SPNet: Automated object detection of antinode regions in oscillating steelpan drums

Abstract

We present a machine learning computer program which applies a neural-network-based object detection algorithm to individual frames of high-speed video of oscillating Caribbean steelpan drums obtained via electronic speckle pattern interferometry (ESPI). This algorithm is trained in a supervised learning context on a dataset of crowd sourced human-annotated images obtained from the Zooniverse Steelpan Vibrations project. The computer code, which we call “SPNet,” is subsequently able to annotate new frames of similar images in a manner consistent with the humans' prior work, albeit much more quickly—hundreds of frames per second. The goal of this annotation work is to better understand the dynamical behavior of these drums such as the coupling of oscillations in different parts of the drum surface, by tracking the development of sympathetic vibration modes and extracting their relevant physics. We present details of the algorithm, performance metrics, and some preliminary physics results.