNet pre-trained on natural images may not fully capture the structural information of various music scores. In this work, we propose to improve the music score classification model based on ResNet50 by using Squeeze-and-Excitation ResNet50(SE-ResNet50). We trained and evaluated the proposed SE-ResNet50 model on a dataset containing different types of music scores, including piano scores, numbered musical notation, guitar and ukulele tabs. The results show that compared to the original ResNet50, the SE-ResNet50 model significantly improved the music score classification accuracy to 99.78%, verifying that the SE network structure can effectively enhance the performance of ResNet50 in music score image classification tasks.