Neural Network Fusion and Selection Techniques for Noise-Efficient Sound Classification
Stelios Α. Mitilineos, Nicolas‐Alexander Tatlas, Stelios M. Potirakis, Maria Rangoussi · Journal of the Audio Engineering Society · 2019
An efficient means for classifying potentially hazardous events using wireless acoustic sensor networks may significantly contribute to the preservation of cultural heritage, artifacts, and architectural sights. However, classification of field-collected sound samples is a demanding task because omnipresent ambient noise severely affects the quality of the recorded samples and the corresponding extracted features. Building on previous work, the authors present a series of fusion or ensemble learning techniques that poll a number of artificial neural network classifiers in order to create class estimates that are significantly more accurate than each isolated classifier or their average. Furthermore, ambient noise effect is simulated by artificially injecting additive white and pink noise to the available sound samples, thus creating a wide range of signal-to-noise (SNR) values. Numerical results demonstrate that the proposed fusion techniques maintain satisfactory accuracy even for negative SNR values, thus demonstrating the applicability of the proposed classification platform for real-world applications.