Low-Data Classification of Historical Music Manuscripts: A Few-Shot Learning Approach

TL;DR

A neural-based feature extractor is trained on unlabelled data, enabling effective classification with minimal samples of musical symbols in historical manuscripts, showcasing the potential of AI-driven methods to ensure the survival of historical music for future generations through advanced digital archiving techniques.

Abstract

In this paper, we explore the intersection of technology and cultural preservation by developing a self-supervised learning framework for the classification of musical symbols in historical manuscripts. Optical Music Recognition (OMR) plays a vital role in digitising and preserving musical heritage, but historical documents often lack the labelled data required by traditional methods. We overcome this challenge by training a neural-based feature extractor on unlabelled data, enabling effective classification with minimal samples. Key contributions include optimising crop preprocessing for a self-supervised Convolutional Neural Network and evaluating classification methods, including SVM, multilayer perceptrons, and prototypical networks. Our experiments yield an accuracy of 87.66\%, showcasing the potential of AI-driven methods to ensure the survival of historical music for future generations through advanced digital archiving techniques.

Figures (5)

• Figure 1: A depiction of the musical symbol classification pipeline based on the pipeline by alfaro-contreras_few-shot_2023.
• Figure 2: Examples of the transformations utilised prior to training. From left to right are Random resizing, Random horizontal flip, Colour jitter, Random greyscale, Gaussian blur, Random rotation, Salt and pepper noise, Elastic distortion and Fade.
• Figure 3: Effect of Augmentations on Classification Accuracy ($K=3$). For the prototypical classifier, the $S=1$ and $Q=2$ configuration is shown.
• Figure 4: Confusion matrix showing the classification of a subset of classes which were significantly misclassified for all classifiers. This example used the kNN algorithm with $K=10$ and $A=2$. This is an average over five bootstraps.
• Figure 5: Confusion matrices for (a) Multilayer Perceptron (MLP) and (b) Prototypical classifier, illustrating the classification performance for various musical symbols in historical manuscripts. The matrices highlight the specific symbols that were frequently misclassified, providing insights into the challenges faced by each model in distinguishing visually similar or degraded symbols.