Ship Type Classification by Convolutional Neural Networks with Auditory-like Mechanisms

Abstract

Ship type classification with radiated noise helps monitor the noise of shipping around the hydrophone deployment site. This paper introduces a convolutional neural network with several auditory-like mechanisms for ship type classification. The proposed model mainly includes a cochlea model and an auditory center model. In cochlea model, acoustic signal decomposition at basement membrane is implemented by time convolutional layer with auditory filters and dilated convolutions. The transformation of neural patterns at hair cells is modeled by a time frequency conversion layer to extract auditory features. In the auditory center model, auditory features are first selectively emphasized in a supervised manner. Then, spectro-temporal patterns are extracted by deep architecture with multistage auditory mechanisms. The whole model is optimized with an objective function of ship type classification to form the plasticity of the auditory system. The contributions compared with an auditory inspired convolutional neural network include the improvements in dilated convolutions, deep architecture and target layer. The proposed model can extract auditory features from a raw hydrophone signal and identify types of ships under different working conditions. The model achieved a classification accuracy of 87.2% on four ship types and ocean background noise.

Keywords: machine learning; neural network; ship radiated noise; underwater acoustics.

Figure 1

Model structure. The model mainly includes cochlea model and auditory center model. In the time convolutional layer, four colors represent four groups. Dilated convolutions are represented by parallel lines at equal intervals. The time frequency conversion layer includes permute layer and max-pooling layer. At the top of the graph, auditory feature recalibration is implemented by global max-pooling and fully connected layers. Frequency convolutional layers are performed by Inception-ResNet. The final stage includes global average pooling and softmax layer.

Figure 2

Signal decomposition of time convolutional layer. Black line at the bottom is the input signal. Four colors (orange, yellow, green and blue) at the upper and middle parts represent the four groups. Four capsules at the top represent the outputs of 4 groups. Four colors of straight lines at the middle part represent dilated convolutional kernels with a dilated rate of 8, 4, 2 and 1.

Figure 3

Operative condition and classification results analysis. Four rows from top to bottom represent Cargo, Tanker, Passenger ship and Tug, respectively. Histograms are the number of samples under different operative conditions. Yellow and green histograms represent the false negative samples and true positive samples, respectively. Orange lines are recall rates.

Figure 4

Visualization of learned auditory filters and their outputs in the time convolutional layer. Blue lines are learned auditory filters and orange lines are the corresponding output feature maps. The 4 panels from top to bottom represent the dilation rate of 1, 2, 4, and 8, respectively. For each panel, vertical axis represents the amplitude of filter and decomposed signal and horizontal axis represents the time domain sampling points of the filter and decomposed signal.

Figure 5

Comparison of learned spectrogram and gammatone spectrogram. (a) Learned spectrogram. (b) Gammatone spectrogram.

Figure 6

Feature visualization by t-distributed stochastic neighbor embedding (t-SNE). Five thousand samples selected randomly from test data were used to perform the experiments. Dots of different colors represent different types of ships. (a) Output of the whole network. (b) Learned spectrogram. (c) Gammatone spectrogram.