Performance Analysis of Active Sonar Target Recognition Based on Deep Learning

Active sonar systems play a pivotal role in underwater target classification and identification. Traditional methods relying on manual feature extraction struggle with the complexity of ocean environments and diverse target structures. Recent advancements in deep learning (DL) offer promising solutions. This study systematically explores DL techniques for intelligent active sonar target recognition, analyzing the impact of preprocessing methods, model complexity, signal-to-noise ratio (SNR), and sample size on performance. Using simulated data of underwater elastic spherical shells and vehicle targets, we evaluate preprocessing methods such as Short-Time Fourier Transform, Wigner-Ville Distribution, and Continuous Wavelet Transform, and compare convolutional residual networks (ResNet) and Transformer-based architectures. Results show that transformed-domain echoes outperform raw time-domain data, with recognition accuracy improving as SNR and sample size increase. We introduce the inflection point SNR (IP-SNR) as a critical threshold: below IP-SNR, performance improves rapidly with higher SNR but saturates with sample size; above IP-SNR, SNR and sample size synergistically enhance accuracy toward the theoretical limit. These insights advance the design and optimization of DL-based active target recognition systems for real-world sonar applications.