Research On Ship Detection Based On Deep Learning In Remote Sensing Images

Recently,there is a new wave for ship detection due to the maturity of remote sensing technology and the development of deep learning.As a fundamental problem in national defense and marine security,ship detection has been attracting significant attention.However,there are still the following two intractable challenges.(1)For existing synthetic aperture radar data,ships in different images are small.And conventional methods cannot effectively deal with complex background,which seriously affects the accuracy of ship detection,i.e.,small object detection and complex background interference.(2)For existing optical remote sensing data,ships in the same image have different sizes and dense distribution,i.e.,different sizes and dense distribution.To address the above challenges,this thesis proposes and implements a ship detection algorithm based on deep learning in remote sensing images.Firstly,to solve the problems in existing synthetic aperture radar data,this thesis proposes a ship detection algorithm under complex background and small object detection.This method is a stable and effective single-stage model,which consists of three parts: feature refinement network,feature fusion network,and reclassification network.Due to the wide imaging of high resolution remote sensing satellite and the small actual size of some ships,feature refinement network is employed to introduce context information to enhance the ability of small object detection,which can perceive the area around small ship to provide auxiliary information for ship detection.Meanwhile,to integrate the local information of different layers,feature fusion network is utilized to extract the global feature of ships by integrating feature pyramid and refined feature with context information.In addition,deformable convolution is added into reclassification network to improve the receptive field and adapt to the geometric deformation of ships,so that the model is able to further distinguish foreground and background.Extensive experimental results show that the proposed method has strong generalization ability.And the accuracy on AIR-SARShip is 6% higher than the latest Center Net.Next,to solve the problems in existing optical data,this thesis proposes a ship detection algorithm under different sizes and dense distribution.The method is an end-to-end deep network,which is mainly composed of three parts: balanced feature pyramid network,proposals balanced sampling network,and rotational region detection network.To extract discriminative features and improve the robustness to different sizes,balanced feature pyramid network is introduced to enhance multi-level features and balance semantic information between high-level features and low-level features.Meanwhile,to provide reliable proposals for feature pyramid,proposals balanced sampling network is employed to mine hard negative samples to guide the model for ship detection.In addition,to cope with dense distributed ships,a rotational region detection network is introduced to eliminate redundant background and make the model converge better.Extensive experimental results show that the proposed method is superior to the start-of-the-art ship detection methods on optical images.This thesis analyzes the challenges affecting its practical application under different remote sensing datasets.And it researches and implements a ship detection method based on deep learning in remote sensing images,which can be utilized for detection and tracking task.In the future,it has important research significance and application value in the field of national defense and marine surveillance.