Contribution
Comparison of Embedded Hardware Platforms for Optimized Machine Learning-Based Acoustic Imaging
* Presenting author
Abstract:
With emerging possibilities for acoustic imaging based on machine learning instead of grid-bound analytical beamforming methods, noise source characterization can be optimized for both efficiency and accuracy. These methods rely on large machine learning frameworks and therefore require immobile desktop computation units, unsuitable for use as mobile measurement equipment. Attempts to bring these machine learning models to lower-end systems have yet to be made, but different embedded hardware platforms such as single-board computers show potential when used with lightweight inference engines. This work compares two hardware platforms with ML-acceleration units and one without regarding precision, timing, usability, and accuracy of results when used as computation units for acoustic imaging. While the chosen platforms' specifications are comparable, their performance varies more widely than expected. Results show that acoustic imaging with machine learning approaches can be performed on embedded devices with sufficient speed and similar output quality to desktop machines, while costing only a fraction of their price.