Effective feature fusion via analysis of quantitative similarity matrices among various acoustic features for underwater active target detection

Abstract

For underwater active target detection, using conventional machine learning technique is limited and unsuitable because of small training data samples and diversity of environment. Therefore, to apply conventional machine learning for underwater acoustics target detection, three methodologies can be manipulated (1) feature, (2) architecture, and (3) learning strategy. In this paper, we implement various acoustic features in terms of feature similarity and feature fusion. From numerous studies in field of acoustics, various acoustic feature extraction methods have been proposed such as Mel-frequency cepstral coefficient, Gammatone-frequency cepstral coefficient, cepstral coefficient, short-time Fourier transform, constant Q transform, and wavelet packet decomposition. In this paper, we calculate a quantitative similarity between acoustic features by interpreting their data distributions with the corresponding probability densities in a reduced dimension. Furthermore, we fuse the acoustic features by simple concatenation. Fusion of a strongly correlated two-dimensional features tends to follow the performance of poor one, whereas the fusion of weakly correlated features improves performance remarkably. The performance improvement by the fusion of weakly correlated features is attributable to complementing acoustic information each other.