The Learned Multi-scale Deep Filters for Underwater Acoustic Target Modeling and Recognition

Underwater acoustic target recognition (UATR) via ship radiated noise faces a big challenge due to the complex ocean environment. On the basis of Auditory system inspired Deep Convolutional Neural Network (ADCNN), an improved ADCNN is proposed to better model and recognize the ship radiated noise. In the improved ADCNN model, deep convolution filter groups are designed to decompose the time domain ship radiated noise into time domain signals with different timescales. Then, the decomposed signals are merged by merging layer and the deep features are extracted by the following full connected layers. Finally, the softmax layer is utilized to obtain the prediction probability for each ship radiated noise sample. In this paper, we expound the decomposition mechanism of deep filters in improved ADCNN taking the view of the time structure of ship radiated noise. Further, the feature fusion method of decomposed signals is improved by weights sharing. In the experiment, the effect of number of deep filter groups is discussed. The experimental results show that setting more filter groups can achieve better classification performance to some extent and that improved ADCNN can achieve better or equivalent recognition results compared to some existing underwater acoustic target recognition methods.