Underwater Target Detection and Localization with Feature Map and CNN-Based Classification

Abstract

The purpose of this paper is to apply the acoustic features, Mel Frequency Cepstral Coefficient (MFCC) and Gammatone Frequency Cepstral Coefficient (GFCC), to underwater signal classification. Underwater acoustic signals are vibration signals, and their characteristics are similar to speech signals. The auditory feature extraction method in speech recognition can also be applied to the underwater environment. For underwater communication, we simulate two models designed for underwater target detection and localization. One is the deterministic model, which is considered as basic model; the other is to combine the deterministic model and statistic model, which is called combined model. The geometric channel model facilitates the generation of the database for different geometric settings. The database is generated by adjusting the parameters of the underwater environment. The classifier adopts a convolutional neural network (CNN). The input to the CNN is the feature maps after feature extraction. We choose continuous wavelet transform (CWT) and short-time Fourier transform (STFT) for comparison. Experiments show the effectiveness of the system architecture and superiority of the proposed algorithm in underwater signal classification and target localization.