Reproduce “Multi-scale Scanning Network for Machine Anomalous Sound Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yucong Zhang, Juan Liu, Ming Li
A Multi-scale Scanning Network (MSN) was developed to improve machine anomalous sound detection by extracting diverse patterns across varying scales within audio spectrograms. This method achieves competitive performance on DCASE 2020 and 2023 Task 2 datasets, demonstrating effectiveness in handling domain shifts without relying on large-scale pre-training.
Key insights
- Machine sound spectrograms contain consistent yet scale-variant patterns, and explicitly capturing these multi-scale patterns is crucial for robust anomalous sound detection.
- Tailored architectural design for feature extraction, such as the multi-scale scanning approach, can achieve high performance in ASD, comparable to or exceeding methods reliant on extensive pre-training.
- Ablation studies reveal that multi-scale analysis, particularly in the frequency domain, is a more critical factor for learning meaningful machine sound patterns than multi-scale analysis in the time domain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.