Anomalous sound detection using sound image and CTF-bilateral filter

This paper introduces an innovative filter algorithm for anomalous sound detection (ASD) inspired by the bilateral filter, which outputs a weighted average of neighboring information at each pixel, smoothing the image and removing anomalies while preserving edges. Our method aims to eliminate background noise in sound signals, retaining only abnormal and prominent points, and improving the detection process to focus on these points’ presence or the emergence of new anomalies. We present the Chroma-Time-Frequency (CTF) Bilateral Filter designed for ASD in non-stationary sound signals. Unlike traditional bilateral filters that use Gaussian chroma and distance factors, our CTF-Bilateral Filter incorporates additional time and frequency factors to better suit acoustic data represented by Log-Mel Spectrograms. To enhance ASD by minimizing the similarity between adjacent pixels, we introduce a Dual Window approach for applying the three weight factors. We further refine this approach with a Flat-Dual Window to better consider the time-frequency information of sound features. The residual image, derived from subtracting the background image from the raw image, undergoes feature processing to highlight prominent points. Experimental validation demonstrates that our method effectively filters out background sound, captures prominent points, improves the detection of subtle anomalies, and exhibits strong generalization capabilities.

Recommended citation: Lu J, Zhang M, Li T. Anomalous sound detection using sound image and CTF-bilateral filter[J]. Neurocomputing, 2025, 637: 130151.
Download Paper