TLDR: Researchers developed a novel digital tool using 3D geometric meshes and contour line analysis to identify whether historical violins have been reduced in size. By fitting contour lines with a parabola-like curve and analyzing parameters, particularly the ‘opening’ parameter (beta), they successfully classified reduced instruments using Support Vector Machines, offering a quantitative and objective method for musicological study.
For centuries, violins have been cherished instruments, evolving from their late 16th-century Italian origins to become standardized around 1750 for European orchestras and conservatories. This standardization led to a fascinating historical practice: the reduction of instruments that fell between two size standards. Luthiers would skillfully reduce these violins to a smaller size, a process that subtly but significantly altered their characteristics, particularly their internal contour lines. Traditionally, identifying these reductions has relied on empirical observations, with reduced instruments often showing ‘V’-shaped contour lines compared to the ‘U’-shaped lines of non-reduced ones. However, a new research paper introduces a quantitative approach to this long-standing musicological question.
The paper, titled “Identification of Violin Reduction via Contour Lines Classification,” by Philémon Beghin, Anne-Emmanuelle Ceulemans, and François Glineur, presents a novel tool designed to classify violin contour lines, thereby distinguishing between reduced and non-reduced instruments. This research addresses a critical need in musicology and has implications for the cultural and economic aspects of historical instruments.
A Quantitative Approach to Violin Geometry
The researchers embarked on a quantitative study using a corpus of 25 violins and violas. Their methodology involved acquiring 3D geometric meshes of these instruments using photogrammetry, a technique that creates precise 3D models from photographs. From these meshes, contour lines were sampled at regular intervals, typically every millimeter, across 10 to 20 levels of the instrument’s soundboard.
Each sampled contour line was then meticulously fitted with a parabola-like curve, represented by the equation y = α|x/λ|^β + γ. This equation uses several parameters to describe the curve’s shape:
- α (alpha): Controls the vertical stretch or compression of the curve.
- β (beta): This crucial exponent is associated with the sharpness or “opening” of the curve. Values smaller than two indicate sharper curves, while larger values suggest a flatter shape.
- γ (gamma) and δ (delta): Represent vertical and horizontal translations, respectively.
- λ (lambda): A normalization term related to the width of the contour, ensuring consistency across different widths.
The researchers found that the ‘beta’ parameter was particularly predictive in distinguishing between reduced and non-reduced instruments, aligning with empirical observations that reduced violins exhibit sharper, more ‘V’-like contours.
From Contour Lines to Classification
To process the extracted contour line data, the team developed a robust classification tool. They employed Support Vector Machines (SVM), a supervised learning algorithm well-suited for classification tasks. In simple terms, an SVM works by finding the optimal “hyperplane” – a decision boundary – that best separates data points belonging to different classes (in this case, reduced vs. non-reduced violins) in a high-dimensional space. The algorithm aims to maximize the “margin,” which is the distance between the data points closest to the decision boundary (known as “support vectors”) and the hyperplane itself.
A key challenge was the varying number of contour levels per instrument. To address this, the data was standardized by resampling each parameter vector (α, β, γ, δ) to 50 equally spaced points, effectively creating a uniform “numerical profile” for each instrument. The researchers experimented with different SVM configurations, including linear and Radial Basis Function (RBF) kernels, and a regularization parameter (C) that controls the trade-off between maximizing the margin and minimizing classification errors.
The study highlighted the importance of a robust method for selecting the regularization parameter C, as its choice significantly impacted the model’s stability and performance. They devised an automated procedure using a leave-one-out cross-validation approach, which involves training the model on all but one instrument and then testing it on the held-out instrument, repeating this for every instrument in the corpus. This rigorous validation ensured a reliable assessment of the classifier’s accuracy, particularly using “balanced accuracy” to account for the imbalance in their dataset (20 non-reduced vs. 5 reduced instruments).
Also Read:
- AI Unlocks Faster Analysis of 3D Object Shapes
- Enhancing Musical Onset Detection in Maracatu Through Adaptive Transfer Learning
Promising Results and Future Directions
The results were highly encouraging, with several combinations of features achieving remarkable classification performance, some even reaching 100% balanced accuracy. Notably, the ‘beta’ profile and features derived from it (such as linear, quadratic, and piecewise linear fittings, or counts/proportions of beta values above/below certain thresholds) proved to be very informative. The resampled alpha and beta profiles also yielded excellent results, suggesting that even a smaller set of derived features can effectively classify instruments.
The researchers recommend using a linear kernel for the SVM model due to its greater interpretability and stability compared to the RBF kernel, despite its higher computational cost. This pioneering work offers a completely new, quantitative, and objective method for identifying violin reduction, moving beyond traditional visual and tactile examinations.
Future work aims to expand this analysis to violin backs (the lower plate of the sound box) and incorporate data from cellos to strengthen the conclusions. The team also plans to explore other classification models, such as decision trees or neural networks, and investigate additional geometric descriptors and non-geometric characteristics like surface texture data to further enhance the classification tool. This research marks a significant step forward in applying digital tools to the study of historical musical instruments. You can read the full paper here.
