Multiresolution Convolutional Neural Network for Underwater Acoustic Target Recognition

Convolution neural network(CNN) has become a popular method for underwater acoustic target recognition because of its advantages. Due to the complexity of underwater acoustic environment, a single-scale CNN shows low robustness for complex broadband target effectively. Therefore, a multi-scale residual convolution neural network for complex broadband underwater acoustic targets is proposed in this paper. In this method, two-dimensional feature map is firstly formed based on auditory feature Mel Cepstrum Coefficient and acoustic information. And then multi-resolution analysis is realized through multi-scale convolution kernel, which improve the robustness of the model to complex broadband targets. The results of verification experiments show that the proposed network achieves recognition accuracy and performs high tolerance to noise interference.