Few-Shot Deep Learning-Based Detection Method for Sonar Pulse Signals

The active sonar systems of different platforms vary in terms of transmission power, transducer size, and operational mode, leading to distinctive sonar pulse parameters. These pulse parameters serve as essential features for identifying platform types. Detecting and estimating the sonar pulse parameters emitted by targets is crucial for accurate target recognition. This paper proposes a detection and parameter estimation method for subsequent pulses based on the concept of few-shot learning. By utilizing the first few labeled pulses or pulse trains, we adopt a U-Net-based embedded learning framework for frame-level pulse detection and parameter estimation. Experimental results demonstrate that the proposed method achieves a high detection rate even under conditions of low signal-to-noise ratio and interference.