Recognition of ships based on vector sensor and bidirectional long short-term memory networks

Abstract

The previous methods of target recognition use feature parameters as the system input. Especially in a better acoustic field, a higher recognition rate can be obtained under the condition of obvious target features. But actually, the acoustic field is complex due to the multipath effect, the time-varying channel characteristics, the variety and complexity of environmental background noise and other factors, which leads to a difficulty in feature extraction. In this paper, a method named bi-LSTM is described and firstly used in the field of underwater acoustic to recognize ships without pre-extracting feature based on vector sensor. Then we compared bi-LSTM and LSTM with SVM and the impact on the recognition rate with some parameters. Experimental results of sea trial are given to demonstrate the robust adaptation. The accuracy of both bi-LSTM and LSTM are higher than SVM and the accuracy of bi-LSTM is higher than LSTM with a long training time. The recognition result of testing set can be received in few seconds which proves the real-time of the method.