Efficient processing of side-scan sonar images and fast detection of sparse targets in large-scale images

Poor feature representation, confusing background topography, and excessive data volume render detecting sparse targets in large-size acoustic imagery challenging. Especially when conducting real-time processing tasks, accuracy and speed are required to be optimized with limited computational resources. Therefore, this paper proposes an efficient method for real-time side-scan sonar (SSS) image processing and detection of sparse targets in large-scale images. Primarily, an intelligent real-time processing method is proposed for the raw SSS data to acquire high-quality SSS images. Aiming at the characteristics of large-size SSS images and sparse targets, we propose an innovative two-stage inference method: The SSS image slices are pre-classified based on the MobileViTv3-XXS model, and then the optimized detection model of RepVGG+YOLOv5m is employed for target detection of image slices containing targets. Experiments show that real-time preprocessing yields SSS images with an average PSNR of 27.112 and SSIM of 0.816, comparable to the post-processing methods. Meanwhile, it maintains high efficiency and achieves 88.2% mAP, significantly outperforming the slice-only method in detection accuracy and efficiency.特征表征不足、背景地形复杂易混淆，加之数据体量庞大，使得在大尺寸声学影像中检测稀疏目标极具挑战性。尤其在开展实时处理任务时，还需在有限计算资源的约束下兼顾精度与速度的优化。为此，本文提出一种高效方法，用于大尺寸图像中的侧扫声呐（side-scan sonar, SSS）实时图像处理与稀疏目标检测。首先，针对原始侧扫声呐数据，本文提出一种智能实时处理方法以获取高质量的侧扫声呐影像。针对大尺寸侧扫声呐影像与稀疏目标的特性，本文提出一种创新的两阶段推理方法：先基于MobileViTv3-XXS模型对侧扫声呐图像切片进行预分类，随后采用优化后的RepVGG+YOLOv5m检测模型，对包含目标的图像切片开展目标检测。实验结果表明，实时预处理得到的侧扫声呐影像平均峰值信噪比（Peak Signal-to-Noise Ratio, PSNR）为27.112，平均结构相似性指数（Structural Similarity Index, SSIM）为0.816，其效果可媲美传统后处理方法。同时，该方法保持了较高的运算效率，平均精度均值（mean Average Precision, mAP）达88.2%，在检测精度与效率两方面均显著优于仅使用切片的基准检测方法。