海洋舰船检测在国防安全、海洋资源管理、环境保护、海上搜救等多个领域发挥着至关重要的作用。本文采用尾迹检测的方式来实现舰船的间接检测，研究了基于偏振图像融合的舰船尾迹检测方法。本文首先介绍了尾迹检测的研究现状，包括尾迹类型（热尾流、水动力学尾迹、气泡尾迹）和探测手段（SAR成像、高光谱成像、可见光成像、偏振成像等）。接着，本文介绍了偏振图像融合技术，包括其优缺点、传统方法和基于深度学习的方法，综述了目标检测与关键点检测技术。偏振图像处理方面，本文对分焦平面相机拍的到的图片进行去马赛克处理，之后开发了基于DenseNet和Transformer网络的偏振图像融合网络，有效融合偏振图像，以提高舰船尾迹的检测准确性和效率。在舰船目标检测方面，本文通过收集和处理大量多样化的偏振尾迹图像，创建了一个多样化的目标检测数据集，这为算法训练和验证奠定了基础。本文提出了基于可编程梯度信息（PGI）的YOLOv8网络改进网络，有效地从图像中提取尾迹特征，同时保持较高的计算效率，尾迹检测准确率达到90%以上，检测帧率达到了96.9FPS，实现了尾迹目标高精度实时检测。在舰船参数反演方面，本文基于SimCC检测头开发了开尔文尾迹关键点检测网络，相比于主流基于Heatmap的检测算法，模型准确率提升了5%。接着本文基于关键点检测结果，采用开尔文尾迹横波速度反演法进行舰船参数反演，实验结果证明，速度反演误差小于10%，航向反演误差小于2.5%。最后本文基于IMAX6ULL开发板实现了算法硬件部署。本文的研究不仅为舰船尾迹检测提供了新的技术手段，而且通过综合利用偏振图像融合和深度学习技术，为提高尾迹检测的准确率和效率提供了有益的探索。

Marine ship inspection plays a vital role in many fields such as national defense and security, marine resource management, environmental protection, and maritime search and rescue. In this paper, the method of wake detection is used to realize the indirect detection of ships, and the method of ship wake detection based on polarization image fusion is studied.In this paper, the research status of wake detection is introduced, including the types of wake (thermal wake, hydrodynamic wake, bubble wake) and detection methods (SAR imaging, hyperspectral imaging, visible light imaging, polarization imaging, etc.). Then, this paper introduces the polarization image fusion technology, including its advantages and disadvantages, traditional methods and deep learning-based methods, and reviews the object detection and key point detection technologies.In terms of polarization image processing, this paper demosaics the images taken by the subfocal plane camera, and then develops a polarization image fusion network based on DenseNet and Transformer network to effectively fuse polarization images to improve the accuracy and efficiency of ship wake detection.In terms of ship target detection, this paper creates a diverse target detection dataset by collecting and processing a large number of diverse polarization wake images, which lays a foundation for algorithm training and validation. In this paper, we propose an improved YOLOv8 network based on Programmable Gradient Information (PGI), which can effectively extract wake features from the image while maintaining high computational efficiency, with an accuracy rate of more than 90% and a detection frame rate of 96.9FPS, which realizes high-precision real-time detection of wake targets.In terms of ship parameter inversion, this paper develops a Kelvin wake key point detection network based on the SimCC detection head, and the accuracy of the model is improved by 5% compared with the mainstream Heatmap-based detection algorithms. Then, based on the key point detection results, the Kelvin wake shear wave velocity inversion method is used to invert the ship parameters, and the experimental results show that the velocity inversion error is less than 10%, and the course inversion error is less than 2.5%. Finally, this paper implements the hardware deployment of the algorithm based on the IMX6ULL development board.This study not only provides a new technical means for ship wake detection, but also provides a useful exploration for improving the accuracy and efficiency of wake detection through the comprehensive use of polarization image fusion and deep learning technology.