Ship Detection And Motion Monitoring Of Remote Sensing Images Based On Deep Learning

Remote sensing technology is an important means for the normal monitoring of key areas at home and abroad,which is because of its large observation area,cyclical revisitability and lack of national border restrictions.As an important carrier of maritime transport and a special military force,ships play a very important role in promoting economic development and safeguarding national rights and interests.Using the unique advantages of remote sensing technology,detection,recognition and motion monitoring of ship targets,whether in the civilian aspects of fishery management,shipping safety,crash rescue,or in the military aspects of the getting deployment of troops,grasping the dynamic conditions of ship formations,and assessing the effectiveness of wartime maritime strikes,are of great practical significance.In recent years,with the increasing attention of countries in the world to remote sensing technology,a large number of professional-level remote sensing satellites have been launched or stepped up to develop,geostationary orbit satellites which have wide observation coverage and high temporal resolution,can achieve long-term continuous detection,low-orbit satellites can obtain sub-meter-level high spatial resolution images,video satellites can obtain color dynamic video of the same level of resolution.How to comprehensively use various types of remote sensing data to accurately,quickly and stably detect and recognize ship targets in massive data has become one of the current research hotspots.Aiming at the problems of ship object detection,recognition and motion monitoring in visible light remote sensing images,this paper mainly conducts research and exploration work on the construction of deep learning object detection network and the improvements of related algorithm,ship target fine-grained classification,and interframe data correlation based on time series images.The main contents of the study are summarized as follows:（1）With the development of remote sensing technology,the amount of remote sensing data is increased exponentially and the image complexity is continuously improved,the necessity of introducing deep learning is analyzed.The advantages and disadvantages of different remote sensing data such as visible light,SAR,infrared and multi/hyperspectral are summarized,and the characteristics of visible light images and ship targets are analyzed in depth.The current status of ship target detection and motion tracking is systematically sorted out,and common algorithms for target detection based on deep learning are analysed and compared.（2）According to the characteristics of ship target rotation,striping and aggregation in remote sensing images,analyze the shortcomings of the horizontal rectangular box detection algorithm commonly used in natural scene images,and then propose a rotation detection box that is more conducive to the specific scene and target.And on the basis of clarifying the angle information by the long-sided definition method,the circular smooth label is used to evolve the angle information from the regression problem to the classification problem,so as to solve the problem of loss mutation caused by the periodicity of the angle.On the HRSC2016 dataset,the three-level finegrained classification of ship target-category-type was realized,and compared with the original horizontal rectangular box algorithm,Recall and m AP were improved by 7.6%and 5.3%,and the improvement is effective.（3）Analyze the development status and future trend of remote sensing technology,comprehensively use the advantages and characteristics of high and low orbit satellites,and design a method to detect and track the target of multi-motion ships at sea by using multi-source remote sensing satellite data,which can accurately locate the target candidate area and reduce the target search scope.On the basis of analyzing the key technologies of Deep SORT target tracking algorithm,the tracking and monitoring of ship targets in remote sensing satellite video is completed,and relevant simulation experiments are carried out to realize specific functions such as track mapping,speed heading estimation,and dynamic track management.The feasibility of introducing apparent features for inter-frame data correlation and stable tracking in the tracking and monitoring tasks of remote sensing images of moving ships is demonstrated.