Robust underwater noise targets classification using auditory inspired time–frequency analysis

• Published in: Applied acoustics Vol. 78; pp. 68 - 76
• Main Authors: Wang, Shuguang, Zeng, Xiangyang
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 01.04.2014; Elsevier
• Subjects: Acoustics; Bark-wavelet analysis; Exact sciences and technology; Fundamental areas of phenomenology (including applications); Hilbert–Huang transform; Physics; Support vector machine; Targets classification; Underwater noise; Underwater sound; Bark-wavelet analysis; Hilbert–Huang transform; Support vector machine; Underwater noise; Targets classification; Time-frequency analysis; Hilbert―Huang transform; Localization; Underwater acoustics; Signal classification; Target detection
• ISSN: 0003-682X; 1872-910X
• DOI: 10.1016/j.apacoust.2013.11.003

•Bark-wavelet analysis is used for the denoising.•Hilbert–Huang transform is used on feature extraction.•Classification experiments under different SNRs are used for the evaluation.•Presented algorithm has better performances than the baseline system under low SNRs. Underwater noise targets classification has variable applications in many fields. During the long range detection, inevitable environmental noise will decrease the recognition accuracy. Thus, robust classification methods need to be developed. Inspired by human auditory perception, a time–frequency analysis method that combines the Bark-wavelet analysis and Hilbert–Huang transform is presented. By using Bark-wavelet analysis, signals are divided into different sub-bands that correspond to the auditory perception. Then denoising is applied to enhance the analyzed signals. With the help of Hilbert–Huang transform, instantaneous frequencies and amplitudes are extracted. Based on these instantaneous parameters, various features are constructed and compared. Support vector machines are used as the classifier. Recorded underwater noise targets signals are used for the experiments. Various signal-to-noise ratios are simulated through the adding of white Gaussian noise at various levels. Cross-validation procedure was used in the experiments. The results showed that proposed method could achieve better recognition performances under different SNRs comparing to other methods.