毕业生知识库

        海上舰船活动监测对我国维护海洋权益和保障海洋安全具有重要意义，在实施海上缉查、打击违法犯罪活动、应急救援、渔业管理等方面有着广泛应用。基于无线电信号的船舶自动识别系统是舰船活动监测的常用方式，但面对非合作舰船时往往失去预期的作用，因此亟需其他有效手段对其进行准确定位与状态估计。光学遥感卫星影像具有观测范围广、空间分辨率高、时效性强、解译相对容易的优点，利用其进行舰船活动监测很大程度上弥补了舰船自动识别系统的不足，极大地加强了相关部门的监测手段。舰船的位置检测和航向航速估计是舰船监测的重要关注内容，二者的结合既可以了解当前舰船的位置及运动状态，又可以预估舰船未来的位置，对舰船活动规律分析及动向预测具有重要的支撑作用。
       当前基于遥感影像的海上舰船检测和航向航速估计方法距实际的应用需求仍有较大差距。一方面，海上场景的遥感影像具有尺寸大、舰船目标分布极其稀疏等特点，利用深度学习方法对大尺寸遥感影像进行处理时，多采用滑动窗口的方式逐窗口进行检测，检测的速度难以满足海量遥感影像数据及时处理的需求。另一方面，尾迹作为静态遥感影像中运动舰船的主要特征，常作为航向航速估计的重要分析对象，但其形态结构复杂且相应的数据集匮乏，已有研究大多基于仿真数据集，在真实数据上的有效性欠佳。

       针对当前海上舰船检测和航向航速估计面临的问题，本文工作如下：
       1. 为了提高海上舰船检测效率和提升航向估计精度，本文将舰船检测任务和航向估计任务结合起来，在 YOLOv5 的基础上改进提出了旋转检测模型YOLOv5-Rotation。针对舰船检测任务，该模型在自建和公开舰船检测数据集上与同类检测模型对比，精度相当但速度约为对比模型的 2 到 4 倍，同时在自建数据集上舰船平均方向偏差小于 4∘。

       2. 为了进一步加快大尺寸遥感影像海上场景中稀疏舰船目标的检测速度，本文提出了基于分类-检测多任务学习的舰船快速检测方法。该方法通过在检测网络模型的浅层引入分类分支，用以判断窗口中是否存在目标，即实现简单背景窗口的预筛选，降低了整幅遥感影像的冗余计算。在自建大尺寸海上舰船数据集上，该方法与原始检测网络相比，在保持相当精度的同时，检测速度提高了约30%。

       3. 针对舰船航向航速数据集匮乏以及舰船尾迹结构复杂导致航向航速难分析问题，本文构建了一个基于遥感影像的航向航速估计数据集并在此基础上针对开尔文尾迹舰船目标提出了航向航速自动估计算法。在航向估计方面，本文在旋转矩形框的检测结果上结合舰船首尾的纹理分析提出了航向自动估计方法，与传统基于 Radon 变换的舰船航向提取方法进行了对比，证明了所提方法的优越性。在航速估计方面，本文充分分析了舰船航速和开尔文尾迹频域内波高谱的关系，提出了一种航速自动估计方法，并和类似方法进行了对比，证明了本文方法的有效性。

       Maritime ship activity monitoring is of great significance for our country to safeguard maritime rights and interests and guarantee maritime security, and is widely used in implementing maritime seizure, combating illegal and criminal activities, emergency rescue, fisheries management, etc. The automatic ship identification system based on radio signals is a common way for ship activity monitoring, but it often loses its expected function when facing non-cooperative ships, so there is an urgent need for other effective means to accurately locate and estimate their status. Optical remote sensing satellite images have the advantages of wide observation range, high spatial resolution, strong timeliness and relatively easy interpretation, and their use for ship activity monitoring largely makes up for the shortage of ship automatic identification system and greatly strengthens the monitoring means of relevant departments. The combination of the two can not only understand the current position and movement status of the ship, but also predict the future position of the ship, which is an important support for the analysis of the ship's activity pattern and movement prediction.

       The current remote sensing image-based maritime ship detection and heading speed estimation methods are still far from the actual application requirements. On the one hand, the remote sensing images of maritime scenes are characterized by large size and extremely sparse distribution of ship targets. When using deep learning methods to process large size remote sensing images, the detection is mostly carried out window by window by sliding window, and the detection speed can hardly meet the demand of timely processing of massive remote sensing image data. On the other hand, as the main feature of moving ships in static remote sensing images, the wake is often used as an important analysis object for heading speed estimation, but its morphological structure is complex and the corresponding dataset is scarce, and most of the existing researches are based on simulated dataset, which is not effective on real data.

       To address the current problems faced by maritime ship detection and heading speed estimation, the work of this thesis is summarized as follows:

       1. In order to improve the efficiency of ship detection at sea and enhance the accuracy of heading estimation, this thesis combines the ship detection task and heading estimation task, and improves on YOLOv5 by proposing a detection model based on rotating rectangular frame, YOLOv5-Rotation, which has comparable accuracy but about 2 to 4 times faster than similar detection models on self-built and public ship detection datasets, while the average ship orientation deviation on the self-built dataset is less than  4∘.

       2. To further speed up the detection of sparse ship targets in large-size remote sensing images of maritime scenes, this thesis proposes a fast ship detection method based on classification-detection multi-task learning. The method reduces the redundant computation of the whole remote sensing image by introducing classifier branches in the shallow layer of the detection network model to determine whether there are targets in the window, i.e., to achieve pre-screening of simple background windows. On the self-built large-size maritime ship dataset, the method improves the detection speed by about 30% compared with the original detection network while maintaining comparable accuracy.

       3. To address the lack of ship heading speed dataset and the difficulty of analyzing the heading speed due to the complex structure of the ship's wake, this thesis constructs a remote sensing image-based heading speed estimation dataset and proposes an automatic heading speed estimation algorithm for Kelvin wake ship targets based on this dataset. In terms of heading estimation, this thesis proposes an automatic heading estimation algorithm based on the detection results of the rotating rectangular frame combined with the texture analysis of the ship's bow and stern, and compares it with the traditional Radon transform-based ship heading extraction method to prove the superiority of the proposed method. In terms of speed estimation, this thesis fully analyzes the relationship between ship speed and Kelvin wake frequency domain wave height spectrum, proposes an automatic speed estimation method, and compares it with other methods to demonstrate the effectiveness of this method.