Multi-category Ship Detection In High Resolution Remote Sensing Images

Multi-category ship detection in high-resolution remote sensing images is of great significance for marine monitoring and management,military coastal defense early warn-ing,etc.,but current ship detection still faces many difficulties,such as ships with large aspect ratios,large scale changes,and arbitrary directions.It is often side-by-side or dense in ports,and the data for marking category information of ships is scarce,which brings challenges to multi-category ship target detection.This thesis focuses on explor-ing the side-by-side dense ship detection.The research is carried out from three aspects:dataset,dense ship detection and semi-supervised multi-category ship detection.The main work and contributions are as follows:（1）Aiming at the problems of insufficient number of samples in the current re-mote sensing ship detection dataset,insufficient ship categories and incomplete labeling information,this thesis uses the rotating rectangle labeling format according to the characteristics of the ship target to construct a multi-category ship detection dataset,named WHU-MCSD.The dataset includes a detection dataset with 8611 images to-tal which has 129904 ship instances,and a multi-category detection dataset with 3947images which has 17850 instances and 16 categories.ALL images in WHU-MCSD are1024×1024 pixels.（2）Aiming at the detection of dense or parallel ship targets in remote sensing im-ages,this thesis learn from the characteristics of the space structure of ship targets and proposes a more representative expression of key points,named bow,stern,and both sides.These points are used as the representation of ship targets.At the same time,a spatial attention mechanism is used to supervise the network to learn to better dis-tinguish the characteristics of side-by-side ships,and then a attention based key point shift network is designed called AKSnet.This thesis conducts multiple sets of compar-ative experiments on the proposed WHU-MCSD dataset,and discusses the effects of attention branch and supervision mask generation methods on detection accuracy.The final experimental results show that the AKSnet network has accuracy 83.0%on ship detection tasks.（3）Aiming at the difficulty of multi-category ship sample detection and recognition task in remote sensing image that it is difficult to label the ship categories,this thesis introduces the semi-supervised learning theory into the multi-category detection task of ship.Pseudo label and data augmentation make full use of ship target samples without category labeling information to effectively improve the accuracy of multi-category ship detection.The experimental results on the proposed WHU-MCSD dataset show the effectiveness of the proposed method.For m AP₅₀,the detector in the experiment has at least 9.5%improvement,while the semi-supervised detection method based on AKSnet has average accuracy of 80.2%and it is the best detection model in multi-category ship detection tasks.