| With the rapid development of earth observation technology and artificial intelligence technology,high-resolution optical imaging satellites keep appearing,and the era of remote sensing big data has arrived.As a commonly used transportation vehicle,the use of massive high-resolution large-scale remote sensing images to achieve rapid and accurate detection of ship targets will play an important role in applications such as sea rescue and route planning.However,due to the limitations of huge image size,sparse target distribution,rich variation of target appearance,insufficient labeled training samples,and other factors,image-based automatic object detection is inefficient and prone to false alarms.In order to meet the practical needs of efficient object detection in large-scale remote sensing images,this paper takes the offshore ships detection on the sea surface as a representative scenario,studies key technologies such as scene classification and target detection,and designs a classification-detection framework that can quickly search large-scale images,locate important ojbect areas,and intelligently assign detection tasks and the main work and research contents are summarized as follows.1.The basic principles of efficient object detection for remote sensing images are introduced,the characteristics of large-scale images are analyzed,and guidance is provided for the research of efficient detection algorithms.The existing algorithms for remote sensing scene classification,object detection and recognition are summarized separately,and the advantages and disadvantages of different algorithms are analyzed.2.A fast four-stage remote sensing offshore ship single class object detection algorithm based on attention mechanism is proposed,which can achieve single class ship target detection with a high recall rate and solve the problems caused by multiscale and sample variability,etc.The effectiveness of the algorithm is verified through experiments.Moreover,the detector has the significant advantages of fewer parameters and low computational complexity.And after model compression and acceleration,the model can be widely applied to different kinds of devices.3.A weakly supervised remote sensing scene classification training method based on noise label distillation is proposed.The method is based on the teacher-student framework and can adapt to data labels with different noise levels.The main purpose of the proposed algorithm is to reduce the human cost while using as much data as possible to enhance the generalization ability and accuracy of the network and improve the model training automaticity and intelligence under the conditions of an insufficient number of remote sensing images labels and frequent data updates.The proposed weakly supervised learning method is also able to be used for completely clean data,used as a knowledge distillation framework,to improve the performance of lightweight models using the efficient learning capability of large models.4.A fast multi-class offshore ship target detection algorithm based on scene encoding attention is proposed for sparse target scenes.Using the classificationdetection framework,target-free regions are identified and filtered by a scene classifier.The object detection task is assigned to the filtered regions,and the foregroundbackground encoding generated by the scene classifier is used to fuse the object detection features to finally achieve fast and efficient localization of offshore ship locations while being able to identify 13 classes of ships. |
