Recognition of Underwater Acoustic Radar Signals Based on Multiresolution and Dense Convolutional Neural Network

Abstract

Recognizing underwater objects based on radiated noise information is one of the most crucial issues in underwater acoustics. Underwater acoustic target signals are altered by elements such as the undersea environment and the ship's operational circumstances; hence, generalizing the recognition model is crucial. Most conventional Machine Learning (ML) algorithms often encounter difficulties when dealing with the costly recognition model for massive data analysis. However, Convolutional Neural Networks (CNNs) can automatically extract features for precise categorization. DenseNet is a powerful CNN network, but it has a data duplication problem, so in this paper, an approach using multi-resolution with a dense CNN model for underwater acoustic radar signal detection is proposed to overcome the DensNet problem. At first, the wavelet decomposition with different levels is applied to the input signal to represent the suitable data. The decomposed signals are inputs to the dense CNN. Our detection approach beats other CNN models and achieves an overall accuracy of 99.5% at 0 dB SNR based on experimental findings evaluated on a real-world passive sonar data set.

Keywords

• Convolution Neural Network
• Deep Learning
• Underwater Acoustic Signal Detection
• Multiresolution