Research On Arbitrary-Oriented Ship Detection In Optical Remote Sensing Images

Ship detection in optical remote sensing scenes is an important research topic of computer vision.In recent years,with the development of remote sensing technology,ship detection has become vital to maritime surveillance,combating illegal smuggling and maritime traffic management.However,the existing ship detection methods failed to detect rotated ships accurately and efficiently.Ship targets have extreme scale distributions in complex optical remote sensing scenes.Besides,arbitrary-oriented ship targets require tightly enclosed rotated bounding boxes for positioning.Therefore,it is extremely challenging to detect ships in optical remote sensing scenes.In this thesis,a complete ship detection framework is proposed to achieve accurate and efficient rotated ship detection in optical remote sensing scenes.In addition,a series of novel methods are proposed to improve the performance of ship detection.The main contributions and novelty of this thesis are summarized as follows:1.In order to accurately and efficiently detect rotated ships in large-scale optical remote sensing images,this thesis proposes a one-stage anchor-free detection framework for rotated ship detection based on deep convolutional neural networks.This framework converts the detection of rotated ships into the prediction of the center keypoint and morphological sizes,including the width,height,and rotation angle.This framework simultaneously predicts the probability of the center keypoint and regresses the bounding box,which has a high detection efficiency.In the inference stage,this framework adopts a non-maximum suppression algorithm for rotated objects to merge redundant detection results.Experimental results show that the proposed method achieves the balance between detection efficiency and accuracy.2.Aiming at the difficulty of feature extraction of rotated ships in complex scenes,this thesis proposes a region pooling method for rotated ships detection in the one-stage detection framework.Different from the general Ro I-Pooling method,this method aggregates the global information of the rotated ship to its center keypoint to enhance the feature response,which is helpful for the prediction of object confidence and morphological size.Experimental results show that the proposed method improves the performance of rotated ship detection by effectively mining its global features.3.Aiming at the problem that multi-scale ships are difficult to detect,this thesis proposes a multi-scale ship detection method based on ”feature pyramid” to effectively detect multi-scale ships in multi-level feature maps.This method dynamically generates self-attention heatmaps in different feature levels to guide the feature map of a certain level to pay more attention to the area where the object is not well detected in the previous level.In order to perform joint learning between different prediction branches,a corresponding supervision method is proposed,which is trained based on the fusion result of multi-scale feature maps and promotes the alignment of multi-level feature spaces and the fusion of object confidence.Experimental results show that the proposed method improves the performance of multi-scale ship detection.4.In order to realize the automatic detection of ships in the optical remote sensing images,this paper designs and develops a ship detection prototype system based on the“Browser/Server”(B/S) architecture.