A Deep-Shallow Network for Passive Underwater Target Recognition

• Published in: 2020 IEEE 22nd International Conference on High Performance Computing and Communications; IEEE 18th International Conference on Smart City; IEEE 6th International Conference on Data Science and Systems (HPCC/SmartCity/DSS) pp. 802 - 807
• Main Authors: Song, Gaoyu, Liu, Xingang, Zeng, Xin, Luo, Hengguang, Wang, Dayu, Zhang, Boxuan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2020
• Subjects: Computational modeling; convolutional neural network; Convolutional neural networks; passive underwater target recognition; residual network; Robustness; Sonar; Target recognition; Training; Transfer learning; weighted voting mechanism
• DOI: 10.1109/HPCC-SmartCity-DSS50907.2020.00105

Passive sonar system is an essential method for underwater target recognition. However, it is also a challenging task due to the interference of ambient noise and multiple operative conditions of different targets. In this paper, a weighted voting mechanism based Deep-shallow Network (DSN) is proposed. It utilizes both Low-Frequency Acquisition and Ranging (LOFAR) spectrum and Mel Frequency Cepstrum Coefficient (MFCC) spectrum as input features. Besides, part of the model training process is optimized with the idea of transfer learning. Finally, we proposed a weighted voting mechanism to take advantage of convolutional neural network (CNN) and residual network (ResNet) at the same time so that shallow and deep information could be integrated. Experiments show that the highest recognition accuracy achieves 93.3% which is at least 6.1% higher than traditional machine learning methods.