Multi-scale spectral feature extraction for underwater acoustic target recognition

Abstract In this work, we proposed a multi-scale spectral (MSS) feature set for underwater acoustic target recognition problem, whose main difficulty lies in the fact that the acoustic signal is often submerged by intense environmental noise. With explicit physical meaning, the proposed MSS feature set fits better with traditional machine learning algorithms, presenting accuracies around 98% for static targets and 90% for moving targets, respectively. The baseline features and algorithm, MFCCs and CNN, show accuracies of only 77.18% and 83.31% for static and moving targets. Thus, the MSS features exhibit “fingerprint” characters of underwater acoustic targets, and are robust in highly noisy and constantly changing environment. MSS features also only need very low computational complexity, thus present great advantages in underwater acoustic target recognition.