A sample augmentation method for simulated mine-like targets based on the imaging mechanism of side-scan sonar

Target detection and recognition in side-scan sonar (SSS) images are often limited by the scarcity of annotated samples, particularly for high-value targets such as mines. To address this problem, this paper proposes a physics-guided SSS sample augmentation framework that combines physics-based simulation with an enhanced generative adversarial network. First, a three-dimensional imaging geometry model, geometric-acoustic ray tracing, and target and seabed scattering models are established to simulate target echoes, acoustic shadows, and seabed backgrounds in accordance with the SSS imaging mechanism. Propagation loss, time-varying gain compensation, and random noise are also incorporated to improve the physical consistency of the simulated SSS images. On this basis, an enhanced GAN integrating a prior physical information extraction module, a background style encoding module, and a multi-scale hybrid attention module is developed. The network preserves the physical characteristics of simulated targets while transferring the grayscale distribution and texture characteristics of real SSS backgrounds without requiring paired training samples. Experimental results show that the proposed method achieves an FID of 165.95, outperforming the evaluated GAN-based comparison methods in terms of distributional similarity to real SSS images. In downstream detection experiments, Dataset3 yields higher mAP@0.5 and mAP@0.5:0.95 than Dataset1 for all six evaluated detectors, including YOLO-based, Transformer-enhanced, two-stage, and non-YOLO single-stage architectures. Seven independent YOLOv12 training runs show relatively stable detection performance across repeated runs. These results support the effectiveness of the proposed framework as a physics-guided augmentation strategy for mine-like SSS target samples under the experimental conditions considered in this study.