Exploratory study of machine learning classification techniques for historical bowed instruments

Details

Supervisors

Glineur, François ; Ceulemans, Anne-Emmanuelle

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enSince the standardization of the violin family in the mid-eighteenth century, numerous instruments have undergone dimensional reductions in order to conform to emerging size conventions. While such reductions have traditionally been identified through qualitative visual inspection by luthiers and musicologists, no fully quantitative and automatic methodology has been established. This thesis explores three machine learning approaches for the identification of violin reduction based on data obtained through photogrammetry on a corpus of historical instruments. The three different approaches investigated are Multiple Kernel Learning (MKL), Convolutional Neural Networks (CNN), and Multi-layer Perceptrons (MLP). The CNN approach constitutes the main focus of this work, while MKL constitutes more of an interpretability extension of the previous SVM attempt (from [BCG25]) and MLP are primarily used as simpler baseline for potential future work. The study discusses their respective strengths, limitations, interpretability, and robustness in the context of an extremely small, imbalanced dataset. The results show that MKL achieves extremely high classification performance, with several configurations reaching up to 100% balanced accuracy. This approach confirms the importance of geometric descriptors derived from contour lines, and highlights the redundancy of the discriminative signal across multiple features. However, it also confirmed a strong sensitivity to hyperparameter tuning, raising concerns regarding model stability and generalization. In contrast, the CNN approach demonstrates that elevation maps contain some useful discriminative information, but fails to achieve high classification performance, with a maximum balanced accuracy of 65.0%. The experiments show that fullresolution inputs introduce significant noise, and that simpler representations, such as resampled or reduced grids, lead to better results. The MLP approach confirms this observation, achieving slightly higher performance (up to 68.0%) with lower computational cost, suggesting that the essential discriminative information is already captured by more compact geometric descriptors. Overall, this work highlights the trade-off between interpretability, performance, and stability across different modeling strategies. While kernel-based methods achieve the highest performance, their instability may limit their reliability. Neural networks approaches, although more flexible, remain constrained by the small dataset and the weak and subtle discriminative signal. These findings suggest that future work should focus on enlarging the dataset, improving input representations, and maybe developing hybrid approaches that combine the interpretability of geometric descriptors with the flexibility of learning-based models.