作为一种主动成像声呐，侧扫声呐由于分辨率高、价格低、能获取连续的水下回波信号并以图像形式进行显示，从而被广泛地应用于水下地貌勘探、环境监测、残骸搜寻、异常目标检测等多个研究领域。声呐目标可分为飞机残骸、沉船、管道等大目标和蛙人、鱼雷等小目标。由于声呐图像原理和光学有所差异，一般的光学图像算法难以直接在声图上应用。且因为水体、海底和电噪声等干扰，声呐回波信号噪声较大，导致在声呐图像中，各类小目标很容易被噪声淹没，对小目标声呐图像处理算法造成了困难。随着神经网络的再度兴起，深度学习算法也被用于声呐图像处理领域。然而，声呐回波信号的获取需要投入大量的人力、物力，导致声呐图像数据集十分缺乏，特别是包含目标的声呐图像数据集，从而阻碍了深度学习方法在声呐图像上的应用。因此，研究声呐回波信号的仿真方法，对深入理解声呐信号特征以及为深度学习算法提供数据支撑具有重要的意义。本文围绕高频声呐的小目标回波信号仿真开展研究。针对现有声学仿真方法在高频下的计算速度慢的问题，本文克服了基于Kirchhoff近似的板块元方法在计算上存在的限制，实现了高频自由场下任意形状目标的理想回波信号仿真。针对实际场景，本文结合波动声学和射线声学理论，提出了一套完整的目标建模仿真方法，其中包含了对于声呐指向性、目标表面粗糙度、阴影渲染和海底混响的建模和仿真。基于本文方法得到的仿真结果在与真实数据的相似性以及仿真速度上，都达到了较好的效果。仿真信号与水池实验得到的真实目标回波信号的在相关性指标NCC上的平均值达到了0.74，优于现有文献中的已知方法。声呐频率等于600 kHz时，半径0.28 m、长度为1.8 m的光滑圆柱单帧回波信号仿真时间小于十分钟。此项工作将为传统的小目标图像处理方法提供新的思路，并有利于推动深度学习算法在声呐图像处理领域的进一步发展。
As a kind of active imaging sonar, side scan sonar is widely used in many research fields such as underwater geomorphological exploration, environmental monitoring, wreckage search, and abnormal target detection. The types of sonar targets include large targets such as aircraft wreckage, sunken ships and pipelines, as well as small targets such as frogmen and torpedoes. Due to the difference between sonar imaging principle and optical imaging, general optical image processing algorithm is difficult to be applied to sonar image directly. Furthermore, because of the interference of water, seabed and electrical noise, the sonar echo signal contains a large amount of noise. As a result, small targets can be easily submerged by noise in the sonar image, which causes difficulties for traditional sonar image processing algorithm. With the resurgence of neural networks, deep learning algorithms are also being utilized in the field of sonar image processing. However, the acquisition of sonar echo signals requires a lot of human and financial resources, resulting in a lack of sonar image data sets, especially those containing targets. This limits the application of deep learning methods in sonar image processing. Therefore, it is of great significance to study the simulation method of sonar images to deeply understand the characteristics of sonar echo signals and provide data support for deep learning algorithms.This paper focuses on the simulation of small target echo signal of high frequency sonar. Aiming at the computational speed limitation of the existing acoustic simulation methods at high frequencies, this paper overcomes the computational limitation of the plate element method based on Kirchhoff approximation, and realizes the ideal echo signal simulation of targets with arbitrary shapes in a high-frequency free field. For the actual scene, this paper proposes a complete set of target modeling and simulation methods based on the theory of wave acoustics and ray acoustics, including the modeling of sonar directivity, target surface roughness, shadow rendering and seabed reverberation. The simulation results obtained based on the method in this paper have achieved good results in the similarity with the real data and the simulation speed. The average value of the correlation index NCC between the simulated signal and the real target echo signal obtained from the pool experiment is 0.74, which is better than the existing methods in the literature. With the sonar frequency of 600 kHz, the simulation time of a single frame echo signal of a smooth cylinder with a radius of 0.28 m and a length of 1.8 m is less than 10 minutes. This work provides a new idea for traditional small target image processing methods, and is conducive to promoting the further development of deep learning algorithms in the field of sonar image processing.
