Underwater Multi-Targets Recognition with ResNet Convolutional Neural Network of SE Block Optimization

Underwater acoustic signals are one of the most effective means of communication in the ocean. In the actual marine environment, the simultaneous existence of multiple targets near the passive sonar is very common, so multi-target recognition is worthy of in-depth research. This paper describes a multi-target signal recognition method using a ResNet convolutional network and optimizes the network by inserting SE blocks. The datasets used include the original audio signal, complex-valued short-time Fourier transform (STFT) spectrum, STFT amplitude spectrum, log-Meier spectrum and Meier frequency cepstrum coefficients. The experimental results show that after optimization, the performance of the network with the amplitude STFT spectrum, complex-valued STFT spectrum and log-Meier spectrum as inputs are improved, and the validation set reaches 88.61%, 95.14%, and 90.66%, respectively, with the most significant improvement of 12.4% when the complex-valued Short-Time Fourier Transform (STFT) spectrum is used as input.