ADAPTIVE DEEP NEURAL NETWORK DESIGN METHOD FOR UNDERWATER ACOUSTIC TARGET RECOGNITION

In recent years, with the amazing achievements of deep learning in the field of computer vision(CV), most researchers have applied it to the field of underwater acoustic target recognition. In order to directly transfer various advanced models in the CV field, researchers chose to use various time-frequency feature extraction methods to turn the ship's radiated noise into three-dimensional data. It still requires effort to design features and requires a large time or frequency range, it also fails to fully utilize the powerful learning ability of deep learning. Based on the limited duration and significant low-frequency effects of ship radiated noise, this paper proposes a Network Design Method of One Dim for Underwater Acoustic Target Recognition (UATR-ND1D), abbreviated as FFT-UATRND1D, which combines fast Fourier transform. Using the entry-level network ResNet as an example using the FFT-UATR-ND1D method, 4320 experiments and 360 experiments were done on two mainstream datasets, ShipsEar and DeepShip, respectively. For the ShipsEar dataset, an extremely lightweight model with only 0.17M parameters and 3.4M FLOPs can achieve an average recognition rate of 97.13% ± 0.43%; When the number of parameters is 2.1M and FLOPs is 5.0M, the optimal level of 98.89% can be achieved. For the DeepShip dataset, an extremely lightweight model with only 0.17M parameters and 6.8M FLOPs is required to achieve an average recognition rate of 95.30% ± 0.28%; When the number of parameters is 2.1M and FLOPs is 13.3M, the optimal level of 98.36% can be achieved. Compared to the methods in existing literature, the methods with parameters similar to this paper have a recognition rate that is more than 3% -5% lower. The papers with recognition rates similar to this paper have parameters and Flops that are at least 1 to 2 orders of magnitude higher than this paper.