Underwater target recognition using convolutional recurrent neural networks with 3-D Mel-spectrogram and data augmentation

Passive recognition of underwater acoustic targets is a hot research issue in acoustic signal processing. The long-term interference of irregular noise in the marine environment caused the relevance of the passive recognition method of underwater targets based on the traditional technical framework to gradually decrease. Due to the interference of irregular noise in the ocean, the passive recognition method used for underwater targets based on the traditional technical framework is gradually becoming less relevant. The feature extraction method that combines deep learning and time-frequency spectrogram can better describe the differences of different targets. In this paper, the proposed model contains three steps to deal with the recognition of underwater targets: feature extraction, data augmentation and deep neural network. For the feature extraction, we use a Mel-spectrogram, as well as the delta and delta-delta features in order to construct 3-D features. In the data augmentation part, we expand the dataset with SpecAugment in the time domain and frequency domain. In deep neural network prediction part, we use the convolutional recurrent neural network (CRNN) for acoustic target recognition. Through a comparison with the ablation test, it is clear that the pipeline in our method is effective in acquiring the recognition result. After evaluating our system through the carrying out of three tasks on the ShipsEar dataset, and the recognition accuracy are 94.6%, 87.5% and 72.6% in task 1, task 2 and task 3 respectively. (C) 2021 Elsevier Ltd. All rights reserved.

Automated summary

Passive recognition of underwater acoustic targets is a hot research issue, and combining deep learning and time-frequency spectrogram for feature extraction, as well as utilizing data augmentation and deep neural network, can effectively enhance the recognition accuracy. Experimental results on the ShipsEar dataset show that the proposed method achieves high recognition performance.