运动舰船及尾迹信息的SAR图像正演、反演与智能识别

With fast advancement of high resolution imagery technology, ocean SAR (Synthetic Aperture Radar) becomes more and more sensitive to the details of ship structure and the texture and roughness of sea surface. Meanwhile, SAR image set enters the era of big data. Therefore, automatic classification, recognition and other analyses of vessels become more and more important for both civilian and military applications, such as maritime traffic safety, fisheries control and border surveillance. . However, due to the long synthetic aperture time and various motion induced effects, e.g., smearing, defocussing and geometric distortion, recognition of ships in motion is still challenging for high resolution SAR. This has greatly reduced the advantages of high resolution SAR and becomes a bottleneck in various applications. . Fortunately, the components of ship wake provide valuable information of ship motion, especially under the imaging mechanism of SAR, which, on the other side, can help the refocusing and information retrieval of the smeared images of moving ship. To retrieve the multi-dimensional information within the SAR images of moving ship, a systematic scheme of forward modeling, backward inversion and machine learning is proposed for the ATR (Automatic Target Recognition) of moving ship. . One of the biggest challenges in the machine learning algorithms, especially in the case of SAR, is the lack of training data. To fill this gap, this project will carry out by training using the simulated SAR data set, then make calibration and test by few real SAR images. The aim of this project is to build a complete chain for the ATR of moving ships, from ship and wake modeling, scattering computing, SAR imaging simulation, data processing, to machine learning based ATR and information retrieval.

随着高分辨率成像技术的发展，海洋SAR对舰船结构、海面纹理的敏感度越来越高；各种类型的SAR图像也进入大数据时代，在海洋资源管理、国防军事领域的价值日益提升。.高分SAR的瓶颈问题之一是孔径合成时间长，目标运动导致其图像发生拖尾、模糊以及几何畸变，运动状态下的舰船因此难以反演基本信息，严重削弱了高分SAR的观测优势。在孔径合成机制下，尾迹波系从侧面提供了丰富的舰船运动信息。为“撬开”SAR图像，实现复杂运动舰船的智能识别，项目提出自尾迹流体模型开始，开展运动舰船SAR图像的正演与反演，从正、反两个方向探索舰船与尾迹的成像机制和识别原理。.机器学习最大的困难是缺乏训练样本。项目通过正演构建SAR模拟数据库，训练舰船识别的机器学习算法，再利用少量真实的SAR图像进行测试和矫正。目标是形成完整的技术链，包括舰船与尾迹建模、散射计算、成像模拟、数据处理、以及基于机器学习的舰船信息反演。

随着SAR图像进入大数据时代，运动舰船与尾迹的自动识别变得越来越重要。为实现复杂运动舰船的智能识别，项目自尾迹流体模型开始，从正、反两个方向探索舰船与尾迹的成像机制和识别原理，在舰船和尾迹的SAR成像正演与反演、多自由度运动舰船SAR成像机理、舰船尾迹SAR图像的分解重建与识别等领域做出系统创新，重要成果和贡献包括:.1.全面阐明运动舰船纵摇、横摇和艏摇引起的SAR图像畸变，为多自由度运动舰船成像机理和识别研究奠定基础。该工作应用于某国防重点项目的成像分析和晃动船模造波池成像试验，取得显著成果。.2.舰船尾迹电磁散射和SAR成像模拟。提出6自由度动态尾迹建模方法，由波高场重建尾迹轨道速度场；提出物理光学相位校正双尺度法，计算动态尾迹雷达回波，应用于某重点预研项目成像分析。.3.舰船湍流尾迹的成像模拟非常困难，项目提出基于能量衰减谱和Swanson半经验模型的湍流建模和SAR图像模拟算法，获IEEE国际计算电磁学会议Ulrich L. Rohde最佳论文奖。.4.提出舰船尾迹波系在SAR图像波数域的频谱轨迹，据此实现尾迹波系SAR图像的逐一分解、逐一重建，显著提高舰船运动参数的反演精度。.5.在舰船成像试验中，由实测风速、风向、海浪方向谱数据及船舶图纸、航行数据等，逐层实现全链路的海面舰船SAR成像仿真和数据验证，被选为Int. J. Numer. Model封面论文。.6.在舰船SAR图像智能检测领域，提出将显著性检测与CNN网络相结合，由显著性检测产生高质量舰船图像切片，在无人工干预的情况下自动提取舰船特征，并有效区分舰船与陆地上的强散射目标。该方法与原始的CenterNet网络相比，在SSDD数据集上总体mAP提高了1.11%，特别地，对近海船舶检测的mAP提高了6.71%。. 项目成果最终形成了完整的技术链，包括舰船和尾迹的建模方法、电磁散射计算方法、雷达成像模拟、雷达成像机理分析、水池试验和外海实测SAR数据处理，以及基于机器学习的舰船检测方法。